Haiku API reference#

Haiku Fundamentals#

Haiku Transforms#

`transform`(f, *[, apply_rng])	Transforms a function using Haiku modules into a pair of pure functions.
`transform_with_state`(f)	Transforms a function using Haiku modules into a pair of pure functions.
`multi_transform`(f)	Transforms a collection of functions using Haiku into pure functions.
`multi_transform_with_state`(f)	Transforms a collection of functions using Haiku into pure functions.
`without_apply_rng`(f)	Removes the rng argument from the apply function.
`without_state`(f)	Wraps a transformed tuple and ignores state in/out.

transform#

haiku.transform(f, *, apply_rng=True)[source]#

Transforms a function using Haiku modules into a pair of pure functions.

For a function out = f(*a, **k) this function returns a pair of two pure functions that call f(*a, **k) explicitly collecting and injecting parameter values:

params = init(rng, *a, **k)
out = apply(params, rng, *a, **k)

Note that the rng argument is typically not required for apply and passing None is accepted.

The first thing to do is to define a Module. A module encapsulates some parameters and a computation on those parameters:

>>> class MyModule(hk.Module):
...   def __call__(self, x):
...     w = hk.get_parameter("w", [], init=jnp.zeros)
...     return x + w

Next, define some function that creates and applies modules. We use transform() to transform that function into a pair of functions that allow us to lift all the parameters out of the function (f.init) and apply the function with a given set of parameters (f.apply):

>>> def f(x):
...   a = MyModule()
...   b = MyModule()
...   return a(x) + b(x)
>>> f = hk.transform(f)

To get the initial state of the module call init with an example input:

>>> params = f.init(None, 1)
>>> params
{'my_module': {'w': ...Array(0., dtype=float32)},
 'my_module_1': {'w': ...Array(0., dtype=float32)}}

You can then apply the function with the given parameters by calling apply (note that since we don’t use Haiku’s random number APIs to apply our network we pass None as an RNG key):

>>> print(f.apply(params, None, 1))
2.0

It is expected that your program will at some point produce updated parameters and you will want to re-apply apply. You can do this by calling apply with different parameters:

>>> new_params = {"my_module": {"w": jnp.array(2.)},
...               "my_module_1": {"w": jnp.array(3.)}}
>>> print(f.apply(new_params, None, 2))
9.0

If your transformed function needs to maintain internal state (e.g. moving averages in batch norm) then see transform_with_state().

Parameters

f – A function closing over Module instances.
apply_rng – In the process of being removed. Can only value True.

Return type

Transformed

Returns

A Transformed tuple with init and apply pure functions.

transform_with_state#

haiku.transform_with_state(f)[source]#

Transforms a function using Haiku modules into a pair of pure functions.

See transform() for general details on Haiku transformations.

For a function out = f(*a, **k) this function returns a pair of two pure functions that call f(*a, **k) explicitly collecting and injecting parameter values and state:

params, state = init(rng, *a, **k)
out, state = apply(params, state, rng, *a, **k)

Note that the rng argument is typically not required for apply and passing None is accepted.

This function is equivalent to transform(), however it allows you to maintain and update internal state (e.g. ExponentialMovingAverage in BatchNorm) via get_state() and set_state():

>>> def f():
...   counter = hk.get_state("counter", shape=[], dtype=jnp.int32,
...                          init=jnp.zeros)
...   hk.set_state("counter", counter + 1)
...   return counter
>>> f = hk.transform_with_state(f)

>>> params, state = f.init(None)
>>> for _ in range(10):
...   counter, state = f.apply(params, state, None)
>>> print(counter)
9

Parameters: f – A function closing over Module instances.
Return type: TransformedWithState
Returns: A TransformedWithState tuple with init and apply pure functions.

multi_transform#

haiku.multi_transform(f)[source]#

Transforms a collection of functions using Haiku into pure functions.

In many scenarios we have several modules which are used either as primitives for several Haiku modules/functions, or whose pure versions are to be reused in downstream code. This utility enables this by applying transform() to an arbitrary tree of Haiku functions which share modules and have a common init function.

f is expected to return a tuple of two elements. First is a template Haiku function which provides an example of how all internal Haiku modules are connected. This function is used to create a common init function (with your parameters).

The second object is an arbitrary tree of Haiku functions all of which reuse the modules connected in the template function. These functions are transformed to pure apply functions.

Example:

>>> def f():
...   encoder = hk.Linear(1, name="encoder")
...   decoder = hk.Linear(1, name="decoder")
...
...   def init(x):
...     z = encoder(x)
...     return decoder(z)
...
...   return init, (encoder, decoder)

>>> f = hk.multi_transform(f)
>>> rng = jax.random.PRNGKey(42)
>>> x = jnp.ones([1, 1])
>>> params = f.init(rng, x)
>>> jax.tree_util.tree_map(jnp.shape, params)
{'decoder': {'b': (1,), 'w': (1, 1)},
 'encoder': {'b': (1,), 'w': (1, 1)}}

>>> encode, decode = f.apply
>>> z = encode(params, None, x)
>>> y = decode(params, None, z)

Parameters

f (Callable[[], tuple[TemplateFn, TreeOfApplyFns]]) – A factory function that returns two functions, firstly a common init function that creates all modules, and secondly a pytree of apply functions which make use of those modules.

Return type

MultiTransformed

Returns

A MultiTransformed instance which contains a pure init function: that creates all parameters, and a pytree of pure apply functions that given the params apply the given function.

See also

multi_transform_with_state(): Equivalent for modules using state.

multi_transform_with_state#

haiku.multi_transform_with_state(f)[source]#

Transforms a collection of functions using Haiku into pure functions.

See multi_transform() for more details.

Example:

>>> def f():
...   encoder = hk.Linear(1, name="encoder")
...   decoder = hk.Linear(1, name="decoder")
...
...   def init(x):
...     z = encoder(x)
...     return decoder(z)
...
...   return init, (encoder, decoder)

>>> f = hk.multi_transform_with_state(f)
>>> rng = jax.random.PRNGKey(42)
>>> x = jnp.ones([1, 1])
>>> params, state = f.init(rng, x)
>>> jax.tree_util.tree_map(jnp.shape, params)
{'decoder': {'b': (1,), 'w': (1, 1)},
 'encoder': {'b': (1,), 'w': (1, 1)}}

>>> encode, decode = f.apply
>>> z, state = encode(params, state, None, x)
>>> y, state = decode(params, state, None, z)

Parameters: f (Callable[[], tuple[TemplateFn, TreeOfApplyFns]]) – Function returning a “template” function and an arbitrary tree of functions using modules connected in the template function.
Return type: MultiTransformedWithState
Returns: An init function and a tree of pure apply functions.

See also

transform_with_state(): Transform a single apply function.
multi_transform(): Transform multiple apply functions without state.

without_apply_rng#

haiku.without_apply_rng(f)[source]#

Removes the rng argument from the apply function.

This is a convenience wrapper that makes the rng argument to f.apply default to None. This is useful when f doesn’t actually use random numbers as part of its computation, such that the rng argument wouldn’t be used. Note that if f does use random numbers, this will cause an error to be thrown complaining that f needs a non-None PRNGKey.

Parameters: f (TransformedT) – A transformed function.
Return type: TransformedT
Returns: The same transformed function, with a modified apply.

without_state#

haiku.without_state(f)[source]#

Wraps a transformed tuple and ignores state in/out.

The example below is equivalent to f = hk.transform(f):

>>> def f(x):
...   mod = hk.Linear(10)
...   return mod(x)
>>> f = hk.without_state(hk.transform_with_state(f))
>>> rng = jax.random.PRNGKey(42)
>>> x = jnp.zeros([1, 1])
>>> params = f.init(rng, x)
>>> print(f.apply(params, rng, x))
[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]

Parameters: f (TransformedWithState) – A transformed function.
Return type: Transformed
Returns: A transformed function that does not take or return state.

with_empty_state#

haiku.with_empty_state(f)[source]#

Wraps a transformed tuple and passes empty state in/out.

The example below is equivalent to f = hk.transform_with_state(f):

>>> def f(x):
...   mod = hk.Linear(10)
...   return mod(x)
>>> f = hk.with_empty_state(hk.transform(f))
>>> rng = jax.random.PRNGKey(42)
>>> x = jnp.zeros([1, 1])
>>> params, state = f.init(rng, x)
>>> state
{}
>>> out, state = f.apply(params, state, rng, x)
>>> print(out)
[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]
>>> state
{}

Parameters: f (Transformed) – A transformed function.
Return type: TransformedWithState
Returns: A transformed function that does accepts and returns state.

Modules, Parameters and State#

`Module`([name])	Base class for Haiku modules.
`to_module`(f)	Converts a function into a callable module class.
`get_parameter`(name, shape[, dtype, init])	Creates or reuses a parameter for the given transformed function.
`get_state`(name[, shape, dtype, init])	Gets the current value for state with an optional initializer.
`set_state`(name, value)	Sets the current value for some state.

Module#

class haiku.Module(name=None)[source]#

Base class for Haiku modules.

A Haiku module is a lightweight container for variables and other modules. Modules typically define one or more “forward” methods (e.g. __call__) which apply operations combining user input and module parameters.

Modules must be initialized inside a transform() call.

For example:

>>> class AddModule(hk.Module):
...   def __call__(self, x):
...     w = hk.get_parameter("w", [], init=jnp.ones)
...     return x + w

>>> def forward_fn(x):
...   mod = AddModule()
...   return mod(x)

>>> forward = hk.transform(forward_fn)
>>> x = 1.
>>> rng = None
>>> params = forward.init(rng, x)
>>> print(forward.apply(params, None, x))
2.0

__init__(name=None)[source]#

Initializes the current module with the given name.

Subclasses should call this constructor before creating other modules or variables such that those modules are named correctly.

Parameters: name (Optional[str]) – An optional string name for the class. Must be a valid Python identifier. If name is not provided then the class name for the current instance is converted to lower_snake_case and used instead.

__post_init__(name=None)[source]#

Initializes the current module with the given name.

Subclasses should call this constructor before creating other modules or variables such that those modules are named correctly.

Parameters: name (Optional[str]) – An optional string name for the class. Must be a valid Python identifier. If name is not provided then the class name for the current instance is converted to lower_snake_case and used instead.

params_dict()[source]#

Returns parameters keyed by name for this module and submodules.

Return type: Mapping[str, jnp.ndarray]

state_dict()[source]#

Returns state keyed by name for this module and submodules.

Return type: Mapping[str, jnp.ndarray]

to_module#

haiku.to_module(f)[source]#

Converts a function into a callable module class.

Sample usage:

>>> def bias_fn(x):
...   b = hk.get_parameter("b", [], init=hk.initializers.RandomNormal())
...   return x + b
>>> Bias = hk.to_module(bias_fn)
>>> def net(x, y):
...   b = Bias(name="my_bias")
...   # Bias x and y by the same amount.
...   return b(x) * b(y)

Parameters: f (Callable[..., Any]) – The function to convert.
Return type: type[CallableModule]
Returns: A module class which runs f when called.

get_parameter#

haiku.get_parameter(name, shape, dtype=<class 'jax.numpy.float32'>, init=None)[source]#

Creates or reuses a parameter for the given transformed function.

>>> print(hk.get_parameter("w", [], init=jnp.ones))
1.0

Parameters within the same transform() and/or Module with the same name have the same value:

>>> w1 = hk.get_parameter("w", [], init=jnp.zeros)
>>> w2 = hk.get_parameter("w", [], init=jnp.zeros)
>>> assert w1 is w2

Parameters

name (str) – A name for the parameter.
shape (Sequence[int]) – The shape of the parameter.
dtype (Any) – The dtype of the parameter.
init (Optional[Initializer]) – A callable of shape, dtype to generate an initial value for the parameter.

Return type

jax.Array

Returns

A jax.Array with the parameter of the given shape.

get_state#

haiku.get_state(name, shape=None, dtype=<class 'jax.numpy.float32'>, init=None)[source]#

Gets the current value for state with an optional initializer.

“State” can be used to represent mutable state in your network. The most common usage of state is to represent the moving averages used in batch normalization (see ExponentialMovingAverage). If your network uses “state” then you are required to use transform_with_state() and pass state into and out of the apply function.

>>> print(hk.get_state("counter", [], init=jnp.zeros))
0.0

If the value for the given state is already defined (e.g. using set_state()) then you can call with just the name:

>>> print(hk.get_state("counter"))
0.0

NOTE: state within the same transform() and/or Module with the same name have the same value:

>>> c1 = hk.get_state("counter")
>>> c2 = hk.get_state("counter")
>>> assert c1 is c2

Parameters

name (str) – A name for the state.
shape (Optional[Sequence[int]]) – The shape of the state.
dtype (Any) – The dtype of the state.
init (Optional[Initializer]) – A callable f(shape, dtype) that returns an initial value for the state.

Return type

jax.Array

Returns

A jax.Array with the state of the given shape.

set_state#

haiku.set_state(name, value)[source]#

Sets the current value for some state.

See get_state().

>>> hk.set_state("counter", jnp.zeros([]))
>>> print(hk.get_state("counter"))
0.0

NOTE: state within the same transform() and/or Module with the same name have the same value:

>>> w1 = hk.get_state("counter")
>>> w2 = hk.get_state("counter")
>>> assert w1 is w2

Parameters

name (str) – A name for the state.
value – A value to set.

Getters and Interceptors#

`custom_creator`(creator, *[, params, state])	Registers a custom parameter and/or state creator.
`custom_getter`(getter, *[, params, state])	Registers a custom parameter or state getter.
`custom_setter`(setter)	Registers a custom state setter.
`GetterContext`(full_name, module, ...)	Context about where parameters are being created.
`SetterContext`(full_name, module, ...)	Context about where state is being set.
`intercept_methods`(interceptor)	Register a new method interceptor.
`MethodContext`(module, method_name, ...)	Read only state showing the calling context for a method.

custom_creator#

haiku.custom_creator(creator, *, params=True, state=False)[source]#

Registers a custom parameter and/or state creator.

When new parameters are created via get_parameter() we first run custom creators passing user defined values through. For example:

>>> def zeros_creator(next_creator, shape, dtype, init, context):
...   init = jnp.zeros
...   return next_creator(shape, dtype, init)

>>> with hk.custom_creator(zeros_creator):
...   z = hk.get_parameter("z", [], jnp.float32, jnp.ones)
>>> print(z)
0.0

If state=True then your creator will additionally run on calls to get_state():

>>> with hk.custom_creator(zeros_creator, state=True):
...   z = hk.get_state("z", [], jnp.float32, jnp.ones)
>>> print(z)
0.0

Parameters

creator (Creator) – A parameter creator.
params (bool) – Whether to intercept parameter creation, defaults to True.
state (bool) – Whether to intercept state creation, defaults to False.

Return type

contextlib.AbstractContextManager

Returns

Context manager under which the creator is active.

custom_getter#

haiku.custom_getter(getter, *, params=True, state=False)[source]#

Registers a custom parameter or state getter.

When parameters are retrieved using get_parameter() we always run all custom getters before returning a value to the user.

>>> def bf16_getter(next_getter, value, context):
...   value = value.astype(jnp.bfloat16)
...   return next_getter(value)

>>> with hk.custom_getter(bf16_getter):
...   w = hk.get_parameter("w", [], jnp.float32, jnp.ones)
>>> w.dtype
dtype(bfloat16)

If state=True the getter will additionally run for calls to get_state():

>>> with hk.custom_getter(bf16_getter, state=True):
...   c = hk.get_state("c", [], jnp.float32, jnp.ones)
>>> c.dtype
dtype(bfloat16)

Parameters

getter (Getter) – A parameter getter.
params (bool) – Whether the getter should run on get_parameter()
state (bool) – Whether the getter should run on get_state().

Return type

contextlib.AbstractContextManager

Returns

Context manager under which the getter is active.

custom_setter#

haiku.custom_setter(setter)[source]#

Registers a custom state setter.

When state is set using set_state() we always run all custom setters before saving the value.

>>> def zero_during_init(next_setter, value, context):
...   if hk.running_init():
...     value = jnp.zeros_like(value)
...   return next_setter(value)

>>> with hk.custom_setter(zero_during_init):
...   hk.set_state("x", jnp.ones([2]))
...   x = hk.get_state("x")
>>> print(x)
[0. 0.]

Parameters: setter (Setter) – A state setter.
Return type: contextlib.AbstractContextManager
Returns: Context manager under which the setter is active.

GetterContext#

class haiku.GetterContext(full_name: str, module: Optional[Module], original_dtype: Any, original_shape: Sequence[int], original_init: Optional[Initializer], lifted_prefix_name: Optional[str])[source]#

Context about where parameters are being created.

full_name#

The full name of the given parameter (e.g. mlp/~/linear_0/w).

Type: str

module#

The module that owns the current parameter, None if this parameter exists outside any module.

Type: Optional[Module]

original_dtype#

The dtype that get_parameter() or get_state() was originally called with.

Type: Any

original_shape#

The shape that get_parameter() or get_state() was originally called with.

Type: Sequence[int]

original_init#

The initializer that get_parameter() or get_state() was originally called with.

Type: Optional[Initializer]

lifted_prefix_name#

The module names of all enclosing lifted modules (see lift() for more context). Adding this string as a prefix to full_name will be equal to the final parameter name in the outer transform’s parameter dictionary. NOTE: When get_parameter() or get_state() is called in an apply context, this name will always be None because only init functions are lifted.

Type: Optional[str]

module_name#: The full name of enclosing modules.

name#: The name of this parameter.

SetterContext#

class haiku.SetterContext(full_name: str, module: Optional[Module], original_dtype: Any, original_shape: Sequence[int], lifted_prefix_name: Optional[str])[source]#

Context about where state is being set.

full_name#

The full name of the given state (e.g. mlp/~/linear_0/w).

Type: str

module#

The module that owns the current state, None if this state exists outside any module.

Type: Optional[Module]

original_dtype#

The dtype that set_state() was originally called with.

Type: Any

original_shape#

The shape that set_state() or get_state() was originally called with.

Type: Sequence[int]

lifted_prefix_name#

Type: Optional[str]

module_name#: The full name of enclosing modules.

name#: The name of this state.

intercept_methods#

haiku.intercept_methods(interceptor)[source]#

Method interceptors allow you to (at a distance) intercept method calls to modules and modify args/kwargs before calling the underlying method. After the underlying method is called you can modify its result before it is passed back to the user.

For example you could intercept method calls to BatchNorm and ensure it is always computed in full precision:

>>> def my_interceptor(next_f, args, kwargs, context):
...   if (type(context.module) is not hk.BatchNorm
...       or context.method_name != "__call__"):
...     # We ignore methods other than BatchNorm.__call__.
...     return next_f(*args, **kwargs)
...
...   def cast_if_array(x):
...     if isinstance(x, jax.Array):
...       x = x.astype(jnp.float32)
...     return x
...
...   args, kwargs = jax.tree_util.tree_map(cast_if_array, (args, kwargs))
...   out = next_f(*args, **kwargs)
...   return out

We can create and use our module in the usual way, we just need to wrap any method calls we want to intercept in the context manager:

>>> mod = hk.BatchNorm(decay_rate=0.9, create_scale=True, create_offset=True)
>>> x = jnp.ones([], jnp.bfloat16)
>>> with hk.intercept_methods(my_interceptor):
...   out = mod(x, is_training=True)
>>> assert out.dtype == jnp.float32

Without the interceptor BatchNorm would compute in bf16, however since we cast x before the underlying method is called we compute in f32.

Parameters: interceptor (MethodGetter) – A method interceptor.
Returns: Context manager under which the interceptor is active.

MethodContext#

class haiku.MethodContext(module: 'Module', method_name: str, orig_method: Callable[..., Any], orig_class: type['Module'])[source]#

Read only state showing the calling context for a method.

For example, let’s define two interceptors and print the values in the context. Additionally, we will make the first interceptor conditionally short circuit, since interceptors stack and are run in order, an earlier interceptor can decide to call the next interceptor, or short circuit and call the underlying method directly:

>>> module = hk.Linear(1, name="method_context_example")
>>> short_circuit = False

>>> def my_interceptor_1(next_fun, args, kwargs, context):
...   print('running my_interceptor_1')
...   print('- module.name: ', context.module.name)
...   print('- method_name: ', context.method_name)
...   if short_circuit:
...     return context.orig_method(*args, **kwargs)
...   else:
...     return next_fun(*args, **kwargs)
>>> def my_interceptor_2(next_fun, args, kwargs, context):
...   print('running my_interceptor_2')
...   print('- module.name: ', context.module.name)
...   print('- method_name: ', context.method_name)
...   return next_fun(*args, **kwargs)

When short_circuit=False the two interceptors will run in order:

>>> with hk.intercept_methods(my_interceptor_1), \
...      hk.intercept_methods(my_interceptor_2):
...   _ = module(jnp.ones([1, 1]))
running my_interceptor_1
- module.name:  method_context_example
- method_name:  __call__
running my_interceptor_2
- module.name:  method_context_example
- method_name:  __call__

Setting short_circuit=True will cause the first interceptor to call the original method (rather than next_fun which will trigger the next interceptor):

>>> short_circuit = True
>>> with hk.intercept_methods(my_interceptor_1), \
...      hk.intercept_methods(my_interceptor_2):
...   _ = module(jnp.ones([1, 1]))
running my_interceptor_1
- module.name:  method_context_example
- method_name:  __call__

module#

A Module instance whose method is being called.

Type: ‘Module’

method_name#

The name of the method being called on the module.

Type: str

orig_method#

The underlying method on the module which when called will not trigger interceptors. You should only call this if you want to short circuit all the other interceptors, in general you should prefer to call the next_fun passed to your interceptor which will run orig_method after running all other interceptors.

Type: Callable[…, Any]

orig_class#

The class which defined orig_method. Note that when using inheritance this is not necessarily the same as type(module).

Type: type[‘Module’]

Random Numbers#

`PRNGSequence`(key_or_seed)	Iterator of JAX random keys.
`next_rng_key`()	Returns a unique JAX random key split from the current global key.
`next_rng_keys`(num)	Returns one or more JAX random keys split from the current global key.
`maybe_next_rng_key`()	`next_rng_key()` if random numbers are available, else `None`.
`reserve_rng_keys`(num)	Pre-allocate some number of JAX RNG keys.
`with_rng`(key)	Provides a new sequence for `next_rng_key()` to draw from.
`maybe_get_rng_sequence_state`()	Returns the internal state of the PRNG sequence.
`replace_rng_sequence_state`(state)	Replaces the internal state of the PRNG sequence with the given state.

PRNGSequence#

class haiku.PRNGSequence(key_or_seed)[source]#

Iterator of JAX random keys.

>>> seq = hk.PRNGSequence(42)  # OR pass a jax.random.PRNGKey
>>> key1 = next(seq)
>>> key2 = next(seq)
>>> assert key1 is not key2

If you know how many keys you will want then you can use reserve() to more efficiently split the keys you need:

>>> seq.reserve(4)
>>> keys = [next(seq) for _ in range(4)]

__init__(key_or_seed)[source]#: Creates a new PRNGSequence.

reserve(num)[source]#: Splits additional num keys for later use.

__next__()[source]#

Return the next item from the iterator. When exhausted, raise StopIteration

Return type: PRNGKey

next()[source]#

Return the next item from the iterator. When exhausted, raise StopIteration

Return type: PRNGKey

next_rng_key#

haiku.next_rng_key()[source]#

Returns a unique JAX random key split from the current global key.

>>> key = hk.next_rng_key()
>>> _ = jax.random.uniform(key, [])

Return type: PRNGKey
Returns: A unique (within a call to init or apply) JAX rng key that can be used with APIs such as jax.random.uniform().

next_rng_keys#

haiku.next_rng_keys(num)[source]#

Returns one or more JAX random keys split from the current global key.

>>> k1, k2 = hk.next_rng_keys(2)
>>> assert (k1 != k2).all()
>>> a = jax.random.uniform(k1, [])
>>> b = jax.random.uniform(k2, [])
>>> assert a != b

Parameters: num (int) – The number of keys to split.
Return type: jax.Array
Returns: An array of shape [num, 2] unique (within a transformed function) JAX rng keys that can be used with APIs such as jax.random.uniform().

maybe_next_rng_key#

haiku.maybe_next_rng_key()[source]#

next_rng_key() if random numbers are available, else None.

Return type: Optional[PRNGKey]

reserve_rng_keys#

haiku.reserve_rng_keys(num)[source]#

Pre-allocate some number of JAX RNG keys.

See next_rng_key().

This API offers a way to micro-optimize how RNG keys are split when using Haiku. It is unlikely that you need it unless you find compilation time of your init function to be a problem, or you sample a lot of random numbers in apply.

>>> hk.reserve_rng_keys(2)  # Pre-allocate 2 keys for us to consume.
>>> _ = hk.next_rng_key()   # Takes the first pre-allocated key.
>>> _ = hk.next_rng_key()   # Takes the second pre-allocated key.
>>> _ = hk.next_rng_key()   # Splits a new key.

Parameters: num (int) – The number of JAX rng keys to allocate.

with_rng#

haiku.with_rng(key)[source]#

Provides a new sequence for next_rng_key() to draw from.

When next_rng_key() is called, it draws a new key from the PRNGSequence defined by the input key to the transformed function. This context manager overrides the sequence for the duration of the scope.

>>> with hk.with_rng(jax.random.PRNGKey(428)):
...   s = jax.random.uniform(hk.next_rng_key(), ())
>>> print("{:.1f}".format(s))
0.5

Parameters: key (PRNGKey) – The key to seed the sequence with.
Returns: Context manager under which the given sequence is active.

maybe_get_rng_sequence_state#

haiku.maybe_get_rng_sequence_state()[source]#

Returns the internal state of the PRNG sequence.

Return type: Optional[PRNGSequenceState]
Returns: The internal state if random numbers are available, else None.

replace_rng_sequence_state#

haiku.replace_rng_sequence_state(state)[source]#

Replaces the internal state of the PRNG sequence with the given state.

Parameters: state (PRNGSequenceState) – The new internal state or None.
Raises: MissingRNGError – If random numbers aren’t available.

Type Hints#

`LSTMState`(hidden, cell)	An LSTM core state consists of hidden and cell vectors.
`Params`	A Mapping is a generic container for associating key/value pairs.
`MutableParams`	A MutableMapping is a generic container for associating key/value pairs.
`State`	A Mapping is a generic container for associating key/value pairs.
`MutableState`	A MutableMapping is a generic container for associating key/value pairs.
`Transformed`(init, apply)	Holds a pair of pure functions.
`TransformedWithState`(init, apply)	Holds a pair of pure functions.
`MultiTransformed`(init, apply)	Holds a collection of pure functions.
`MultiTransformedWithState`(init, apply)	Holds a collection of pure functions.
`ModuleProtocol`(args, *kwargs)	Protocol for Module like types.
`SupportsCall`(args, *kwargs)	Protocol for Module like types that are Callable.

LSTMState#

class haiku.LSTMState(hidden: jax.Array, cell: jax.Array)[source]#

An LSTM core state consists of hidden and cell vectors.

hidden#

Hidden state.

Type: jax.Array

cell#

Cell state.

Type: jax.Array

Params#

haiku.Params#: alias of collections.abc.Mapping[str, collections.abc.Mapping[str, jax.Array]]

MutableParams#

haiku.MutableParams#: alias of collections.abc.MutableMapping[str, collections.abc.MutableMapping[str, jax.Array]]

State#

haiku.State#: alias of collections.abc.Mapping[str, collections.abc.Mapping[str, jax.Array]]

MutableState#

haiku.MutableState#: alias of collections.abc.MutableMapping[str, collections.abc.MutableMapping[str, jax.Array]]

Transformed#

class haiku.Transformed(init: Callable[..., hk.MutableParams], apply: Callable[..., Any])[source]#

Holds a pair of pure functions.

init#

A pure function: params = init(rng, *a, **k)

Type: Callable[…, hk.MutableParams]

apply#

A pure function: out = apply(params, rng, *a, **k)

Type: Callable[…, Any]

TransformedWithState#

class haiku.TransformedWithState(init: Callable[..., tuple[hk.MutableParams, hk.MutableState]], apply: Callable[..., tuple[Any, hk.MutableState]])[source]#

Holds a pair of pure functions.

init#

A pure function: params, state = init(rng, *a, **k)

Type: Callable[…, tuple[hk.MutableParams, hk.MutableState]]

apply#

A pure function: out, state = apply(params, state, rng, *a, **k)

Type: Callable[…, tuple[Any, hk.MutableState]]

MultiTransformed#

class haiku.MultiTransformed(init: Callable[..., hk.MutableParams], apply: Any)[source]#

Holds a collection of pure functions.

init#

A pure function: params = init(rng, *a, **k)

Type: Callable[…, hk.MutableParams]

apply#

A JAX tree of pure functions each with the signature: out = apply(params, rng, *a, **k).

Type: Any

See also

Transformed: Single apply variant of multi-transform.
MultiTransformedWithState: Multi apply with state variant.

MultiTransformedWithState#

class haiku.MultiTransformedWithState(init: Callable[..., tuple[hk.MutableParams, hk.MutableState]], apply: Any)[source]#

Holds a collection of pure functions.

init#

A pure function: params, state = init(rng, *a, **k)

Type: Callable[…, tuple[hk.MutableParams, hk.MutableState]]

apply#

A JAX tree of pure functions each with the signature: out, state = apply(params, state, rng, *a, **k).

Type: Any

See also

TransformedWithState: Single apply variant of multi-transform.
MultiTransformed: Multi apply with state variant.

ModuleProtocol#

class haiku.ModuleProtocol(*args, **kwargs)[source]#: Protocol for Module like types.

SupportsCall#

class haiku.SupportsCall(*args, **kwargs)[source]#

Protocol for Module like types that are Callable.

Being a protocol means you don’t need to explicitly extend this type in order to support instance checks with it. For example, Linear only extends Module, however since it conforms (e.g. implements __call__) to this protocol you can instance check using it:

>>> assert isinstance(hk.Linear(1), hk.SupportsCall)

Flax Interop#

Haiku inside Flax#

Module#

class haiku.experimental.flax.Module(transformed, parent=<flax.linen.module._Sentinel object>, name=None)[source]#

A Flax nn.Module that runs a Haiku transformed function.

This type is designed to make it easy to take a Haiku transformed function and/or a Haiku module and use it inside a program that otherwise uses Flax.

Given a Haiku transformed function

>>> def f(x):
...   return hk.Linear(1)(x)
>>> f = hk.transform(f)

You can convert it into a Flax module using:

>>> mod = hk.experimental.flax.Module(f)

Calling this module is the same as calling any regular Flax module:

>>> rng = jax.random.PRNGKey(42)
>>> x = jnp.ones([1, 1])
>>> variables = mod.init(rng, x)
>>> out = mod.apply(variables, x)

If you just want to convert a Haiku module class such that it can be used with Flax you can use the create class method:

>>> mod = hk.experimental.flax.Module.create(hk.Linear, 1)
>>> variables = mod.init(rng, x)
>>> out = mod.apply(variables, x)

flatten_flax_to_haiku#

haiku.experimental.flax.flatten_flax_to_haiku(collection)[source]#

Flattens a Flax variable collection (e.g. params) to a Haiku dict.

Return type: HaikuParamsOrState

Flax inside Haiku#

lift#

haiku.experimental.flax.lift(mod, *, name)[source]#

Lifts a flax nn.Module into a Haiku transformed function.

For a Flax Module (e.g. mod = nn.Dense(10)), mod = lift(mod) allows you to run the call method of the module as if the module was a regular Haiku module.

Parameters and state from the Flax module are registered with Haiku and become part of the params/state dictionaries (as returned from init/apply).

>>> def f(x):
...   # Create and "lift" a Flax module.
...   mod = hk.experimental.flax.lift(nn.Dense(300), name='dense')
...   x = mod(x)                  # Any params/state will be registered
...                               # with Haiku when applying the module.
...   x = jax.nn.relu(x)
...   x = hk.nets.MLP([100, 10])  # You can of course mix Haiku modules in.
...   return x
>>> f = hk.transform(f)
>>> rng = jax.random.PRNGKey(42)
>>> x = jnp.ones([1, 1])
>>> params = f.init(rng, x)
>>> out = f.apply(params, None, x)

Parameters

mod (nn.Module) – Any Flax nn.Module instance.
name (str) – Name scope to prefix entries in the outer params/state dict.

Return type

Callable[…, Any]

Returns

A function that when applied calls the call method of the given Flax module and returns its output. As a side effect of calling the module any module parameters and state variables are registered with Haiku.

Advanced State Management#

Lifting#

`lift`(init_fn, *[, allow_reuse, name])	Registers parameters from an inner init function in an outer transform.
`lift_with_state`(init_fn, *[, allow_reuse, name])	Registers params and state from an init function in an outer transform.
`transparent_lift`(init_fn, *[, allow_reuse])	Registers parameters in an outer transform without adding a name scope.
`transparent_lift_with_state`(init_fn, *[, ...])	Registers params and state in an outer transform without adding scope.
`LiftWithStateUpdater`(name)	Handles updating the state for a lift_with_state computation.

lift#

haiku.lift(init_fn, *, allow_reuse=False, name='lifted')[source]#

Registers parameters from an inner init function in an outer transform.

HINT: lift() is for when you want to make non-trivial use of JAX transforms (e.g. jax.vmap) inside of a transform() or transform_with_state(). We generally recommend trying to use JAX transforms on the pure functions returned by transform(), in which case you do not need lift().

Use lift()when nesting Haiku transforms to register the parameters of the inner transform in any outer transform. This is mainly useful when using JAX functions inside of a Haiku module (eg. using jax.vmap on a layer). See https://dm-haiku.readthedocs.io/en/latest/notebooks/transforms.html#Using-hk.lift for more explanation of when to use lift(). (If you’re not using JAX functions inside of a module or don’t need access to your parameters inside of a transform, you probably don’t need to use lift())

Must be called inside transform(), and be passed the init member of a Transformed.

During init, the returned callable will run the given init_fn, and include the resulting params in the outer transform’s dictionaries. During apply, the returned callable will instead pull the relevant parameters from the outer transform’s dictionaries.

By default, users must ensure that the given init does not accidentally catch modules from an outer transform() via functional closure. If this behavior is desirable, set allow_reuse to True.

Example:

A common usage of lift() is to use JAX transformations like vmap in non-trivial ways, inside a transform(). For example, we can use lift() and jax.vmap to create an ensemble.

First we’ll create a helper function that uses lift() to apply vmap to our model. As you can see from the comments, we are using vmap to change how parameters should be created (in this case we create a unique set of parameters for each member of the ensemble) and we change how apply works (we “map” the parameters meaning JAX will compute the forward pass separately, in parallel, for each member of the ensemble):

>>> def create_ensemble(model, size: int):
...   init_rng = hk.next_rng_keys(size) if hk.running_init() else None
...   model = hk.transform(model)
...   # in_axes: rng is mapped, data is not.
...   init_model = jax.vmap(model.init, in_axes=(0, None))
...   # Use hk.lift to "lift" parameters created by `init_model` into the
...   # outer transform.
...   init_model = hk.lift(init_model, name="ensemble")
...   def ensemble(x):
...     params = init_model(init_rng, x)
...     # in_axes: params are mapped, rng/data are not.
...     return jax.vmap(model.apply, in_axes=(0, None, None))(params, None, x)
...   return ensemble

We can now use this function to ensemble any Haiku module(s), inside of a transform. First we define a function for each member of the ensemble:

>>> def member_fn(x):
...   return hk.nets.MLP([300, 100, 10])(x)

Secondly we can combine our two functions, inside a transform() to create an ensemble:

>>> def f(x):
...   ensemble = create_ensemble(member_fn, size=4)
...   x = ensemble(x)
...   # You could create other modules here which were not ensembled.
...   return x
>>> f = hk.transform(f)

When we initialize the network, our ensemble member’s parameters have a leading dimension the size of the ensemble:

>>> rng = jax.random.PRNGKey(777)
>>> x = jnp.ones([32, 128])
>>> params = f.init(rng, x)
>>> jax.tree_util.tree_map(lambda x: x.shape, params)
{'ensemble/mlp/~/linear_0': {'b': (4, 300), 'w': (4, 128, 300)},
 'ensemble/mlp/~/linear_1': {'b': (4, 100), 'w': (4, 300, 100)},
 'ensemble/mlp/~/linear_2': {'b': (4, 10), 'w': (4, 100, 10)}}

When we apply the network, we get an output for each member of the ensemble for the entire batch:

>>> y = f.apply(params, None, x)
>>> y.shape
(4, 32, 10)

Parameters

init_fn (Callable[..., hk.Params]) – The init function from an Transformed.
allow_reuse (bool) – Allows lifted parameters and state to be reused from the outer transform(). This can be desirable when using lift within control flow (e.g. hk.scan).
name (str) – A string name to prefix parameters with.

Return type

Callable[…, hk.Params]

Returns

A callable that during init injects parameter values into the outer context and during apply retrieves parameters from the outer context. In both cases returns parameter values to be used with an apply function.

See also

lift_with_state(): Register params and state with an outer transform.
transparent_lift(): Register params with an outer transform without a namespace.
transparent_lift_with_state(): Register params and state with an outer transform without a namespace.

lift_with_state#

haiku.lift_with_state(init_fn, *, allow_reuse=False, name='lifted')[source]#

Registers params and state from an init function in an outer transform.

See lift()for more context on when to use lift.

This function returns two objects. The first is a callable that runs your init function with slightly different behaviour based on if it’s run during init vs. apply time. The second is an updater that can be used to pass updated state values that result from running your apply function. See later in the docs for a worked example.

During init, the returned callable will run the given init_fn, and include the resulting params/state in the outer transform’s dictionaries. During apply, the returned callable will instead pull the relevant params/state from the outer transform’s dictionaries.

Must be called inside transform_with_state(), and be passed the init member of a TransformedWithState.

By default, users must ensure that the given init does not accidentally catch modules from an outer transform_with_state() via functional closure. If this behavior is desirable, set allow_reuse to True.

Example

>>> def g(x):
...   return hk.nets.ResNet50(1)(x, True)
>>> g = hk.transform_with_state(g)
>>> params_and_state_fn, updater = (
...   hk.lift_with_state(g.init, name='f_lift'))
>>> init_rng = hk.next_rng_key() if hk.running_init() else None
>>> x = jnp.ones([1, 224, 224, 3])
>>> params, state = params_and_state_fn(init_rng, x)
>>> out, state = g.apply(params, state, None, x)
>>> updater.update(state)

Parameters

init_fn (Callable[..., tuple[hk.Params, hk.State]]) – The init function from an TransformedWithState.
allow_reuse (bool) – Allows lifted parameters and state to be reused from the outer transform_with_state(). This can be desirable when using lift_with_state within control flow (e.g. hk.scan).
name (str) – A string name to prefix parameters with.

Return type

tuple[Callable[…, tuple[hk.Params, hk.State]], LiftWithStateUpdater]

Returns

A callable that during init injects parameter values into the outer context and during apply reuses parameters from the outer context. In both cases returns parameter values to be used with an apply function. The init function additionally returns an object used to update the outer context with new state after apply is called.

See also

lift(): Register parameters with an outer transform.
transparent_lift(): Register parameters with an outer transform without a namespace.
transparent_lift_with_state(): Register parameters and state with an outer transform without a namespace.

transparent_lift#

haiku.transparent_lift(init_fn, *, allow_reuse=False)[source]#

Registers parameters in an outer transform without adding a name scope.

Functionally this is equivalent to lift()but without automatically adding an additional variable scoping. Note that closing over a module from an outer scope is disallowed.

See lift()for more context on when to use lift.

Parameters

init_fn (Callable[..., hk.Params]) – The init function from an Transformed.
allow_reuse (bool) – Allows lifted parameters to be reused from the outer transform_with_state(). This can be desirable when e.g. within control flow (e.g. hk.scan).

Return type

Callable[…, hk.Params]

Returns

See also

lift(): Register params with an outer transform.
lift_with_state(): Register params and state with an outer transform.
transparent_lift_with_state(): Register params and state with an outer transform without a namespace.

transparent_lift_with_state#

haiku.transparent_lift_with_state(init_fn, *, allow_reuse=False)[source]#

Registers params and state in an outer transform without adding scope.

Functionally this is equivalent to lift_with_state()but without automatically adding an additional variable scoping.

See lift_with_state()for more context on when to use lift_with_state.

Parameters

init_fn (Callable[..., tuple[hk.Params, hk.State]]) – The init function from an TransformedWithState.
allow_reuse (bool) – Allows lifted parameters and state to be reused from the outer transform_with_state(). This can be desirable when e.g. within control flow (e.g. hk.scan).

Return type

tuple[Callable[…, tuple[hk.Params, hk.State]], LiftWithStateUpdater]

Returns

See also

lift(): Register params with an outer transform.
lift_with_state(): Register params and state with an outer transform.
transparent_lift(): Register params with an outer transform without a namespace.

LiftWithStateUpdater#

class haiku.LiftWithStateUpdater(name)[source]#: Handles updating the state for a lift_with_state computation.

Layer Stack#

`layer_stack`(num_layers[, ...])	Utility to wrap a Haiku function and recursively apply it to an input.
`LayerStackTransparencyMapping`(args, *kwargs)	Module name mapping for transparent layer_stack.

layer_stack#

class haiku.layer_stack(num_layers, with_per_layer_inputs=False, unroll=1, pass_reverse_to_layer_fn=False, transparent=False, transparency_map=None, name=None)[source]#

Utility to wrap a Haiku function and recursively apply it to an input.

This can be used to improve model compile times.

A function is valid if it uses only explicit position parameters, and its return type matches its input type. The position parameters can be arbitrarily nested structures with jax.Array at the leaf nodes. Note that kwargs are not supported, neither are functions with variable number of parameters (specified by *args).

Note that layer_stack cannot at the moment be used with functions that build Haiku modules with state.

If with_per_layer_inputs=False then the new, wrapped function can be understood as performing the following:

>>> f = lambda x: x+1
>>> num_layers = 4
>>> x = 0
>>> for i in range(num_layers):
...   x = f(x)
>>> x
4

And if with_per_layer_inputs=True, assuming f takes two arguments on top of x:

>>> f = lambda x, y0, y1: (x+1, y0+y1)
>>> num_layers = 4
>>> x = 0
>>> ys_0 = [1, 2, 3, 4]
>>> ys_1 = [5, 6, 7, 8]
>>> zs = []
>>> for i in range(num_layers):
...   x, z = f(x, ys_0[i], ys_1[i])
...   zs.append(z)
>>> x, zs
(4, [6, 8, 10, 12])

The code using layer_stack for the above function would be:

>>> f = lambda x, y0, y1: (x+1, y0+y1)
>>> num_layers = 4
>>> x = 0
>>> ys_0 = jnp.array([1, 2, 3, 4])
>>> ys_1 = jnp.array([5, 6, 7, 8])
>>> stack = hk.layer_stack(num_layers, with_per_layer_inputs=True)
>>> x, zs = stack(f)(x, ys_0, ys_1)
>>> print(x, zs)
4 [ 6  8 10 12]

Check the tests in layer_stack_test.py for further examples.

Crucially, any parameters created inside f will not be shared across iterations.

Parameters

num_layers (int) – The number of times to iterate the wrapped function.
with_per_layer_inputs – Whether or not to pass per-layer inputs to the wrapped function.
unroll (int) – the unroll used by scan.
pass_reverse_to_layer_fn (bool) – Whether or not to pass the reverse keyword to the function f, so that it is aware if the layer stack is being run forward or in reverse (and the underlying scan). To run the layer stack in reverse you need to pass in reverse=True to the call to the layer stack.
transparent (bool) – Whether to apply layer_stack transparently. When this is True, and a correct transparency_map is provided, the parameters are generated in such a way that layer_stack can be replaced by a regular for loop without changing the parameter tree.
transparency_map (Optional[LayerStackTransparencyMapping]) – How to map stacked module names to flat names and reverse. See LayerStackTransparencyMapping and layer_stack_test.py for an example.
name (Optional[str]) – name of the Haiku context.

Returns

Callable that will produce a layer stack when called with a valid function.

LayerStackTransparencyMapping#

class haiku.LayerStackTransparencyMapping(*args, **kwargs)[source]#: Module name mapping for transparent layer_stack.

Naming#

`name_scope`(name, *[, method_name])	Context manager which adds a prefix to all new modules, params or state.
`current_name`()	Returns the currently active module name.
`DO_NOT_STORE`	Causes a parameter or state value to not be stored.
`get_params`()	Returns the parameters for the current `transform()`.
`get_current_state`()	Returns the current state for the current `transform_with_state()`.
`get_initial_state`()	Returns the initial state for the current `transform_with_state()`.
`force_name`(name)	Forces Haiku to use this name, ignoring all context information.
`name_like`(method_name)	Allows a method to be named like some other method.
`transparent`(method)	Decorator to wrap a method, preventing automatic variable scope wrapping.

name_scope#

haiku.name_scope(name, *, method_name='__call__')[source]#

Context manager which adds a prefix to all new modules, params or state.

>>> with hk.name_scope("my_name_scope"):
...   net = hk.Linear(1, name="my_linear")
>>> net.module_name
'my_name_scope/my_linear'

When used inside a module, any submodules, parameters or state created inside the name scope will have a prefix added to their names:

>>> class MyModule(hk.Module):
...   def __call__(self, x):
...     with hk.name_scope("my_name_scope"):
...       submodule = hk.Linear(1, name="submodule")
...       w = hk.get_parameter("w", [], init=jnp.ones)
...     return submodule(x) + w

>>> f = hk.transform(lambda x: MyModule()(x))
>>> params = f.init(jax.random.PRNGKey(42), jnp.ones([1, 1]))
>>> jax.tree_util.tree_map(jnp.shape, params)
{'my_module/my_name_scope': {'w': ()},
 'my_module/my_name_scope/submodule': {'b': (1,), 'w': (1, 1)}}

Name scopes are very similar to putting all of the code inside the context manager inside a method on a Module with the name you provide. Behind the scenes this is precisely how name scopes are implemented.

If you are familiar with TensorFlow then Haiku’s name_scope() is similar to tf.variable_scope(..) in TensorFlow 1 and tf.name_scope(..) in TensorFlow 1 and 2 in that it changes the names associated with modules, parameters and state.

Parameters

name (str) – The name scope to use (e.g. "foo" or "foo/bar").
method_name (str) – (Advanced uses only). Since name scopes are equivalent to calling methods on modules the method name attribute allows you to specify which method name you want to simulate. Most users should leave this as the default value (“__call__”).

Return type

ContextManager[None]

Returns

A single use context manager that when active prefixes new modules, parameters or state with the given name.

current_name#

haiku.current_name()[source]#

Returns the currently active module name.

Outside of a Haiku module (but inside a Haiku transform) this will return ~ which matches the key in the params/state dict where top level values are stored.

>>> hk.current_name()
'~'

Inside a module this returns the current module name:

>>> class ExampleModule(hk.Module):
...   def __call__(self):
...     return hk.current_name()
>>> ExampleModule()()
'example_module'

Inside a name scope this returns the current name scope:

>>> with hk.name_scope('example_name_scope'):
...   print(hk.current_name())
example_name_scope

Return type: str
Returns: The currently active module or name scope name. If modules or name scopes are in use returns ~.

DO_NOT_STORE#

haiku.DO_NOT_STORE = <haiku._src.base.DoNotStore object>#

Causes a parameter or state value to not be stored.

By default, Haiku will put the value returned from get_parameter(), get_state() and set_state() into the dictionaries returned by init. This is not always desirable.

For example, a user may want to have part of their network come from a pretrained checkpoint, and they may want to freeze those values (aka. have them not appear in the params dict passed later to grad). You can achieve this by manipulating the params dict, however sometimes it is more convenient to do this using custom creators/getters/setters.

Consider the following function:

>>> def f(x):
...   x = hk.Linear(300, name='torso')(x)
...   x = hk.Linear(10, name='tail')(x)
...   return x

Imagine you have a pre-trained set of weights for the torso:

>>> pretrained = {'torso': {'w': jnp.ones([28 * 28, 300]),
...                         'b': jnp.ones([300])}}

First we define a creator, that tells Haiku to not store any parameters that are part of the pretrained dict:

>>> def my_creator(next_creator, shape, dtype, init, context):
...   if context.module_name in pretrained:
...     return hk.DO_NOT_STORE
...   return next_creator(shape, dtype, init)

Then we need a getter that provides the parameter value from the pretrained dict:

>>> def my_getter(next_getter, value, context):
...   if context.module_name in pretrained:
...     assert value is hk.DO_NOT_STORE
...     value = pretrained[context.module_name][context.name]
...   return next_getter(value)

Finally we’ll wrap our function in context managers activating our creator and getter:

>>> def f_with_pretrained_torso(x):
...   with hk.custom_creator(my_creator), \
...        hk.custom_getter(my_getter):
...     return f(x)

You can see that when we run our function we only get parameters from modules that were not in the pretrained dict:

>>> f_with_pretrained_torso = hk.transform(f_with_pretrained_torso)
>>> rng = jax.random.PRNGKey(42)
>>> x = jnp.ones([1, 28 * 28])
>>> params = f_with_pretrained_torso.init(rng, x)
>>> assert list(params) == ['tail']

This value can be used in initialisers, custom_creator() or custom_setter().

get_params#

haiku.get_params()[source]#

Returns the parameters for the current transform().

>>> def report(when):
...   shapes = jax.tree_util.tree_map(jnp.shape, hk.get_params())
...   print(f'{when}: {shapes}')
>>> def f(x):
...   report('Before call')
...   x = hk.Linear(1)(x)
...   report('After call')
...   return x
>>> f = hk.transform(f)
>>> rng = jax.random.PRNGKey(42)
>>> x = jnp.ones([1, 1])

During init the parameters dictionary will get populated as modules are called:

>>> params = f.init(rng, x)
Before call: {}
After call: {'linear': {'b': (1,), 'w': (1, 1)}}

During apply the parameters dictionary will remain unchanged:

>>> _ = f.apply(params, None, x)
Before call: {'linear': {'b': (1,), 'w': (1, 1)}}
After call: {'linear': {'b': (1,), 'w': (1, 1)}}

NOTE: Does not run custom_getters() or parameter initializers.

Return type: Params
Returns: A copy of the parameters dictionary. During init this dictionary will be populated with any parameters that have been created so far. During apply this will contain all parameters of all modules (the params dict does not change during apply).

See also

get_initial_state(): The initial state for the function.
get_current_state(): The current state for the function.

get_current_state#

haiku.get_current_state()[source]#

Returns the current state for the current transform_with_state().

Example:

>>> def report(when):
...   state = jax.tree_util.tree_map(int, hk.get_current_state())
...   print(f'{when}: {state}')
>>> def f():
...   report('Before get_state')
...   x = hk.get_state('x', [], init=jnp.zeros)
...   report('After get_state')
...   hk.set_state('x', x + 1)
...   report('After set_state')
>>> f = hk.transform_with_state(f)

During init, the most recently set value (either directly via set_state() or via the init argument to get_state()) will be returned:

>>> _, state = f.init(None)
Before get_state: {}
After get_state: {'~': {'x': 0}}
After set_state: {'~': {'x': 1}}

During apply the most recently set value will be used, if no value has been set then the value that is passed into apply will be used:

>>> state = {'~': {'x': 10}}
>>> _ = f.apply({}, state, None)
Before get_state: {'~': {'x': 10}}
After get_state: {'~': {'x': 10}}
After set_state: {'~': {'x': 11}}

NOTE: Does not run custom_getters() or state initializers.

Return type: State
Returns: A copy of the state dictionary that would be returned from init or apply.

See also

get_params(): The current parameters for the function.
get_initial_state(): The initial state for the function.

get_initial_state#

haiku.get_initial_state()[source]#

Returns the initial state for the current transform_with_state().

Example:

>>> def report(when):
...   state = jax.tree_util.tree_map(int, hk.get_initial_state())
...   print(f'{when}: {state}')
>>> def f():
...   report('Before get_state')
...   x = hk.get_state('x', [], init=jnp.zeros)
...   report('After get_state')
...   hk.set_state('x', x + 1)
...   report('After set_state')
>>> f = hk.transform_with_state(f)

During init, the first set value (either directly via set_state() or via the init argument to get_state()) will be returned:

>>> _, state = f.init(None)
Before get_state: {}
After get_state: {'~': {'x': 0}}
After set_state: {'~': {'x': 0}}

During apply the value passed into the apply function will be used:

>>> state = {'~': {'x': 10}}
>>> _ = f.apply({}, state, None)
Before get_state: {'~': {'x': 10}}
After get_state: {'~': {'x': 10}}
After set_state: {'~': {'x': 10}}

NOTE: Does not run custom_getters() or state initializers.

Return type: State
Returns: A copy of the state dictionary that would be returned from init or passed into apply.

See also

get_params(): The current parameters for the function.
get_current_state(): The current state for the function.

force_name#

haiku.force_name(name)[source]#

Forces Haiku to use this name, ignoring all context information.

NOTE: This method is intended for advanced use cases only and should be avoided whenever possible as it effectively enforces a singleton pattern when setting absolute names.

Haiku names modules according to where they are created (e.g. the stack of modules that created them, or the current name_scope()). This function allows you to create modules that ignore all of this and have precisely the name you provide.

This might be useful in the case that you have two modules and you want to force them to share parameters:

>>> mod0 = hk.Linear(1)
>>> some_hyperparameter = True
>>> if some_hyperparameter:
...   # Force mod1 and mod0 to have shared weights.
...   mod1 = hk.Linear(1, name=hk.force_name(mod0.module_name))
... else:
...   # mod0 and mod1 are independent.
...   mod1 = hk.Linear(1)

(A simpler version of this snippet would do mod1 = mod0 instead of using force_name, however in real examples it can be simpler to use force_name, especially in cases where you may not have access to the module instance without lots of plumbing, but getting the module name is easy [e.g. it is a hyperparameter]).

Parameters: name (str) – String name for the module. For example "foo" or "foo/bar".
Return type: str
Returns: A value suitable to pass into the name argument of any Haiku module constructor.

name_like#

haiku.name_like(method_name)[source]#

Allows a method to be named like some other method.

In Haiku submodules are named based on the name of their parent module and the method in which they are created. When refactoring code it may be desirable to maintain previous names in order to keep checkpoint compatibility, this can be achieved using name_like().

As an example, consider the following toy autoencoder:

>>> class Autoencoder(hk.Module):
...   def __call__(self, x):
...     z = hk.Linear(10, name="enc")(x)  # name: autoencoder/enc
...     y = hk.Linear(10, name="dec")(z)  # name: autoencoder/dec
...     return y

If we want to refactor this such that users can encode or decode, we would create two methods (encode, decode) which would create and apply our modules. In order to retain checkpoint compatibility with the original module we can use name_like() to name those submodules as if they were created inside __call__:

>>> class Autoencoder(hk.Module):
...   @hk.name_like("__call__")
...   def encode(self, x):
...     return hk.Linear(10, name="enc")(x)  # name: autoencoder/enc
...
...   @hk.name_like("__call__")
...   def decode(self, z):
...     return hk.Linear(10, name="dec")(z)  # name: autoencoder/dec
...
...   def __call__(self, x):
...     return self.decode(self.encode(x))

One sharp edge is if users rely on Haiku’s numbering to take care of giving unique names and refactor using name_like(). For example when refactoring the following:

>>> class Autoencoder(hk.Module):
...   def __call__(self, x):
...     y = hk.Linear(10)(z)  # name: autoencoder/linear_1
...     z = hk.Linear(10)(x)  # name: autoencoder/linear
...     return y

To use name_like(), the unnamed linear modules in encode/decode will end up with the same name (both: autoencoder/linear) because module numbering is only applied within a method:

>>> class Autoencoder(hk.Module):
...   @hk.name_like("__call__")
...   def encode(self, x):
...     return hk.Linear(10)(x)  # name: autoencoder/linear
...
...   @hk.name_like("__call__")
...   def decode(self, z):
...     return hk.Linear(10)(z)  # name: autoencoder/linear  <-- NOT INTENDED

To fix this case you need to explicitly name the modules within the method with their former name:

>>> class Autoencoder(hk.Module):
...   @hk.name_like("__call__")
...   def encode(self, x):
...     return hk.Linear(10, name="linear")(x)    # name: autoencoder/linear
...
...   @hk.name_like("__call__")
...   def decode(self, z):
...     return hk.Linear(10, name="linear_1")(z)  # name: autoencoder/linear_1

Parameters: method_name (str) – The name of a method whose name we should adopt. This method does not actually have to be defined on the class.
Return type: Callable[[T], T]
Returns: A decorator that when applied to a method marks it as having a different name.

transparent#

haiku.transparent(method)[source]#

Decorator to wrap a method, preventing automatic variable scope wrapping.

By default, all variables and modules created in a method are scoped by the module and method names. This is undesirable in some cases. Any method decorated with transparent() will create variables and modules in the scope in which it was called.

Parameters: method (T) – the method to wrap.
Return type: T
Returns: The method, with a flag indicating no name scope wrapping should occur.

Visualisation#

to_dot(fun)

Converts a function using Haiku modules to a dot graph.

to_dot#

haiku.to_dot(fun)[source]#

Converts a function using Haiku modules to a dot graph.

To view the resulting graph in Google Colab or an iPython notebook use the graphviz package:

dot = hk.to_dot(f)(x)
import graphviz
graphviz.Source(dot)

Parameters: fun (Callable[..., Any]) – A function using Haiku modules.
Return type: Callable[…, str]
Returns: A function that returns the source code string to a graphviz graph describing the operations executed by the given function clustered by Haiku module.

See also

abstract_to_dot(): Generates a graphviz graph using abstract inputs.

Common Modules#

Linear#

`Linear`(output_size[, with_bias, w_init, ...])	Linear module.
`Bias`([output_size, bias_dims, b_init, name])	Adds a bias to inputs.

Linear#

class haiku.Linear(output_size, with_bias=True, w_init=None, b_init=None, name=None)[source]#

Linear module.

__init__(output_size, with_bias=True, w_init=None, b_init=None, name=None)[source]#

Constructs the Linear module.

Parameters

output_size (int) – Output dimensionality.
with_bias (bool) – Whether to add a bias to the output.
w_init (Optional[hk.initializers.Initializer]) – Optional initializer for weights. By default, uses random values from truncated normal, with stddev 1 / sqrt(fan_in). See https://arxiv.org/abs/1502.03167v3.
b_init (Optional[hk.initializers.Initializer]) – Optional initializer for bias. By default, zero.
name (Optional[str]) – Name of the module.

__call__(inputs, *, precision=None)[source]#

Computes a linear transform of the input.

Return type: jax.Array

Bias#

class haiku.Bias(output_size=None, bias_dims=None, b_init=None, name=None)[source]#

Adds a bias to inputs.

Example Usage:

>>> N, H, W, C = 1, 2, 3, 4
>>> x = jnp.ones([N, H, W, C])
>>> scalar_bias = hk.Bias(bias_dims=[])
>>> scalar_bias_output = scalar_bias(x)
>>> assert scalar_bias.bias_shape == ()

Create a bias over all non-minibatch dimensions:

>>> all_bias = hk.Bias()
>>> all_bias_output = all_bias(x)
>>> assert all_bias.bias_shape == (H, W, C)

Create a bias over the last non-minibatch dimension:

>>> last_bias = hk.Bias(bias_dims=[-1])
>>> last_bias_output = last_bias(x)
>>> assert last_bias.bias_shape == (C,)

Create a bias over the first non-minibatch dimension:

>>> first_bias = hk.Bias(bias_dims=[1])
>>> first_bias_output = first_bias(x)
>>> assert first_bias.bias_shape == (H, 1, 1)

Subtract and later add the same learned bias:

>>> bias = hk.Bias()
>>> h1 = bias(x, multiplier=-1)
>>> h2 = bias(x)
>>> h3 = bias(x, multiplier=-1)
>>> reconstructed_x = bias(h3)
>>> assert (x == reconstructed_x).all()

__init__(output_size=None, bias_dims=None, b_init=None, name=None)[source]#

Constructs a Bias module that supports broadcasting.

Parameters

output_size (Optional[Sequence[int]]) – Output size (output shape without batch dimension). If output_size is left as None, the size will be directly inferred by the input.
bias_dims (Optional[Sequence[int]]) – Sequence of which dimensions to retain from the input shape when constructing the bias. The remaining dimensions will be broadcast over (given size of 1), and leading dimensions will be removed completely. See class doc for examples.
b_init (Optional[hk.initializers.Initializer]) – Optional initializer for the bias. Default to zeros.
name (Optional[str]) – Name of the module.

__call__(inputs, multiplier=None)[source]#

Adds bias to inputs and optionally multiplies by multiplier.

Parameters

inputs (jax.Array) – A Tensor of size [batch_size, input_size1, ...].
multiplier (Optional[Union[float, jax.Array]]) – A scalar or Tensor which the bias term is multiplied by before adding it to inputs. Anything which works in the expression bias * multiplier is acceptable here. This may be useful if you want to add a bias in one place and subtract the same bias in another place via multiplier=-1.

Return type

jax.Array

Returns

A Tensor of size [batch_size, input_size1, ...].

Pooling#

`avg_pool`(value, window_shape, strides, padding)	Average pool.
`AvgPool`(window_shape, strides, padding[, ...])	Average pool.
`max_pool`(value, window_shape, strides, padding)	Max pool.
`MaxPool`(window_shape, strides, padding[, ...])	Max pool.

Average Pool#

haiku.avg_pool(value, window_shape, strides, padding, channel_axis=- 1)[source]#

Average pool.

Parameters

value (jax.Array) – Value to pool.
window_shape (Union[int, Sequence[int]]) – Shape of the pooling window, same rank as value.
strides (Union[int, Sequence[int]]) – Strides of the pooling window, same rank as value.
padding (str) – Padding algorithm. Either VALID or SAME.
channel_axis (Optional[int]) – Axis of the spatial channels for which pooling is skipped.

Return type

jax.Array

Returns

Pooled result. Same rank as value.

Raises

ValueError – If the padding is not valid.

class haiku.AvgPool(window_shape, strides, padding, channel_axis=- 1, name=None)[source]#

Average pool.

Equivalent to partial application of avg_pool().

__init__(window_shape, strides, padding, channel_axis=- 1, name=None)[source]#

Average pool.

Parameters

window_shape (Union[int, Sequence[int]]) – Shape of the pooling window, same rank as value.
strides (Union[int, Sequence[int]]) – Strides of the pooling window, same rank as value.
padding (str) – Padding algorithm. Either VALID or SAME.
channel_axis (Optional[int]) – Axis of the spatial channels for which pooling is skipped.
name (Optional[str]) – String name for the module.

__call__(value)[source]#

Call self as a function.

Return type: jax.Array

Max Pool#

haiku.max_pool(value, window_shape, strides, padding, channel_axis=- 1)[source]#

Max pool.

Parameters

value (jax.Array) – Value to pool.
window_shape (Union[int, Sequence[int]]) – Shape of the pooling window, same rank as value.
strides (Union[int, Sequence[int]]) – Strides of the pooling window, same rank as value.
padding (str) – Padding algorithm. Either VALID or SAME.
channel_axis (Optional[int]) – Axis of the spatial channels for which pooling is skipped.

Return type

jax.Array

Returns

Pooled result. Same rank as value.

class haiku.MaxPool(window_shape, strides, padding, channel_axis=- 1, name=None)[source]#

Max pool.

Equivalent to partial application of max_pool().

__init__(window_shape, strides, padding, channel_axis=- 1, name=None)[source]#

Max pool.

Parameters

window_shape (Union[int, Sequence[int]]) – Shape of the pooling window, same rank as value.
strides (Union[int, Sequence[int]]) – Strides of the pooling window, same rank as value.
padding (str) – Padding algorithm. Either VALID or SAME.
channel_axis (Optional[int]) – Axis of the spatial channels for which pooling is skipped.
name (Optional[str]) – String name for the module.

__call__(value)[source]#

Call self as a function.

Return type: jax.Array

Dropout#

dropout(rng, rate, x[, broadcast_dims])

Randomly drop units in the input at a given rate.

dropout#

haiku.dropout(rng, rate, x, broadcast_dims=())[source]#

Randomly drop units in the input at a given rate.

See: http://www.cs.toronto.edu/~hinton/absps/dropout.pdf

Parameters

rng (PRNGKey) – A JAX random key.
rate (float) – Probability that each element of x is discarded. Must be a scalar in the range [0, 1).
x (jax.Array) – The value to be dropped out.
broadcast_dims (Sequence[int]) – specifies dimensions that will share the same dropout mask.

Return type

jax.Array

Returns

x, but dropped out and scaled by 1 / (1 - rate).

Note

This involves generating x.size pseudo-random samples from U([0, 1)) computed with the full precision required to compare them with rate. When rate is a Python float, this is typically 32 bits, which is often more than what applications require. A work-around is to pass rate with a lower precision, e.g. using np.float16(rate).

Combinator#

Sequential(layers[, name])

Sequentially calls the given list of layers.

Sequential#

class haiku.Sequential(layers, name=None)[source]#

Sequentially calls the given list of layers.

Note that Sequential is limited in the range of possible architectures it can handle. This is a deliberate design decision; Sequential is only meant to be used for the simple case of fusing together modules/ops where the input of a particular module/op is the output of the previous one.

Another restriction is that it is not possible to have extra arguments in the __call__() method that are passed to the constituents of the module - for example, if there is a BatchNorm module in Sequential and the user wishes to switch the is_training flag. If this is the desired use case, the recommended solution is to subclass Module and implement __call__:

>>> class CustomModule(hk.Module):
...   def __call__(self, x, is_training):
...     x = hk.Conv2D(32, 4, 2)(x)
...     x = hk.BatchNorm(True, True, 0.9)(x, is_training)
...     x = jax.nn.relu(x)
...     return x

__init__(layers, name=None)[source]#

Initializes the current module with the given name.

Subclasses should call this constructor before creating other modules or variables such that those modules are named correctly.

Parameters: name (Optional[str]) – An optional string name for the class. Must be a valid Python identifier. If name is not provided then the class name for the current instance is converted to lower_snake_case and used instead.

__call__(inputs, *args, **kwargs)[source]#: Calls all layers sequentially.

Convolutional#

`ConvND`(num_spatial_dims, output_channels, ...)	General N-dimensional convolutional.
`Conv1D`(output_channels, kernel_shape[, ...])	One dimensional convolution.
`Conv2D`(output_channels, kernel_shape[, ...])	Two dimensional convolution.
`Conv3D`(output_channels, kernel_shape[, ...])	Three dimensional convolution.
`ConvNDTranspose`(num_spatial_dims, ...[, ...])	General n-dimensional transposed convolution (aka.
`Conv1DTranspose`(output_channels, kernel_shape)	One dimensional transposed convolution (aka.
`Conv2DTranspose`(output_channels, kernel_shape)	Two dimensional transposed convolution (aka.
`Conv3DTranspose`(output_channels, kernel_shape)	Three dimensional transposed convolution (aka.
`DepthwiseConv1D`(channel_multiplier, kernel_shape)	One dimensional convolution.
`DepthwiseConv2D`(channel_multiplier, kernel_shape)	Two dimensional convolution.
`DepthwiseConv3D`(channel_multiplier, kernel_shape)	Three dimensional convolution.
`get_channel_index`(data_format)	Returns the channel index when given a valid data format.

ConvND#

class haiku.ConvND(num_spatial_dims, output_channels, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='channels_last', mask=None, feature_group_count=1, name=None)[source]#

General N-dimensional convolutional.

__init__(num_spatial_dims, output_channels, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='channels_last', mask=None, feature_group_count=1, name=None)[source]#

Initializes the module.

Parameters

num_spatial_dims (int) – The number of spatial dimensions of the input.
output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length num_spatial_dims.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length num_spatial_dims. Defaults to 1.
rate (Union[int, Sequence[int]]) – Optional kernel dilation rate. Either an integer or a sequence of length num_spatial_dims. 1 corresponds to standard ND convolution, rate > 1 corresponds to dilated convolution. Defaults to 1.
padding (Union[str, Sequence[tuple[int, int]], hk.pad.PadFn, Sequence[hk.pad.PadFn]]) – Optional padding algorithm. Either VALID or SAME or a sequence of n (low, high) integer pairs that give the padding to apply before and after each spatial dimension. or a callable or sequence of callables of size num_spatial_dims. Any callables must take a single integer argument equal to the effective kernel size and return a sequence of two integers representing the padding before and after. See haiku.pad.* for more details and example functions. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Can be either channels_first, channels_last, N...C or NC.... By default, channels_last. See get_channel_index().
mask (Optional[jax.Array]) – Optional mask of the weights.
feature_group_count (int) – Optional number of groups in group convolution. Default value of 1 corresponds to normal dense convolution. If a higher value is used, convolutions are applied separately to that many groups, then stacked together. This reduces the number of parameters and possibly the compute for a given output_channels. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
name (Optional[str]) – The name of the module.

__call__(inputs, *, precision=None)[source]#

Connects ConvND layer.

Parameters

inputs (jax.Array) – An array of shape [spatial_dims, C] and rank-N+1 if unbatched, or an array of shape [N, spatial_dims, C] and rank-N+2 if batched.
precision (Optional[lax.Precision]) – Optional jax.lax.Precision to pass to jax.lax.conv_general_dilated().

Return type

jax.Array

Returns

An array of shape [spatial_dims, output_channels] and rank-N+1 if: unbatched, or an array of shape [N, spatial_dims, output_channels] and rank-N+2 if batched.

Conv1D#

class haiku.Conv1D(output_channels, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NWC', mask=None, feature_group_count=1, name=None)[source]#

One dimensional convolution.

__init__(output_channels, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NWC', mask=None, feature_group_count=1, name=None)[source]#

Initializes the module.

Parameters

output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length 1.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length 1. Defaults to 1.
rate (Union[int, Sequence[int]]) – Optional kernel dilation rate. Either an integer or a sequence of length 1. 1 corresponds to standard ND convolution, rate > 1 corresponds to dilated convolution. Defaults to 1.
padding (Union[str, Sequence[tuple[int, int]], hk.pad.PadFn, Sequence[hk.pad.PadFn]]) – Optional padding algorithm. Either VALID or SAME or a callable or sequence of callables of length 1. Any callables must take a single integer argument equal to the effective kernel size and return a list of two integers representing the padding before and after. See haiku.pad.* for more details and example functions. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Either NWC or NCW. By default, NWC.
mask (Optional[jax.Array]) – Optional mask of the weights.
feature_group_count (int) – Optional number of groups in group convolution. Default value of 1 corresponds to normal dense convolution. If a higher value is used, convolutions are applied separately to that many groups, then stacked together. This reduces the number of parameters and possibly the compute for a given output_channels. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
name (Optional[str]) – The name of the module.

Conv2D#

class haiku.Conv2D(output_channels, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NHWC', mask=None, feature_group_count=1, name=None)[source]#

Two dimensional convolution.

__init__(output_channels, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NHWC', mask=None, feature_group_count=1, name=None)[source]#

Initializes the module.

Parameters

output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length 2.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length 2. Defaults to 1.
rate (Union[int, Sequence[int]]) – Optional kernel dilation rate. Either an integer or a sequence of length 2. 1 corresponds to standard ND convolution, rate > 1 corresponds to dilated convolution. Defaults to 1.
padding (Union[str, Sequence[tuple[int, int]], hk.pad.PadFn, Sequence[hk.pad.PadFn]]) – Optional padding algorithm. Either VALID or SAME or a callable or sequence of callables of length 2. Any callables must take a single integer argument equal to the effective kernel size and return a list of two integers representing the padding before and after. See haiku.pad.* for more details and example functions. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Either NHWC or NCHW. By default, NHWC.
mask (Optional[jax.Array]) – Optional mask of the weights.
feature_group_count (int) – Optional number of groups in group convolution. Default value of 1 corresponds to normal dense convolution. If a higher value is used, convolutions are applied separately to that many groups, then stacked together. This reduces the number of parameters and possibly the compute for a given output_channels. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
name (Optional[str]) – The name of the module.

Conv3D#

class haiku.Conv3D(output_channels, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NDHWC', mask=None, feature_group_count=1, name=None)[source]#

Three dimensional convolution.

__init__(output_channels, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NDHWC', mask=None, feature_group_count=1, name=None)[source]#

Initializes the module.

Parameters

output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length 3.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length 3. Defaults to 1.
rate (Union[int, Sequence[int]]) – Optional kernel dilation rate. Either an integer or a sequence of length 3. 1 corresponds to standard ND convolution, rate > 1 corresponds to dilated convolution. Defaults to 1.
padding (Union[str, Sequence[tuple[int, int]], hk.pad.PadFn, Sequence[hk.pad.PadFn]]) – Optional padding algorithm. Either VALID or SAME or a callable or sequence of callables of length 3. Any callables must take a single integer argument equal to the effective kernel size and return a list of two integers representing the padding before and after. See haiku.pad.* for more details and example functions. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Either NDHWC or NCDHW. By default, NDHWC.
mask (Optional[jax.Array]) – Optional mask of the weights.
feature_group_count (int) – Optional number of groups in group convolution. Default value of 1 corresponds to normal dense convolution. If a higher value is used, convolutions are applied separately to that many groups, then stacked together. This reduces the number of parameters and possibly the compute for a given output_channels. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
name (Optional[str]) – The name of the module.

ConvNDTranspose#

class haiku.ConvNDTranspose(num_spatial_dims, output_channels, kernel_shape, stride=1, output_shape=None, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='channels_last', mask=None, name=None)[source]#

General n-dimensional transposed convolution (aka. deconvolution).

__init__(num_spatial_dims, output_channels, kernel_shape, stride=1, output_shape=None, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='channels_last', mask=None, name=None)[source]#

Initializes the module.

Parameters

num_spatial_dims (int) – The number of spatial dimensions of the input.
output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length num_spatial_dims.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length num_spatial_dims. Defaults to 1.
output_shape (Optional[Union[int, Sequence[int]]]) – Output shape of the spatial dimensions of a transpose convolution. Can be either an integer or an iterable of integers. If a None value is given, a default shape is automatically calculated.
padding (Union[str, Sequence[tuple[int, int]]]) – Optional padding algorithm. Either “VALID” or “SAME”. Defaults to “SAME”. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Can be either channels_first, channels_last, N...C or NC.... By default, channels_last.
mask (Optional[jax.Array]) – Optional mask of the weights.
name (Optional[str]) – The name of the module.

__call__(inputs, *, precision=None)[source]#

Computes the transposed convolution of the input.

Parameters

inputs (jax.Array) – An array of shape [spatial_dims, C] and rank-N+1 if unbatched, or an array of shape [N, spatial_dims, C] and rank-N+2 if batched.
precision (Optional[lax.Precision]) – Optional jax.lax.Precision to pass to jax.lax.conv_transpose().

Return type

jax.Array

Returns

An array of shape [spatial_dims, output_channels] and rank-N+1 if: unbatched, or an array of shape [N, spatial_dims, output_channels] and rank-N+2 if batched.

Conv1DTranspose#

class haiku.Conv1DTranspose(output_channels, kernel_shape, stride=1, output_shape=None, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NWC', mask=None, name=None)[source]#

One dimensional transposed convolution (aka. deconvolution).

__init__(output_channels, kernel_shape, stride=1, output_shape=None, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NWC', mask=None, name=None)[source]#

Initializes the module.

Parameters

output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length 1.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length 1. Defaults to 1.
output_shape (Optional[Union[int, Sequence[int]]]) – Output shape of the spatial dimensions of a transpose convolution. Can be either an integer or an iterable of integers. If a None value is given, a default shape is automatically calculated.
padding (Union[str, Sequence[tuple[int, int]]]) – Optional padding algorithm. Either VALID or SAME. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Either NWC or NCW. By default, NWC.
mask (Optional[jax.Array]) – Optional mask of the weights.
name (Optional[str]) – The name of the module.

Conv2DTranspose#

class haiku.Conv2DTranspose(output_channels, kernel_shape, stride=1, output_shape=None, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NHWC', mask=None, name=None)[source]#

Two dimensional transposed convolution (aka. deconvolution).

Initializes the module.

Parameters

output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length 2.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length 2. Defaults to 1.
output_shape (Optional[Union[int, Sequence[int]]]) – Output shape of the spatial dimensions of a transpose convolution. Can be either an integer or an iterable of integers. If a None value is given, a default shape is automatically calculated.
padding (Union[str, Sequence[tuple[int, int]]]) – Optional padding algorithm. Either VALID or SAME. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Either NHWC or NCHW. By default, NHWC.
mask (Optional[jax.Array]) – Optional mask of the weights.
name (Optional[str]) – The name of the module.

Conv3DTranspose#

class haiku.Conv3DTranspose(output_channels, kernel_shape, stride=1, output_shape=None, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NDHWC', mask=None, name=None)[source]#

Three dimensional transposed convolution (aka. deconvolution).

Initializes the module.

Parameters

output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length 3.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length 3. Defaults to 1.
output_shape (Optional[Union[int, Sequence[int]]]) – Output shape of the spatial dimensions of a transpose convolution. Can be either an integer or an iterable of integers. If a None value is given, a default shape is automatically calculated.
padding (Union[str, Sequence[tuple[int, int]]]) – Optional padding algorithm. Either VALID or SAME. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Either NDHWC or NCDHW. By default, NDHWC.
mask (Optional[jax.Array]) – Optional mask of the weights.
name (Optional[str]) – The name of the module.

DepthwiseConv1D#

class haiku.DepthwiseConv1D(channel_multiplier, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NWC', name=None)[source]#

One dimensional convolution.

__init__(channel_multiplier, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NWC', name=None)[source]#

Construct a 1D Depthwise Convolution.

Parameters

channel_multiplier (int) – Multiplicity of output channels. To keep the number of output channels the same as the number of input channels, set 1.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length 1.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length 1. Defaults to 1.
rate (Union[int, Sequence[int]]) – Optional kernel dilation rate. Either an integer or a sequence of length 1. 1 corresponds to standard ND convolution, rate > 1 corresponds to dilated convolution. Defaults to 1.
padding (Union[str, Sequence[tuple[int, int]]]) – Optional padding algorithm. Either VALID, SAME or a sequence of before, after pairs. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Can be either channels_first, channels_last, N...C or NC.... By default, channels_last. See get_channel_index().
name (Optional[str]) – The name of the module.

DepthwiseConv2D#

class haiku.DepthwiseConv2D(channel_multiplier, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NHWC', name=None)[source]#

Two dimensional convolution.

__init__(channel_multiplier, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NHWC', name=None)[source]#

Construct a 2D Depthwise Convolution.

Parameters

channel_multiplier (int) – Multiplicity of output channels. To keep the number of output channels the same as the number of input channels, set 1.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length 2.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length 2. Defaults to 1.
rate (Union[int, Sequence[int]]) – Optional kernel dilation rate. Either an integer or a sequence of length 1. 1 corresponds to standard ND convolution, rate > 1 corresponds to dilated convolution. Defaults to 1.
padding (Union[str, Sequence[tuple[int, int]]]) – Optional padding algorithm. Either VALID, SAME or a sequence of before, after pairs. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Can be either channels_first, channels_last, N...C or NC.... By default, channels_last. See get_channel_index().
name (Optional[str]) – The name of the module.

DepthwiseConv3D#

class haiku.DepthwiseConv3D(channel_multiplier, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NDHWC', name=None)[source]#

Three dimensional convolution.

__init__(channel_multiplier, kernel_shape, stride=1, rate=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NDHWC', name=None)[source]#

Construct a 3D Depthwise Convolution.

Parameters

channel_multiplier (int) – Multiplicity of output channels. To keep the number of output channels the same as the number of input channels, set 1.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length 3.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length 3. Defaults to 1.
rate (Union[int, Sequence[int]]) – Optional kernel dilation rate. Either an integer or a sequence of length 1. 1 corresponds to standard ND convolution, rate > 1 corresponds to dilated convolution. Defaults to 1.
padding (Union[str, Sequence[tuple[int, int]]]) – Optional padding algorithm. Either VALID, SAME or a sequence of before, after pairs. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Can be either channels_first, channels_last, N...C or NC.... By default, channels_last. See get_channel_index().
name (Optional[str]) – The name of the module.

SeparableDepthwiseConv2D#

class haiku.SeparableDepthwiseConv2D(channel_multiplier, kernel_shape, stride=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NHWC', name=None)[source]#

Separable 2-D Depthwise Convolution Module.

__init__(channel_multiplier, kernel_shape, stride=1, padding='SAME', with_bias=True, w_init=None, b_init=None, data_format='NHWC', name=None)[source]#

Construct a Separable 2D Depthwise Convolution module.

Parameters

channel_multiplier (int) – Multiplicity of output channels. To keep the number of output channels the same as the number of input channels, set 1.
kernel_shape (Union[int, Sequence[int]]) – The shape of the kernel. Either an integer or a sequence of length num_spatial_dims.
stride (Union[int, Sequence[int]]) – Optional stride for the kernel. Either an integer or a sequence of length num_spatial_dims. Defaults to 1.
padding (Union[str, Sequence[tuple[int, int]]]) – Optional padding algorithm. Either VALID, SAME or a sequence of before, after pairs. Defaults to SAME. See: https://www.tensorflow.org/xla/operation_semantics#conv_convolution.
with_bias (bool) – Whether to add a bias. By default, true.
w_init (Optional[hk.initializers.Initializer]) – Optional weight initialization. By default, truncated normal.
b_init (Optional[hk.initializers.Initializer]) – Optional bias initialization. By default, zeros.
data_format (str) – The data format of the input. Can be either channels_first, channels_last, N...C or NC.... By default, channels_last.
name (Optional[str]) – The name of the module.

__call__(inputs)[source]#

Call self as a function.

Return type: jax.Array

get_channel_index#

haiku.get_channel_index(data_format)[source]#

Returns the channel index when given a valid data format.

>>> hk.get_channel_index('channels_last')
-1
>>> hk.get_channel_index('channels_first')
1
>>> hk.get_channel_index('N...C')
-1
>>> hk.get_channel_index('NCHW')
1

Parameters: data_format (str) – String, the data format to get the channel index from. Valid data formats are spatial (e.g.``NCHW``), sequential (e.g. BTHWD), channels_first and channels_last).
Return type: int
Returns: The channel index as an int, either 1 or -1.
Raises: ValueError – If the data format is unrecognised.

Normalization#

`BatchNorm`(create_scale, create_offset, ...)	Normalizes inputs to maintain a mean of ~0 and stddev of ~1.
`GroupNorm`(groups[, axis, create_scale, ...])	Group normalization module.
`InstanceNorm`(create_scale, create_offset[, ...])	Normalizes inputs along the spatial dimensions.
`LayerNorm`(axis, create_scale, create_offset)	LayerNorm module.
`RMSNorm`(axis[, eps, scale_init, name, ...])	RMSNorm module.
`SpectralNorm`([eps, n_steps, name])	Normalizes an input by its first singular value.
`ExponentialMovingAverage`(decay[, ...])	Maintains an exponential moving average.
`SNParamsTree`([eps, n_steps, ignore_regex, name])	Applies Spectral Normalization to all parameters in a tree.
`EMAParamsTree`(decay[, zero_debias, ...])	Maintains an exponential moving average for all parameters in a tree.

BatchNorm#

class haiku.BatchNorm(create_scale, create_offset, decay_rate, eps=1e-05, scale_init=None, offset_init=None, axis=None, cross_replica_axis=None, cross_replica_axis_index_groups=None, data_format='channels_last', name=None)[source]#

Normalizes inputs to maintain a mean of ~0 and stddev of ~1.

See: https://arxiv.org/abs/1502.03167.

There are many different variations for how users want to manage scale and offset if they require them at all. These are:

No scale/offset in which case create_* should be set to False and scale/offset aren’t passed when the module is called.

Trainable scale/offset in which case create_* should be set to True and again scale/offset aren’t passed when the module is called. In this case this module creates and owns the scale/offset variables.

Externally generated scale/offset, such as for conditional normalization, in which case create_* should be set to False and then the values fed in at call time.

NOTE: jax.vmap(hk.transform(BatchNorm)) will update summary statistics and normalize values on a per-batch basis; we currently do not support normalizing across a batch axis introduced by vmap.

__init__(create_scale, create_offset, decay_rate, eps=1e-05, scale_init=None, offset_init=None, axis=None, cross_replica_axis=None, cross_replica_axis_index_groups=None, data_format='channels_last', name=None)[source]#

Constructs a BatchNorm module.

Parameters

create_scale (bool) – Whether to include a trainable scaling factor.
create_offset (bool) – Whether to include a trainable offset.
decay_rate (float) – Decay rate for EMA.
eps (float) – Small epsilon to avoid division by zero variance. Defaults 1e-5, as in the paper and Sonnet.
scale_init (Optional[hk.initializers.Initializer]) – Optional initializer for gain (aka scale). Can only be set if create_scale=True. By default, 1.
offset_init (Optional[hk.initializers.Initializer]) – Optional initializer for bias (aka offset). Can only be set if create_offset=True. By default, 0.
axis (Optional[Sequence[int]]) – Which axes to reduce over. The default (None) signifies that all but the channel axis should be normalized. Otherwise this is a list of axis indices which will have normalization statistics calculated.
cross_replica_axis (Optional[Union[str, Sequence[str]]]) – If not None, it should be a string (or sequence of strings) representing the axis name(s) over which this module is being run within a jax map (e.g. jax.pmap or jax.vmap). Supplying this argument means that batch statistics are calculated across all replicas on the named axes.
cross_replica_axis_index_groups (Optional[Sequence[Sequence[int]]]) – Specifies how devices are grouped. Valid only within jax.pmap collectives.
data_format (str) – The data format of the input. Can be either channels_first, channels_last, N...C or NC.... By default it is channels_last. See get_channel_index().
name (Optional[str]) – The module name.

__call__(inputs, is_training, test_local_stats=False, scale=None, offset=None)[source]#

Computes the normalized version of the input.

Parameters

inputs (jax.Array) – An array, where the data format is [..., C].
is_training (bool) – Whether this is during training.
test_local_stats (bool) – Whether local stats are used when is_training=False.
scale (Optional[jax.Array]) – An array up to n-D. The shape of this tensor must be broadcastable to the shape of inputs. This is the scale applied to the normalized inputs. This cannot be passed in if the module was constructed with create_scale=True.
offset (Optional[jax.Array]) – An array up to n-D. The shape of this tensor must be broadcastable to the shape of inputs. This is the offset applied to the normalized inputs. This cannot be passed in if the module was constructed with create_offset=True.

Return type

jax.Array

Returns

The array, normalized across all but the last dimension.

GroupNorm#

class haiku.GroupNorm(groups, axis=slice(1, None, None), create_scale=True, create_offset=True, eps=1e-05, scale_init=None, offset_init=None, data_format='channels_last', name=None)[source]#

Group normalization module.

This applies group normalization to the x. This involves splitting the channels into groups before calculating the mean and variance. The default behaviour is to compute the mean and variance over the spatial dimensions and the grouped channels. The mean and variance will never be computed over the created groups axis.

It transforms the input x into:

\[\d{outputs} = \d{scale} \dfrac{x - \mu}{\sigma + \epsilon} + \d{offset}\]

Where \(\mu\) and \(\sigma\) are respectively the mean and standard deviation of x.

There are many different variations for how users want to manage scale and offset if they require them at all. These are:

No scale/offset in which case create_* should be set to False and scale/offset aren’t passed when the module is called.

Trainable scale/offset in which case create_* should be set to True and again scale/offset aren’t passed when the module is called. In this case this module creates and owns the scale/offset parameters.

Externally generated scale/offset, such as for conditional normalization, in which case create_* should be set to False and then the values fed in at call time.

__init__(groups, axis=slice(1, None, None), create_scale=True, create_offset=True, eps=1e-05, scale_init=None, offset_init=None, data_format='channels_last', name=None)[source]#

Constructs a GroupNorm module.

Parameters

groups (int) – number of groups to divide the channels by. The number of channels must be divisible by this.
axis (Union[int, slice, Sequence[int]]) – int, slice or sequence of ints representing the axes which should be normalized across. By default this is all but the first dimension. For time series data use slice(2, None) to average over the none Batch and Time data.
create_scale (bool) – whether to create a trainable scale per channel applied after the normalization.
create_offset (bool) – whether to create a trainable offset per channel applied after normalization and scaling.
eps (float) – Small epsilon to add to the variance to avoid division by zero. Defaults to 1e-5.
scale_init (Optional[hk.initializers.Initializer]) – Optional initializer for the scale parameter. Can only be set if create_scale=True. By default scale is initialized to 1.
offset_init (Optional[hk.initializers.Initializer]) – Optional initializer for the offset parameter. Can only be set if create_offset=True. By default offset is initialized to 0.
data_format (str) – The data format of the input. Can be either channels_first, channels_last, N...C or NC.... By default it is channels_last. See get_channel_index().
name (Optional[str]) – Name of the module.

__call__(x, scale=None, offset=None)[source]#

Returns normalized inputs.

Parameters

x (jax.Array) – An n-D tensor of the data_format specified in the constructor on which the transformation is performed.
scale (Optional[jax.Array]) – A tensor up to n-D. The shape of this tensor must be broadcastable to the shape of x. This is the scale applied to the normalized x. This cannot be passed in if the module was constructed with create_scale=True.
offset (Optional[jax.Array]) – A tensor up to n-D. The shape of this tensor must be broadcastable to the shape of x. This is the offset applied to the normalized x. This cannot be passed in if the module was constructed with create_offset=True.

Return type

jax.Array

Returns

An n-d tensor of the same shape as x that has been normalized.

InstanceNorm#

class haiku.InstanceNorm(create_scale, create_offset, eps=1e-05, scale_init=None, offset_init=None, data_format='channels_last', name=None)[source]#

Normalizes inputs along the spatial dimensions.

See LayerNorm for more details.

__init__(create_scale, create_offset, eps=1e-05, scale_init=None, offset_init=None, data_format='channels_last', name=None)[source]#

Constructs an InstanceNorm module.

This method creates a module which normalizes over the spatial dimensions.

Parameters

create_scale (bool) – bool representing whether to create a trainable scale per channel applied after the normalization.
create_offset (bool) – bool representing whether to create a trainable offset per channel applied after normalization and scaling.
eps (float) – Small epsilon to avoid division by zero variance. Defaults to 1e-5.
scale_init (Optional[hk.initializers.Initializer]) – Optional initializer for the scale variable. Can only be set if create_scale=True. By default scale is initialized to 1.
offset_init (Optional[hk.initializers.Initializer]) – Optional initializer for the offset variable. Can only be set if create_offset=True. By default offset is initialized to 0.
data_format (str) – The data format of the input. Can be either channels_first, channels_last, N...C or NC.... By default it is channels_last. See get_channel_index().
name (Optional[str]) – Name of the module.

LayerNorm#

class haiku.LayerNorm(axis, create_scale, create_offset, eps=1e-05, scale_init=None, offset_init=None, use_fast_variance=False, name=None, *, param_axis=None)[source]#

LayerNorm module.

See: https://arxiv.org/abs/1607.06450.

Example usage:

>>> ln = hk.LayerNorm(axis=-1, param_axis=-1,
...                   create_scale=True, create_offset=True)
>>> x = ln(jnp.ones([8, 224, 224, 3]))

__init__(axis, create_scale, create_offset, eps=1e-05, scale_init=None, offset_init=None, use_fast_variance=False, name=None, *, param_axis=None)[source]#

Constructs a LayerNorm module.

Parameters

axis (AxisOrAxes) – Integer, list of integers, or slice indicating which axes to normalize over. Note that the shape of the scale/offset parameters are controlled by the param_axis argument.
create_scale (bool) – Bool, defines whether to create a trainable scale per channel applied after the normalization.
create_offset (bool) – Bool, defines whether to create a trainable offset per channel applied after normalization and scaling.
eps (float) – Small epsilon to avoid division by zero variance. Defaults 1e-5, as in the paper and Sonnet.
scale_init (Optional[hk.initializers.Initializer]) – Optional initializer for gain (aka scale). By default, one.
offset_init (Optional[hk.initializers.Initializer]) – Optional initializer for bias (aka offset). By default, zero.
use_fast_variance (bool) – If true, use a faster but less numerically stable formulation for computing variance.
name (Optional[str]) – The module name.
param_axis (Optional[AxisOrAxes]) – Axis used to determine the parameter shape of the learnable scale/offset. Sonnet sets this to the channel/feature axis (e.g. to -1 for NHWC). Other libraries set this to the same as the reduction axis (e.g. axis=param_axis).

__call__(inputs, scale=None, offset=None)[source]#

Connects the layer norm.

Parameters

inputs (jax.Array) – An array, where the data format is [N, ..., C].
scale (Optional[jax.Array]) – An array up to n-D. The shape of this tensor must be broadcastable to the shape of inputs. This is the scale applied to the normalized inputs. This cannot be passed in if the module was constructed with create_scale=True.
offset (Optional[jax.Array]) – An array up to n-D. The shape of this tensor must be broadcastable to the shape of inputs. This is the offset applied to the normalized inputs. This cannot be passed in if the module was constructed with create_offset=True.

Return type

jax.Array

Returns

The array, normalized.

RMSNorm#

class haiku.RMSNorm(axis, eps=1e-05, scale_init=None, name=None, create_scale=True, *, param_axis=None)[source]#

RMSNorm module.

RMSNorm provides an alternative that can be both faster and more stable than LayerNorm. The inputs are normalized by the root-mean-squared (RMS) and scaled by a learned parameter, but they are not recentered around their mean.

See https://arxiv.org/pdf/1910.07467.pdf

__init__(axis, eps=1e-05, scale_init=None, name=None, create_scale=True, *, param_axis=None)[source]#

Constructs a RMSNorm module.

Parameters

axis (AxisOrAxes) – Integer, list of integers, or slice indicating which axes to normalize over.
eps (float) – Small epsilon to avoid division by zero variance. Defaults to 1e-5.
scale_init (Optional[hk.initializers.Initializer]) – Optional initializer for gain (aka scale). By default, one.
name (Optional[str]) – The module name.
create_scale (bool) – Bool, defines whether to create a trainable scale per channel applied after the normalization.
param_axis (Optional[AxisOrAxes]) – Axis used to determine the parameter shape of the learnable scale/offset. Sonnet sets this to the channel/feature axis (e.g. to -1 for NHWC). Other libraries set this to the same as the reduction axis (e.g. axis=param_axis). None defaults to (-1,).

__call__(inputs)[source]#

Connects the layer norm.

Parameters: inputs (jax.Array) – An array, where the data format is [N, ..., C].
Returns: The normalized array, of the same shape as the inputs.

SpectralNorm#

class haiku.SpectralNorm(eps=0.0001, n_steps=1, name=None)[source]#

Normalizes an input by its first singular value.

This module uses power iteration to calculate this value based on the input and an internal hidden state.

__init__(eps=0.0001, n_steps=1, name=None)[source]#

Initializes an SpectralNorm module.

Parameters

eps (float) – The constant used for numerical stability.
n_steps (int) – How many steps of power iteration to perform to approximate the singular value of the input.
name (Optional[str]) – The name of the module.

__call__(value, update_stats=True, error_on_non_matrix=False)[source]#

Performs Spectral Normalization and returns the new value.

Parameters

value – The array-like object for which you would like to perform an spectral normalization on.
update_stats (bool) – A boolean defaulting to True. Regardless of this arg, this function will return the normalized input. When update_stats is True, the internal state of this object will also be updated to reflect the input value. When update_stats is False the internal stats will remain unchanged.
error_on_non_matrix (bool) – Spectral normalization is only defined on matrices. By default, this module will return scalars unchanged and flatten higher-order tensors in their leading dimensions. Setting this flag to True will instead throw errors in those cases.

Return type

jax.Array

Returns

The input value normalized by it’s first singular value.

Raises

ValueError – If error_on_non_matrix is True and value has ndims > 2.

ExponentialMovingAverage#

class haiku.ExponentialMovingAverage(decay, zero_debias=True, warmup_length=0, name=None)[source]#

Maintains an exponential moving average.

This uses the Adam debiasing procedure. See https://arxiv.org/pdf/1412.6980.pdf for details.

__init__(decay, zero_debias=True, warmup_length=0, name=None)[source]#

Initializes an ExponentialMovingAverage module.

Parameters

decay – The chosen decay. Must in [0, 1). Values close to 1 result in slow decay; values close to 0 result in fast decay.
zero_debias (bool) – Whether to run with zero-debiasing.
warmup_length (int) – A positive integer, EMA has no effect until the internal counter has reached warmup_length at which point the initial value for the decaying average is initialized to the input value after warmup_length iterations.
name (Optional[str]) – The name of the module.

initialize(shape, dtype=<class 'jax.numpy.float32'>)[source]#: If uninitialized sets the average to zeros of the given shape/dtype.

__call__(value, update_stats=True)[source]#

Updates the EMA and returns the new value.

Parameters

value (Union[float, jax.Array]) – The array-like object for which you would like to perform an exponential decay on.
update_stats (bool) – A Boolean, whether to update the internal state of this object to reflect the input value. When update_stats is False the internal stats will remain unchanged.

Return type

jax.Array

Returns

The exponentially weighted average of the input value.

SNParamsTree#

class haiku.SNParamsTree(eps=0.0001, n_steps=1, ignore_regex='', name=None)[source]#

Applies Spectral Normalization to all parameters in a tree.

This is isomorphic to EMAParamsTree in moving_averages.py.

__init__(eps=0.0001, n_steps=1, ignore_regex='', name=None)[source]#

Initializes an SNParamsTree module.

Parameters

eps (float) – The constant used for numerical stability.
n_steps (int) – How many steps of power iteration to perform to approximate the singular value of the input.
ignore_regex (str) – A string. Any parameter in the tree whose name matches this regex will not have spectral normalization applied to it. The empty string means this module applies to all parameters.
name (Optional[str]) – The name of the module.

__call__(tree, update_stats=True)[source]#: Call self as a function.

EMAParamsTree#

class haiku.EMAParamsTree(decay, zero_debias=True, warmup_length=0, ignore_regex='', name=None)[source]#

Maintains an exponential moving average for all parameters in a tree.

While ExponentialMovingAverage is meant to be applied to single parameters within a function, this class is meant to be applied to the entire tree of parameters for a function.

Given a set of parameters for some network:

>>> network_fn = lambda x: hk.Linear(10)(x)
>>> x = jnp.ones([1, 1])
>>> params = hk.transform(network_fn).init(jax.random.PRNGKey(428), x)

You might use the EMAParamsTree like follows:

>>> ema_fn = hk.transform_with_state(lambda x: hk.EMAParamsTree(0.2)(x))
>>> _, ema_state = ema_fn.init(None, params)
>>> ema_params, ema_state = ema_fn.apply(None, ema_state, None, params)

Here, we are transforming a Haiku function and constructing its parameters via an init_fn as normal, but are creating a second transformed function which expects a tree of parameters as input. This function is then called with the current parameters as input, which then returns an identical tree with every parameter replaced with its exponentially decayed average. This ema_params object can then be passed into the network_fn as usual, and will cause it to run with EMA weights.

__init__(decay, zero_debias=True, warmup_length=0, ignore_regex='', name=None)[source]#

Initializes an EMAParamsTree module.

Parameters

decay – The chosen decay. Must in [0, 1). Values close to 1 result in slow decay; values close to 0 result in fast decay.
zero_debias (bool) – Whether to run with zero-debiasing.
warmup_length (int) – A positive integer, EMA has no effect until the internal counter has reached warmup_length at which point the initial value for the decaying average is initialized to the input value after warmup_length iterations.
ignore_regex (str) – A string. Any parameter in the tree whose name matches this regex will not have any moving average applied to it. The empty string means this module will EMA all parameters.
name (Optional[str]) – The name of the module.

__call__(tree, update_stats=True)[source]#: Call self as a function.

Recurrent#

`RNNCore`([name])	Base class for RNN cores.
`dynamic_unroll`(core, input_sequence, ...[, ...])	Performs a dynamic unroll of an RNN.
`static_unroll`(core, input_sequence, ...[, ...])	Performs a static unroll of an RNN.
`expand_apply`(f[, axis])	Wraps f to temporarily add a size-1 axis to its inputs.
`VanillaRNN`(hidden_size[, double_bias, name])	Basic fully-connected RNN core.
`LSTM`(hidden_size[, name])	Long short-term memory (LSTM) RNN core.
`GRU`(hidden_size[, w_i_init, w_h_init, ...])	Gated Recurrent Unit.
`DeepRNN`(layers[, name])	Wraps a sequence of cores and callables as a single core.
`deep_rnn_with_skip_connections`(layers[, name])	Constructs a `DeepRNN` with skip connections.
`ResetCore`(core[, name])	A wrapper for managing state resets during unrolls.
`IdentityCore`([name])	A recurrent core that forwards the inputs and an empty state.
`Conv1DLSTM`(input_shape, output_channels, ...)	1-D convolutional LSTM.
`Conv2DLSTM`(input_shape, output_channels, ...)	2-D convolutional LSTM.
`Conv3DLSTM`(input_shape, output_channels, ...)	3-D convolutional LSTM.

RNNCore#

class haiku.RNNCore(name=None)[source]#

Base class for RNN cores.

This class defines the basic functionality that every core should implement: initial_state(), used to construct an example of the core state; and __call__() which applies the core parameterized by a previous state to an input.

Cores may be used with dynamic_unroll() and static_unroll() to iteratively construct an output sequence from the given input sequence.

abstract __call__(inputs, prev_state)[source]#

Run one step of the RNN.

Parameters

inputs – An arbitrarily nested structure.
prev_state – Previous core state.

Return type

tuple[Any, Any]

Returns

A tuple with two elements output, next_state. output is an arbitrarily nested structure. next_state is the next core state, this must be the same shape as prev_state.

abstract initial_state(batch_size)[source]#

Constructs an initial state for this core.

Parameters: batch_size (Optional[int]) – Optional int or an integral scalar tensor representing batch size. If None, the core may either fail or (experimentally) return an initial state without a batch dimension.
Returns: Arbitrarily nested initial state for this core.

dynamic_unroll#

haiku.dynamic_unroll(core, input_sequence, initial_state, time_major=True, reverse=False, return_all_states=False, unroll=1)[source]#

Performs a dynamic unroll of an RNN.

An unroll corresponds to calling the core on each element of the input sequence in a loop, carrying the state through:

state = initial_state
for t in range(len(input_sequence)):
   outputs, state = core(input_sequence[t], state)

A dynamic unroll preserves the loop structure when executed inside jax.jit(). See static_unroll() for an unroll function which replaces a loop with its body repeated multiple times.

Parameters

core – An RNNCore to unroll.
input_sequence – An arbitrarily nested structure of tensors of shape [T, ...] if time-major=True, or [B, T, ...] if time_major=False, where T is the number of time steps.
initial_state – An initial state of the given core.
time_major – If True, inputs are expected time-major, otherwise they are expected batch-major.
reverse – If True, inputs are scanned in the reversed order. Equivalent to reversing the time dimension in both inputs and outputs. See https://jax.readthedocs.io/en/latest/_autosummary/jax.lax.scan.html for more details.
return_all_states – If True, all intermediate states are returned rather than only the last one in time.
unroll – How many scan iterations to unroll within a single iteration of a loop.

Returns

output_sequence - An arbitrarily nested structure of tensors of shape [T, ...] if time-major, otherwise [B, T, ...].
state_sequence - If return_all_states is True, returns the sequence of core states. Otherwise, core state at time step T.

Return type

A tuple with two elements

static_unroll#

haiku.static_unroll(core, input_sequence, initial_state, time_major=True)[source]#

Performs a static unroll of an RNN.

An unroll corresponds to calling the core on each element of the input sequence in a loop, carrying the state through:

state = initial_state
for t in range(len(input_sequence)):
   outputs, state = core(input_sequence[t], state)

A static unroll replaces a loop with its body repeated multiple times when executed inside jax.jit():

state = initial_state
outputs0, state = core(input_sequence[0], state)
outputs1, state = core(input_sequence[1], state)
outputs2, state = core(input_sequence[2], state)
...

See dynamic_unroll() for a loop-preserving unroll function.

Parameters

core – An RNNCore to unroll.
input_sequence – An arbitrarily nested structure of tensors of shape [T, ...] if time-major=True, or [B, T, ...] if time_major=False, where T is the number of time steps.
initial_state – An initial state of the given core.
time_major – If True, inputs are expected time-major, otherwise they are expected batch-major.

Returns

output_sequence - An arbitrarily nested structure of tensors of shape [T, ...] if time-major, otherwise [B, T, ...].
final_state - Core state at time step T.

Return type

A tuple with two elements

expand_apply#

haiku.expand_apply(f, axis=0)[source]#

Wraps f to temporarily add a size-1 axis to its inputs.

Syntactic sugar for:

ins = jax.tree_util.tree_map(lambda t: np.expand_dims(t, axis=axis), ins)
out = f(ins)
out = jax.tree_util.tree_map(lambda t: np.squeeze(t, axis=axis), out)

This may be useful for applying a function built for [Time, Batch, ...] arrays to a single timestep.

Parameters

f – The callable to be applied to the expanded inputs.
axis – Where to add the extra axis.

Returns

f, wrapped as described above.

VanillaRNN#

class haiku.VanillaRNN(hidden_size, double_bias=True, name=None)[source]#

Basic fully-connected RNN core.

Given \(x_t\) and the previous hidden state \(h_{t-1}\) the core computes

\[h_t = \operatorname{ReLU}(w_i x_t + b_i + w_h h_{t-1} + b_h)\]

The output is equal to the new state, \(h_t\).

__init__(hidden_size, double_bias=True, name=None)[source]#

Constructs a vanilla RNN core.

Parameters

hidden_size (int) – Hidden layer size.
double_bias (bool) – Whether to use a bias in the two linear layers. This changes nothing to the learning performance of the cell. However, doubling will create two sets of bias parameters rather than one.
name (Optional[str]) – Name of the module.

__call__(inputs, prev_state)[source]#

Run one step of the RNN.

Parameters

inputs – An arbitrarily nested structure.
prev_state – Previous core state.

Returns

A tuple with two elements output, next_state. output is an arbitrarily nested structure. next_state is the next core state, this must be the same shape as prev_state.

initial_state(batch_size)[source]#

Constructs an initial state for this core.

Parameters: batch_size (Optional[int]) – Optional int or an integral scalar tensor representing batch size. If None, the core may either fail or (experimentally) return an initial state without a batch dimension.
Returns: Arbitrarily nested initial state for this core.

LSTM#

class haiku.LSTM(hidden_size, name=None)[source]#

Long short-term memory (LSTM) RNN core.

The implementation is based on [1]. Given \(x_t\) and the previous state \((h_{t-1}, c_{t-1})\) the core computes

\[\begin{array}{ll} i_t = \sigma(W_{ii} x_t + W_{hi} h_{t-1} + b_i) \\ f_t = \sigma(W_{if} x_t + W_{hf} h_{t-1} + b_f) \\ g_t = \tanh(W_{ig} x_t + W_{hg} h_{t-1} + b_g) \\ o_t = \sigma(W_{io} x_t + W_{ho} h_{t-1} + b_o) \\ c_t = f_t c_{t-1} + i_t g_t \\ h_t = o_t \tanh(c_t) \end{array}\]

where \(i_t\), \(f_t\), \(o_t\) are input, forget and output gate activations, and \(g_t\) is a vector of cell updates.

The output is equal to the new hidden, \(h_t\).

Notes

Forget gate initialization:: Following [2] we add 1.0 to \(b_f\) after initialization in order to reduce the scale of forgetting in the beginning of the training.

__init__(hidden_size, name=None)[source]#

Constructs an LSTM.

Parameters

hidden_size (int) – Hidden layer size.
name (Optional[str]) – Name of the module.

__call__(inputs, prev_state)[source]#

Run one step of the RNN.

Parameters

inputs (jax.Array) – An arbitrarily nested structure.
prev_state (LSTMState) – Previous core state.

Return type

tuple[jax.Array, LSTMState]

Returns

A tuple with two elements output, next_state. output is an arbitrarily nested structure. next_state is the next core state, this must be the same shape as prev_state.

initial_state(batch_size)[source]#

Constructs an initial state for this core.

Parameters: batch_size (Optional[int]) – Optional int or an integral scalar tensor representing batch size. If None, the core may either fail or (experimentally) return an initial state without a batch dimension.
Return type: LSTMState
Returns: Arbitrarily nested initial state for this core.

GRU#

class haiku.GRU(hidden_size, w_i_init=None, w_h_init=None, b_init=None, name=None)[source]#

Gated Recurrent Unit.

The implementation is based on: https://arxiv.org/pdf/1412.3555v1.pdf with biases.

Given \(x_t\) and the previous state \(h_{t-1}\) the core computes

\[\begin{array}{ll} z_t &= \sigma(W_{iz} x_t + W_{hz} h_{t-1} + b_z) \\ r_t &= \sigma(W_{ir} x_t + W_{hr} h_{t-1} + b_r) \\ a_t &= \tanh(W_{ia} x_t + W_{ha} (r_t \bigodot h_{t-1}) + b_a) \\ h_t &= (1 - z_t) \bigodot h_{t-1} + z_t \bigodot a_t \end{array}\]

where \(z_t\) and \(r_t\) are reset and update gates.

The output is equal to the new hidden state, \(h_t\).

__init__(hidden_size, w_i_init=None, w_h_init=None, b_init=None, name=None)[source]#

Initializes the current module with the given name.

Subclasses should call this constructor before creating other modules or variables such that those modules are named correctly.

Parameters: name (Optional[str]) – An optional string name for the class. Must be a valid Python identifier. If name is not provided then the class name for the current instance is converted to lower_snake_case and used instead.

__call__(inputs, state)[source]#

Run one step of the RNN.

Parameters

inputs – An arbitrarily nested structure.
prev_state – Previous core state.

Returns

A tuple with two elements output, next_state. output is an arbitrarily nested structure. next_state is the next core state, this must be the same shape as prev_state.

initial_state(batch_size)[source]#

Constructs an initial state for this core.

Parameters: batch_size (Optional[int]) – Optional int or an integral scalar tensor representing batch size. If None, the core may either fail or (experimentally) return an initial state without a batch dimension.
Returns: Arbitrarily nested initial state for this core.

DeepRNN#

class haiku.DeepRNN(layers, name=None)[source]#

Wraps a sequence of cores and callables as a single core.

>>> deep_rnn = hk.DeepRNN([
...     hk.LSTM(hidden_size=4),
...     jax.nn.relu,
...     hk.LSTM(hidden_size=2),
... ])

The state of a DeepRNN is a tuple with one element per RNNCore. If no layers are RNNCores, the state is an empty tuple.

__init__(layers, name=None)[source]#

Initializes the current module with the given name.

Subclasses should call this constructor before creating other modules or variables such that those modules are named correctly.

Parameters: name (Optional[str]) – An optional string name for the class. Must be a valid Python identifier. If name is not provided then the class name for the current instance is converted to lower_snake_case and used instead.

haiku.deep_rnn_with_skip_connections(layers, name=None)[source]#

Constructs a DeepRNN with skip connections.

Skip connections alter the dependency structure within a DeepRNN. Specifically, input to the i-th layer (i > 0) is given by a concatenation of the core’s inputs and the outputs of the (i-1)-th layer.

The output of the DeepRNN is the concatenation of the outputs of all cores.

outputs0, ... = layers[0](inputs, ...)
outputs1, ... = layers[1](tf.concat([inputs, outputs0], axis=-1], ...)
outputs2, ... = layers[2](tf.concat([inputs, outputs1], axis=-1], ...)
...

Parameters

layers (Sequence[RNNCore]) – List of RNNCores.
name (Optional[str]) – Name of the module.

Return type

RNNCore

Returns

A _DeepRNN with skip connections.

Raises

ValueError – If any of the layers is not an RNNCore.

ResetCore#

class haiku.ResetCore(core, name=None)[source]#

A wrapper for managing state resets during unrolls.

When unrolling an RNNCore on a batch of inputs sequences it may be necessary to reset the core’s state at different timesteps for different elements of the batch. The ResetCore class enables this by taking a batch of should_reset booleans in addition to the batch of inputs, and conditionally resetting the core’s state for individual elements of the batch. You may also reset individual entries of the state by passing a should_reset nest compatible with the state structure.

__init__(core, name=None)[source]#

Initializes the current module with the given name.

Subclasses should call this constructor before creating other modules or variables such that those modules are named correctly.

Parameters: name (Optional[str]) – An optional string name for the class. Must be a valid Python identifier. If name is not provided then the class name for the current instance is converted to lower_snake_case and used instead.

__call__(inputs, state)[source]#

Run one step of the wrapped core, handling state reset.

Parameters

inputs – Tuple with two elements, inputs, should_reset, where should_reset is the signal used to reset the wrapped core’s state. should_reset can be either tensor or nest. If nest, should_reset must match the state structure, and its components’ shapes must be prefixes of the corresponding entries tensors’ shapes in the state nest. If tensor, supported shapes are all commom shape prefixes of the state component tensors, e.g. [batch_size].
state – Previous wrapped core state.

Returns

Tuple of the wrapped core’s output, next_state.

initial_state(batch_size)[source]#

Constructs an initial state for this core.

Parameters: batch_size (Optional[int]) – Optional int or an integral scalar tensor representing batch size. If None, the core may either fail or (experimentally) return an initial state without a batch dimension.
Returns: Arbitrarily nested initial state for this core.

IdentityCore#

class haiku.IdentityCore(name=None)[source]#

A recurrent core that forwards the inputs and an empty state.

This is commonly used when switching between recurrent and feedforward versions of a model while preserving the same interface.

__call__(inputs, state)[source]#

Run one step of the RNN.

Parameters

inputs – An arbitrarily nested structure.
prev_state – Previous core state.

Returns

A tuple with two elements output, next_state. output is an arbitrarily nested structure. next_state is the next core state, this must be the same shape as prev_state.

initial_state(batch_size)[source]#

Constructs an initial state for this core.

Parameters: batch_size (Optional[int]) – Optional int or an integral scalar tensor representing batch size. If None, the core may either fail or (experimentally) return an initial state without a batch dimension.
Returns: Arbitrarily nested initial state for this core.

Conv1DLSTM#

class haiku.Conv1DLSTM(input_shape, output_channels, kernel_shape, name=None)[source]#

1-D convolutional LSTM.

The implementation is based on [3]. Given \(x_t\) and the previous state \((h_{t-1}, c_{t-1})\) the core computes

\[\begin{array}{ll} i_t = \sigma(W_{ii} * x_t + W_{hi} * h_{t-1} + b_i) \\ f_t = \sigma(W_{if} * x_t + W_{hf} * h_{t-1} + b_f) \\ g_t = \tanh(W_{ig} * x_t + W_{hg} * h_{t-1} + b_g) \\ o_t = \sigma(W_{io} * x_t + W_{ho} * h_{t-1} + b_o) \\ c_t = f_t c_{t-1} + i_t g_t \\ h_t = o_t \tanh(c_t) \end{array}\]

where \(*\) denotes the convolution operator; \(i_t\), \(f_t\), \(o_t\) are input, forget and output gate activations, and \(g_t\) is a vector of cell updates.

The output is equal to the new hidden state, \(h_t\).

Notes

Forget gate initialization:: Following [2] we add 1.0 to \(b_f\) after initialization in order to reduce the scale of forgetting in the beginning of the training.

__init__(input_shape, output_channels, kernel_shape, name=None)[source]#

Constructs a 1-D convolutional LSTM.

Parameters

input_shape (Sequence[int]) – Shape of the inputs excluding batch size.
output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – Sequence of kernel sizes (of length 1), or an int. kernel_shape will be expanded to define a kernel size in all dimensions.
name (Optional[str]) – Name of the module.

Conv2DLSTM#

class haiku.Conv2DLSTM(input_shape, output_channels, kernel_shape, name=None)[source]#

2-D convolutional LSTM.

The implementation is based on [3]. Given \(x_t\) and the previous state \((h_{t-1}, c_{t-1})\) the core computes

where \(*\) denotes the convolution operator; \(i_t\), \(f_t\), \(o_t\) are input, forget and output gate activations, and \(g_t\) is a vector of cell updates.

The output is equal to the new hidden state, \(h_t\).

Notes

Forget gate initialization:: Following [2] we add 1.0 to \(b_f\) after initialization in order to reduce the scale of forgetting in the beginning of the training.

__init__(input_shape, output_channels, kernel_shape, name=None)[source]#

Constructs a 2-D convolutional LSTM.

Parameters

input_shape (Sequence[int]) – Shape of the inputs excluding batch size.
output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – Sequence of kernel sizes (of length 2), or an int. kernel_shape will be expanded to define a kernel size in all dimensions.
name (Optional[str]) – Name of the module.

Conv3DLSTM#

class haiku.Conv3DLSTM(input_shape, output_channels, kernel_shape, name=None)[source]#

3-D convolutional LSTM.

The implementation is based on [3]. Given \(x_t\) and the previous state \((h_{t-1}, c_{t-1})\) the core computes

where \(*\) denotes the convolution operator; \(i_t\), \(f_t\), \(o_t\) are input, forget and output gate activations, and \(g_t\) is a vector of cell updates.

The output is equal to the new hidden state, \(h_t\).

Notes

Forget gate initialization:: Following [2] we add 1.0 to \(b_f\) after initialization in order to reduce the scale of forgetting in the beginning of the training.

__init__(input_shape, output_channels, kernel_shape, name=None)[source]#

Constructs a 3-D convolutional LSTM.

Parameters

input_shape (Sequence[int]) – Shape of the inputs excluding batch size.
output_channels (int) – Number of output channels.
kernel_shape (Union[int, Sequence[int]]) – Sequence of kernel sizes (of length 3), or an int. kernel_shape will be expanded to define a kernel size in all dimensions.
name (Optional[str]) – Name of the module.

Attention#

MultiHeadAttention#

class haiku.MultiHeadAttention(num_heads, key_size, w_init_scale=None, *, w_init=None, with_bias=True, b_init=None, value_size=None, model_size=None, name=None)[source]#

Multi-headed attention (MHA) module.

This module is intended for attending over sequences of vectors.

Rough sketch: - Compute keys (K), queries (Q), and values (V) as projections of inputs. - Attention weights are computed as W = softmax(QK^T / sqrt(key_size)). - Output is another projection of WV^T.

For more detail, see the original Transformer paper:: “Attention is all you need” https://arxiv.org/abs/1706.03762.

Glossary of shapes: - T: Sequence length. - D: Vector (embedding) size. - H: Number of attention heads.

__init__(num_heads, key_size, w_init_scale=None, *, w_init=None, with_bias=True, b_init=None, value_size=None, model_size=None, name=None)[source]#

Initialises the module.

Parameters

num_heads (int) – Number of independent attention heads (H).
key_size (int) – The size of keys (K) and queries used for attention.
w_init_scale (Optional[float]) – DEPRECATED. Please use w_init instead.
w_init (Optional[hk.initializers.Initializer]) – Initialiser for weights in the linear map. Once w_init_scale is fully deprecated w_init will become mandatory. Until then it has a default value of None for backwards compatability.
with_bias (bool) – Whether to add a bias when computing various linear projections.
b_init (Optional[hk.initializers.Initializer]) – Optional initializer for bias. By default, zero.
value_size (Optional[int]) – Optional size of the value projection (V). If None, defaults to the key size (K).
model_size (Optional[int]) – Optional size of the output embedding (D’). If None, defaults to the key size multiplied by the number of heads (K * H).
name (Optional[str]) – Optional name for this module.

__call__(query, key, value, mask=None)[source]#

Computes (optionally masked) MHA with queries, keys & values.

This module broadcasts over zero or more ‘batch-like’ leading dimensions.

Parameters

query (jax.Array) – Embeddings sequence used to compute queries; shape […, T’, D_q].
key (jax.Array) – Embeddings sequence used to compute keys; shape […, T, D_k].
value (jax.Array) – Embeddings sequence used to compute values; shape […, T, D_v].
mask (Optional[jax.Array]) – Optional mask applied to attention weights; shape […, H=1, T’, T].

Return type

jax.Array

Returns

A new sequence of embeddings, consisting of a projection of the: attention-weighted value projections; shape […, T’, D’].

Batch#

`Reshape`(output_shape[, preserve_dims, name])	Reshapes input Tensor, preserving the batch dimension.
`Flatten`([preserve_dims, name])	Flattens the input, preserving the batch dimension(s).
`BatchApply`(f[, num_dims])	Temporarily merges leading dimensions of input tensors.

Reshape#

class haiku.Reshape(output_shape, preserve_dims=1, name=None)[source]#

Reshapes input Tensor, preserving the batch dimension.

For example, given an input tensor with shape [B, H, W, C, D]:

>>> B, H, W, C, D = range(1, 6)
>>> x = jnp.ones([B, H, W, C, D])

The default behavior when output_shape is (-1, D) is to flatten all dimensions between B and D:

>>> mod = hk.Reshape(output_shape=(-1, D))
>>> assert mod(x).shape == (B, H*W*C, D)

You can change the number of preserved leading dimensions via preserve_dims:

>>> mod = hk.Reshape(output_shape=(-1, D), preserve_dims=2)
>>> assert mod(x).shape == (B, H, W*C, D)

>>> mod = hk.Reshape(output_shape=(-1, D), preserve_dims=3)
>>> assert mod(x).shape == (B, H, W, C, D)

>>> mod = hk.Reshape(output_shape=(-1, D), preserve_dims=4)
>>> assert mod(x).shape == (B, H, W, C, 1, D)

Alternatively, a negative value of preserve_dims specifies the number of trailing dimensions to replace with output_shape:

>>> mod = hk.Reshape(output_shape=(-1, D), preserve_dims=-3)
>>> assert mod(x).shape == (B, H, W*C, D)

This is useful in the case of applying the same module to batched and unbatched outputs:

>>> mod = hk.Reshape(output_shape=(-1, D), preserve_dims=-3)
>>> assert mod(x[0]).shape == (H, W*C, D)

__init__(output_shape, preserve_dims=1, name=None)[source]#

Constructs a Reshape module.

Parameters

output_shape (Sequence[int]) – Shape to reshape the input tensor to while preserving its first preserve_dims dimensions. When the special value -1 appears in output_shape the corresponding size is automatically inferred. Note that -1 can only appear once in output_shape. To flatten all non-batch dimensions use Flatten.
preserve_dims (int) – Number of leading dimensions that will not be reshaped. If negative, this is interpreted instead as the number of trailing dimensions to replace with the new shape.
name (Optional[str]) – Name of the module.

Raises

ValueError – If preserve_dims is zero.

__call__(inputs)[source]#: Call self as a function.

Flatten#

class haiku.Flatten(preserve_dims=1, name=None)[source]#

Flattens the input, preserving the batch dimension(s).

By default, Flatten combines all dimensions except the first. Additional leading dimensions can be preserved by setting preserve_dims.

>>> x = jnp.ones([3, 2, 4])
>>> flat = hk.Flatten()
>>> flat(x).shape
(3, 8)

When the input to flatten has fewer than preserve_dims dimensions it is returned unchanged:

>>> x = jnp.ones([3])
>>> flat(x).shape
(3,)

Alternatively, a negative value of preserve_dims specifies the number of trailing dimensions flattened:

>>> x = jnp.ones([3, 2, 4])
>>> negative_flat = hk.Flatten(preserve_dims=-2)
>>> negative_flat(x).shape
(3, 8)

This allows the same module to be seamlessly applied to a single element or a batch of elements with the same element shape:

>> negative_flat(x[0]).shape (8,)

__init__(preserve_dims=1, name=None)[source]#

Constructs a Reshape module.

Parameters

output_shape – Shape to reshape the input tensor to while preserving its first preserve_dims dimensions. When the special value -1 appears in output_shape the corresponding size is automatically inferred. Note that -1 can only appear once in output_shape. To flatten all non-batch dimensions use Flatten.
preserve_dims (int) – Number of leading dimensions that will not be reshaped. If negative, this is interpreted instead as the number of trailing dimensions to replace with the new shape.
name (Optional[str]) – Name of the module.

Raises

ValueError – If preserve_dims is zero.

BatchApply#

class haiku.BatchApply(f, num_dims=2)[source]#

Temporarily merges leading dimensions of input tensors.

Merges the leading dimensions of a tensor into a single dimension, runs the given callable, then splits the leading dimension of the result to match the input.

Input arrays whose rank is smaller than the number of dimensions to collapse are passed unmodified.

This may be useful for applying a module to each timestep of e.g. a [Time, Batch, ...] array.

For some fs and platforms, this may be more efficient than jax.vmap(), especially when combined with other transformations like jax.grad().

__init__(f, num_dims=2)[source]#

Constructs a BatchApply module.

Parameters

f – The callable to be applied to the reshaped array.
num_dims – The number of dimensions to merge.

__call__(*args, **kwargs)[source]#: Call self as a function.

Embedding#

`Embed`([vocab_size, embed_dim, ...])	Module for embedding tokens in a low-dimensional space.
`EmbedLookupStyle`(value[, names, module, ...])	How to return the embedding matrices given IDs.

Embed#

class haiku.Embed(vocab_size=None, embed_dim=None, embedding_matrix=None, w_init=None, lookup_style='ARRAY_INDEX', name=None, precision=<Precision.HIGHEST: 2>)[source]#

Module for embedding tokens in a low-dimensional space.

__init__(vocab_size=None, embed_dim=None, embedding_matrix=None, w_init=None, lookup_style='ARRAY_INDEX', name=None, precision=<Precision.HIGHEST: 2>)[source]#

Constructs an Embed module.

Parameters

vocab_size (Optional[int]) – The number of unique tokens to embed. If not provided, an existing vocabulary matrix from which vocab_size can be inferred must be provided as embedding_matrix.
embed_dim (Optional[int]) – Number of dimensions to assign to each embedding. If an existing vocabulary matrix initializes the module, this should not be provided as it will be inferred.
embedding_matrix (Optional[Union[np.ndarray, jax.Array]]) – A matrix-like object equivalent in size to [vocab_size, embed_dim]. If given, it is used as the initial value for the embedding matrix and neither vocab_size or embed_dim need be given. If they are given, their values are checked to be consistent with the dimensions of embedding_matrix.
w_init (Optional[hk.initializers.Initializer]) – An initializer for the embeddings matrix. As a default, embeddings are initialized via a truncated normal distribution.
lookup_style (Union[str, EmbedLookupStyle]) – One of the enum values of EmbedLookupStyle determining how to access the value of the embeddings given an ID. Regardless the input should be a dense array of integer values representing ids. This setting changes how internally this module maps those ids to embeddings. The result is the same, but the speed and memory tradeoffs are different. It defaults to using NumPy-style array indexing. This value is only the default for the module, and at any given invocation can be overridden in __call__().
name (Optional[str]) – Optional name for this module.
precision (jax.lax.Precision) – Only used when lookup_style is ONE_HOT. The precision to use for the dot-product between the one-hot-encoded inputs and the embedding vectors. It is possible to attain a ~2x speedup on TPU using jax.lax.Precision.DEFAULT at the cost of a slightly lower precision.

Raises

ValueError – If none of embed_dim, embedding_matrix and vocab_size are supplied, or if embedding_matrix is supplied and embed_dim or vocab_size is not consistent with the supplied matrix.

__call__(ids, lookup_style=None, precision=None)[source]#

Lookup embeddings.

Looks up an embedding vector for each value in ids. All ids must be within [0, vocab_size) to prevent NaNs from propagating.

Parameters

ids (Union[jax.Array, Sequence[int]]) – integer array.
lookup_style (Optional[Union[str, hk.EmbedLookupStyle]]) – Overrides the lookup_style given in the constructor.
precision (Optional[jax.lax.Precision]) – Overrides the precision given in the constructor.

Return type

jax.Array

Returns

Tensor of ids.shape + [embedding_dim].

Raises

AttributeError – If lookup_style is not valid.
ValueError – If ids is not an integer array.

EmbedLookupStyle#

class haiku.EmbedLookupStyle(value, names=None, *, module=None, qualname=None, type=None, start=1, boundary=None)[source]#

How to return the embedding matrices given IDs.

ARRAY_INDEX = 1#

ONE_HOT = 2#

Utilities#

Deferred(factory[, call_methods])

Defers the construction of another module until the first call.

Deferred#

class haiku.Deferred(factory, call_methods=('__call__',))[source]#

Defers the construction of another module until the first call.

Deferred can be used to declare modules that depend on computed properties of other modules before those modules are defined. This allows users to separate the declaration and use of modules. For example at the start of your program you can declare two modules which are coupled:

>>> encoder = hk.Linear(64)
>>> decoder = hk.Deferred(lambda: hk.Linear(encoder.input_size))

Later you can use these naturally (note: that using decoder first would cause an error since encoder.input_size is only defined after encoder has been called):

>>> x = jnp.ones([8, 32])
>>> y = encoder(x)
>>> z = decoder(y)  # Constructs the Linear encoder by calling the lambda.

The result will satisfy the following conditions:

>>> assert x.shape == z.shape
>>> assert y.shape == (8, 64)
>>> assert decoder.input_size == encoder.output_size
>>> assert decoder.output_size == encoder.input_size

__init__(factory, call_methods=('__call__',))[source]#

Initializes the Deferred module.

Parameters

factory (Callable[[], T]) – A no argument callable which constructs the module to defer to. The first time one of the call_methods are called the factory will be run and then the constructed module will be called with the same method and arguments as the deferred module.
call_methods (Sequence[str]) – Methods which should trigger construction of the target module. The default value configures this module to construct the first time __call__ is run. If you want to add methods other than call you should explicitly pass them (optionally), for example call_methods=(“__call__”, “encode”, “decode”).

property target: T#

Returns the target module.

If the factory has not already run this will trigger construction. Subsequent calls to target will return the same instance.

Return type: T
Returns: A Module instance as created by the factory function passed into the constructor.

__call__(*args, **kwargs)[source]#: Call self as a function.

__setattr__(name, value)[source]#: Implement setattr(self, name, value).

__delattr__(name)[source]#: Implement delattr(self, name).

Initializers#

`Initializer`	alias of `Callable`[[`collections.abc.Sequence`[`int`], `Any`], `jax.Array`]
`Constant`(constant)	Initializes with a constant.
`Identity`([gain])	Initializer that generates the identity matrix.
`Orthogonal`([scale, axis])	Uniform scaling initializer.
`RandomNormal`([stddev, mean])	Initializes by sampling from a normal distribution.
`RandomUniform`([minval, maxval])	Initializes by sampling from a uniform distribution.
`TruncatedNormal`([stddev, mean, lower, upper])	Initializes by sampling from a truncated normal distribution.
`VarianceScaling`([scale, mode, distribution, ...])	Initializer which adapts its scale to the shape of the initialized array.
`UniformScaling`([scale])	Uniform scaling initializer.

Initializer#

haiku.initializers.Initializer#: alias of Callable[[collections.abc.Sequence[int], Any], jax.Array]

Constant#

class haiku.initializers.Constant(constant)[source]#

Initializes with a constant.

__init__(constant)[source]#

Constructs a Constant initializer.

Parameters: constant (Union[float, int, complex, np.ndarray, jax.Array]) – Constant to initialize with.

__call__(shape, dtype)[source]#

Call self as a function.

Return type: jax.Array

Identity#

class haiku.initializers.Identity(gain=1.0)[source]#

Initializer that generates the identity matrix.

Constructs a 2D identity matrix or batches of these.

__init__(gain=1.0)[source]#

Constructs an Identity initializer.

Parameters: gain (Union[float, np.ndarray, jax.Array]) – Multiplicative factor to apply to the identity matrix.

__call__(shape, dtype)[source]#

Call self as a function.

Return type: jax.Array

Orthogonal#

class haiku.initializers.Orthogonal(scale=1.0, axis=- 1)[source]#

Uniform scaling initializer.

__init__(scale=1.0, axis=- 1)[source]#

Construct an initializer for uniformly distributed orthogonal matrices.

These matrices will be row-orthonormal along the access specified by axis. If the rank of the weight is greater than 2, the shape will be flattened in all other dimensions and then will be row-orthonormal along the final dimension. Note that this only works if the axis dimension is larger, otherwise the matrix will be transposed (equivalently, it will be column orthonormal instead of row orthonormal).

If the shape is not square, the matrices will have orthonormal rows or columns depending on which side is smaller.

Parameters

scale – Scale factor.
axis – Which axis corresponds to the “output dimension” of the tensor.

Returns

An orthogonally initialized parameter.

__call__(shape, dtype)[source]#

Call self as a function.

Return type: jax.Array

RandomNormal#

class haiku.initializers.RandomNormal(stddev=1.0, mean=0.0)[source]#

Initializes by sampling from a normal distribution.

__init__(stddev=1.0, mean=0.0)[source]#

Constructs a RandomNormal initializer.

Parameters

stddev – The standard deviation of the normal distribution to sample from.
mean – The mean of the normal distribution to sample from.

__call__(shape, dtype)[source]#

Call self as a function.

Return type: jax.Array

RandomUniform#

class haiku.initializers.RandomUniform(minval=0.0, maxval=1.0)[source]#

Initializes by sampling from a uniform distribution.

__init__(minval=0.0, maxval=1.0)[source]#

Constructs a RandomUniform initializer.

Parameters

minval – The lower limit of the uniform distribution.
maxval – The upper limit of the uniform distribution.

__call__(shape, dtype)[source]#

Call self as a function.

Return type: jax.Array

TruncatedNormal#

class haiku.initializers.TruncatedNormal(stddev=1.0, mean=0.0, lower=- 2.0, upper=2.0)[source]#

Initializes by sampling from a truncated normal distribution.

__init__(stddev=1.0, mean=0.0, lower=- 2.0, upper=2.0)[source]#

Constructs a TruncatedNormal initializer.

Parameters

stddev (Union[float, jax.Array]) – The standard deviation parameter of the truncated normal distribution.
mean (Union[float, complex, jax.Array]) – The mean of the truncated normal distribution.
lower (Union[float, jax.Array]) – Float or array representing the lower bound for truncation.
upper (Union[float, jax.Array]) – Float or array representing the upper bound for truncation.

__call__(shape, dtype)[source]#

Call self as a function.

Return type: jax.Array

VarianceScaling#

class haiku.initializers.VarianceScaling(scale=1.0, mode='fan_in', distribution='truncated_normal', fan_in_axes=None)[source]#

Initializer which adapts its scale to the shape of the initialized array.

The initializer first computes the scaling factor s = scale / n, where n is:

Number of input units in the weight tensor, if mode = fan_in.

Number of output units, if mode = fan_out.

Average of the numbers of input and output units, if mode = fan_avg.

Then, with distribution="truncated_normal" or "normal", samples are drawn from a distribution with a mean of zero and a standard deviation (after truncation, if used) stddev = sqrt(s).

With distribution=uniform, samples are drawn from a uniform distribution within [-limit, limit], with limit = sqrt(3 * s).

The variance scaling initializer can be configured to generate other standard initializers using the scale, mode and distribution arguments. Here are some example configurations:

Name	Parameters
glorot_uniform	VarianceScaling(1.0, “fan_avg”, “uniform”)
glorot_normal	VarianceScaling(1.0, “fan_avg”, “truncated_normal”)
lecun_uniform	VarianceScaling(1.0, “fan_in”, “uniform”)
lecun_normal	VarianceScaling(1.0, “fan_in”, “truncated_normal”)
he_uniform	VarianceScaling(2.0, “fan_in”, “uniform”)
he_normal	VarianceScaling(2.0, “fan_in”, “truncated_normal”)

__init__(scale=1.0, mode='fan_in', distribution='truncated_normal', fan_in_axes=None)[source]#

Constructs the VarianceScaling initializer.

Parameters

scale – Scale to multiply the variance by.
mode – One of fan_in, fan_out, fan_avg
distribution – Random distribution to use. One of truncated_normal, normal or uniform.
fan_in_axes – Optional sequence of int specifying which axes of the shape are part of the fan-in. If none provided, then the weight is assumed to be like a convolution kernel, where all leading dimensions are part of the fan-in, and only the trailing dimension is part of the fan-out. Useful if instantiating multi-headed attention weights.

__call__(shape, dtype)[source]#

Call self as a function.

Return type: jax.Array

UniformScaling#

class haiku.initializers.UniformScaling(scale=1.0)[source]#

Uniform scaling initializer.

Initializes by sampling from a uniform distribution, but with the variance scaled by the inverse square root of the number of input units, multiplied by the scale.

__init__(scale=1.0)[source]#

Constructs the UniformScaling initializer.

Parameters: scale – Scale to multiply the upper limit of the uniform distribution by.

__call__(shape, dtype)[source]#

Call self as a function.

Return type: jax.Array

Paddings#

`PadFn`	alias of `Callable`[[`int`], `tuple`[`int`, `int`]]
`is_padfn`(padding)	Tests whether the given argument is a single or sequence of PadFns.
`create`(padding, kernel, rate, n)	Generates the padding required for a given padding algorithm.
`create_from_padfn`(padding, kernel, rate, n)	Generates the padding required for a given padding algorithm.
`create_from_tuple`(padding, n)	Create a padding tuple using partially specified padding tuple.
`causal`(effective_kernel_size)	Pre-padding such that output has no dependence on the future.
`full`(effective_kernel_size)	Maximal padding whilst not convolving over just padded elements.
`reverse_causal`(effective_kernel_size)	Post-padding such that output has no dependence on the past.
`same`(effective_kernel_size)	Pads such that the output size matches input size for stride=1.
`valid`(effective_kernel_size)	No padding.

PadFn#

haiku.pad.PadFn#: alias of Callable[[int], tuple[int, int]]

is_padfn#

haiku.pad.is_padfn(padding)[source]#

Tests whether the given argument is a single or sequence of PadFns.

Return type: bool

create#

haiku.pad.create(padding, kernel, rate, n)[source]#

Generates the padding required for a given padding algorithm.

Parameters

padding (Union[hk.pad.PadFn, Sequence[hk.pad.PadFn]]) – callable/tuple or a sequence of callables/tuples. The callables take an integer representing the effective kernel size (kernel size when the rate is 1) and return a sequence of two integers representing the padding before and padding after for that dimension. The tuples are defined with two elements, padding before and after. If padding is a sequence it must be of length 1 or n.
kernel (Union[int, Sequence[int]]) – int or sequence of ints of length n. The size of the kernel for each dimension. If it is an int it will be replicated for the non channel and batch dimensions.
rate (Union[int, Sequence[int]]) – int or sequence of ints of length n. The dilation rate for each dimension. If it is an int it will be replicated for the non channel and batch dimensions.
n (int) – the number of spatial dimensions.

Return type

Sequence[tuple[int, int]]

Returns

A sequence of length n containing the padding for each element. These are of the form [pad_before, pad_after].

create_from_padfn#

haiku.pad.create_from_padfn(padding, kernel, rate, n)[source]#

Generates the padding required for a given padding algorithm.

Parameters

padding (Union[hk.pad.PadFn, Sequence[hk.pad.PadFn]]) – callable/tuple or a sequence of callables/tuples. The callables take an integer representing the effective kernel size (kernel size when the rate is 1) and return a sequence of two integers representing the padding before and padding after for that dimension. The tuples are defined with two elements, padding before and after. If padding is a sequence it must be of length 1 or n.
kernel (Union[int, Sequence[int]]) – int or sequence of ints of length n. The size of the kernel for each dimension. If it is an int it will be replicated for the non channel and batch dimensions.
rate (Union[int, Sequence[int]]) – int or sequence of ints of length n. The dilation rate for each dimension. If it is an int it will be replicated for the non channel and batch dimensions.
n (int) – the number of spatial dimensions.

Return type

Sequence[tuple[int, int]]

Returns

A sequence of length n containing the padding for each element. These are of the form [pad_before, pad_after].

create_from_tuple#

haiku.pad.create_from_tuple(padding, n)[source]#

Create a padding tuple using partially specified padding tuple.

Return type: Sequence[tuple[int, int]]

causal#

haiku.pad.causal(effective_kernel_size)[source]#

Pre-padding such that output has no dependence on the future.

Return type: tuple[int, int]

full#

haiku.pad.full(effective_kernel_size)[source]#

Maximal padding whilst not convolving over just padded elements.

Return type: tuple[int, int]

reverse_causal#

haiku.pad.reverse_causal(effective_kernel_size)[source]#

Post-padding such that output has no dependence on the past.

Return type: tuple[int, int]

same#

haiku.pad.same(effective_kernel_size)[source]#

Pads such that the output size matches input size for stride=1.

Return type: tuple[int, int]

valid#

haiku.pad.valid(effective_kernel_size)[source]#

No padding.

Return type: tuple[int, int]

Full Networks#

MLP#

class haiku.nets.MLP(output_sizes, w_init=None, b_init=None, with_bias=True, activation=<jax._src.custom_derivatives.custom_jvp object>, activate_final=False, name=None)[source]#

A multi-layer perceptron module.

__init__(output_sizes, w_init=None, b_init=None, with_bias=True, activation=<jax._src.custom_derivatives.custom_jvp object>, activate_final=False, name=None)[source]#

Constructs an MLP.

Parameters

output_sizes (Iterable[int]) – Sequence of layer sizes.
w_init (Optional[hk.initializers.Initializer]) – Initializer for Linear weights.
b_init (Optional[hk.initializers.Initializer]) – Initializer for Linear bias. Must be None if with_bias=False.
with_bias (bool) – Whether or not to apply a bias in each layer.
activation (Callable[[jax.Array], jax.Array]) – Activation function to apply between Linear layers. Defaults to ReLU.
activate_final (bool) – Whether or not to activate the final layer of the MLP.
name (Optional[str]) – Optional name for this module.

Raises

ValueError – If with_bias is False and b_init is not None.

__call__(inputs, dropout_rate=None, rng=None)[source]#

Connects the module to some inputs.

Parameters

inputs (jax.Array) – A Tensor of shape [batch_size, input_size].
dropout_rate (Optional[float]) – Optional dropout rate.
rng – Optional RNG key. Require when using dropout.

Return type

jax.Array

Returns

The output of the model of size [batch_size, output_size].

reverse(activate_final=None, name=None)[source]#

Returns a new MLP which is the layer-wise reverse of this MLP.

NOTE: Since computing the reverse of an MLP requires knowing the input size of each linear layer this method will fail if the module has not been called at least once.

The contract of reverse is that the reversed module will accept the output of the parent module as input and produce an output which is the input size of the parent.

>>> mlp = hk.nets.MLP([1, 2, 3])
>>> mlp_in = jnp.ones([1, 2])
>>> y = mlp(mlp_in)
>>> rev = mlp.reverse()
>>> rev_mlp_out = rev(y)
>>> mlp_in.shape == rev_mlp_out.shape
True

Parameters

activate_final (Optional[bool]) – Whether the final layer of the MLP should be activated.
name (Optional[str]) – Optional name for the new module. The default name will be the name of the current module prefixed with "reversed_".

Return type

MLP

Returns

An MLP instance which is the reverse of the current instance. Note these instances do not share weights and, apart from being symmetric to each other, are not coupled in any way.

MobileNet#

MobileNetV1#

class haiku.nets.MobileNetV1(strides=(1, 2, 1, 2, 1, 2, 1, 1, 1, 1, 1, 2, 1), channels=(64, 128, 128, 256, 256, 512, 512, 512, 512, 512, 512, 1024, 1024), num_classes=1000, use_bn=True, name=None)[source]#

MobileNetV1 model.

__init__(strides=(1, 2, 1, 2, 1, 2, 1, 1, 1, 1, 1, 2, 1), channels=(64, 128, 128, 256, 256, 512, 512, 512, 512, 512, 512, 1024, 1024), num_classes=1000, use_bn=True, name=None)[source]#

Constructs a MobileNetV1 model.

Parameters

strides (Sequence[int]) – The stride to use the in depthwise convolution in each mobilenet block.
channels (Sequence[int]) – Number of output channels from the pointwise convolution to use in each block.
num_classes (int) – Number of classes.
use_bn (bool) – Whether or not to use batch normalization. Defaults to True. When true, biases are not used. When false, biases are used.
name (Optional[str]) – Name of the module.

__call__(inputs, is_training)[source]#

Call self as a function.

Return type: jax.Array

ResNet#

`ResNet`(blocks_per_group, num_classes[, ...])	ResNet model.
`ResNet.BlockGroup`(channels, num_blocks, ...)	Higher level block for ResNet implementation.
`ResNet.BlockV1`(channels, stride, ...[, name])	ResNet V1 block with optional bottleneck.
`ResNet.BlockV2`(channels, stride, ...[, name])	ResNet V2 block with optional bottleneck.
`ResNet18`(num_classes[, bn_config, ...])	ResNet18.
`ResNet34`(num_classes[, bn_config, ...])	ResNet34.
`ResNet50`(num_classes[, bn_config, ...])	ResNet50.
`ResNet101`(num_classes[, bn_config, ...])	ResNet101.
`ResNet152`(num_classes[, bn_config, ...])	ResNet152.
`ResNet200`(num_classes[, bn_config, ...])	ResNet200.

ResNet#

class haiku.nets.ResNet(blocks_per_group, num_classes, bn_config=None, resnet_v2=False, bottleneck=True, channels_per_group=(256, 512, 1024, 2048), use_projection=(True, True, True, True), logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

ResNet model.

class BlockGroup(channels, num_blocks, stride, bn_config, resnet_v2, bottleneck, use_projection, name=None)[source]#

Higher level block for ResNet implementation.

__call__(inputs, is_training, test_local_stats)[source]#: Call self as a function.

__init__(channels, num_blocks, stride, bn_config, resnet_v2, bottleneck, use_projection, name=None)[source]#

Initializes the current module with the given name.

Subclasses should call this constructor before creating other modules or variables such that those modules are named correctly.

Parameters: name (Optional[str]) – An optional string name for the class. Must be a valid Python identifier. If name is not provided then the class name for the current instance is converted to lower_snake_case and used instead.

class BlockV1(channels, stride, use_projection, bn_config, bottleneck, name=None)[source]#

ResNet V1 block with optional bottleneck.

__call__(inputs, is_training, test_local_stats)[source]#: Call self as a function.

__init__(channels, stride, use_projection, bn_config, bottleneck, name=None)[source]#

Initializes the current module with the given name.

Subclasses should call this constructor before creating other modules or variables such that those modules are named correctly.

Parameters: name (Optional[str]) – An optional string name for the class. Must be a valid Python identifier. If name is not provided then the class name for the current instance is converted to lower_snake_case and used instead.

class BlockV2(channels, stride, use_projection, bn_config, bottleneck, name=None)[source]#

ResNet V2 block with optional bottleneck.

__call__(inputs, is_training, test_local_stats)[source]#: Call self as a function.

__init__(channels, stride, use_projection, bn_config, bottleneck, name=None)[source]#

Initializes the current module with the given name.

Subclasses should call this constructor before creating other modules or variables such that those modules are named correctly.

Parameters: name (Optional[str]) – An optional string name for the class. Must be a valid Python identifier. If name is not provided then the class name for the current instance is converted to lower_snake_case and used instead.

__init__(blocks_per_group, num_classes, bn_config=None, resnet_v2=False, bottleneck=True, channels_per_group=(256, 512, 1024, 2048), use_projection=(True, True, True, True), logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

Constructs a ResNet model.

Parameters

blocks_per_group (Sequence[int]) – A sequence of length 4 that indicates the number of blocks created in each group.
num_classes (int) – The number of classes to classify the inputs into.
bn_config (Optional[Mapping[str, FloatStrOrBool]]) – A dictionary of two elements, decay_rate and eps to be passed on to the BatchNorm layers. By default the decay_rate is 0.9 and eps is 1e-5.
resnet_v2 (bool) – Whether to use the v1 or v2 ResNet implementation. Defaults to False.
bottleneck (bool) – Whether the block should bottleneck or not. Defaults to True.
channels_per_group (Sequence[int]) – A sequence of length 4 that indicates the number of channels used for each block in each group.
use_projection (Sequence[bool]) – A sequence of length 4 that indicates whether each residual block should use projection.
logits_config (Optional[Mapping[str, Any]]) – A dictionary of keyword arguments for the logits layer.
name (Optional[str]) – Name of the module.
initial_conv_config (Optional[Mapping[str, FloatStrOrBool]]) – Keyword arguments passed to the constructor of the initial Conv2D module.
strides (Sequence[int]) – A sequence of length 4 that indicates the size of stride of convolutions for each block in each group.

__call__(inputs, is_training, test_local_stats=False)[source]#: Call self as a function.

ResNet18#

class haiku.nets.ResNet18(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

ResNet18.

__init__(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

Constructs a ResNet model.

Parameters

num_classes (int) – The number of classes to classify the inputs into.
bn_config (Optional[Mapping[str, FloatStrOrBool]]) – A dictionary of two elements, decay_rate and eps to be passed on to the BatchNorm layers.
resnet_v2 (bool) – Whether to use the v1 or v2 ResNet implementation. Defaults to False.
logits_config (Optional[Mapping[str, Any]]) – A dictionary of keyword arguments for the logits layer.
name (Optional[str]) – Name of the module.
initial_conv_config (Optional[Mapping[str, FloatStrOrBool]]) – Keyword arguments passed to the constructor of the initial Conv2D module.
strides (Sequence[int]) – A sequence of length 4 that indicates the size of stride of convolutions for each block in each group.

ResNet34#

class haiku.nets.ResNet34(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

ResNet34.

__init__(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

Constructs a ResNet model.

Parameters

num_classes (int) – The number of classes to classify the inputs into.
bn_config (Optional[Mapping[str, FloatStrOrBool]]) – A dictionary of two elements, decay_rate and eps to be passed on to the BatchNorm layers.
resnet_v2 (bool) – Whether to use the v1 or v2 ResNet implementation. Defaults to False.
logits_config (Optional[Mapping[str, Any]]) – A dictionary of keyword arguments for the logits layer.
name (Optional[str]) – Name of the module.
initial_conv_config (Optional[Mapping[str, FloatStrOrBool]]) – Keyword arguments passed to the constructor of the initial Conv2D module.
strides (Sequence[int]) – A sequence of length 4 that indicates the size of stride of convolutions for each block in each group.

ResNet50#

class haiku.nets.ResNet50(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

ResNet50.

__init__(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

Constructs a ResNet model.

Parameters

num_classes (int) – The number of classes to classify the inputs into.
bn_config (Optional[Mapping[str, FloatStrOrBool]]) – A dictionary of two elements, decay_rate and eps to be passed on to the BatchNorm layers.
resnet_v2 (bool) – Whether to use the v1 or v2 ResNet implementation. Defaults to False.
logits_config (Optional[Mapping[str, Any]]) – A dictionary of keyword arguments for the logits layer.
name (Optional[str]) – Name of the module.
initial_conv_config (Optional[Mapping[str, FloatStrOrBool]]) – Keyword arguments passed to the constructor of the initial Conv2D module.
strides (Sequence[int]) – A sequence of length 4 that indicates the size of stride of convolutions for each block in each group.

ResNet101#

class haiku.nets.ResNet101(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

ResNet101.

__init__(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

Constructs a ResNet model.

Parameters

num_classes (int) – The number of classes to classify the inputs into.
bn_config (Optional[Mapping[str, FloatStrOrBool]]) – A dictionary of two elements, decay_rate and eps to be passed on to the BatchNorm layers.
resnet_v2 (bool) – Whether to use the v1 or v2 ResNet implementation. Defaults to False.
logits_config (Optional[Mapping[str, Any]]) – A dictionary of keyword arguments for the logits layer.
name (Optional[str]) – Name of the module.
initial_conv_config (Optional[Mapping[str, FloatStrOrBool]]) – Keyword arguments passed to the constructor of the initial Conv2D module.
strides (Sequence[int]) – A sequence of length 4 that indicates the size of stride of convolutions for each block in each group.

ResNet152#

class haiku.nets.ResNet152(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

ResNet152.

__init__(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

Constructs a ResNet model.

Parameters

num_classes (int) – The number of classes to classify the inputs into.
bn_config (Optional[Mapping[str, FloatStrOrBool]]) – A dictionary of two elements, decay_rate and eps to be passed on to the BatchNorm layers.
resnet_v2 (bool) – Whether to use the v1 or v2 ResNet implementation. Defaults to False.
logits_config (Optional[Mapping[str, Any]]) – A dictionary of keyword arguments for the logits layer.
name (Optional[str]) – Name of the module.
initial_conv_config (Optional[Mapping[str, FloatStrOrBool]]) – Keyword arguments passed to the constructor of the initial Conv2D module.
strides (Sequence[int]) – A sequence of length 4 that indicates the size of stride of convolutions for each block in each group.

ResNet200#

class haiku.nets.ResNet200(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

ResNet200.

__init__(num_classes, bn_config=None, resnet_v2=False, logits_config=None, name=None, initial_conv_config=None, strides=(1, 2, 2, 2))[source]#

Constructs a ResNet model.

Parameters

num_classes (int) – The number of classes to classify the inputs into.
bn_config (Optional[Mapping[str, FloatStrOrBool]]) – A dictionary of two elements, decay_rate and eps to be passed on to the BatchNorm layers.
resnet_v2 (bool) – Whether to use the v1 or v2 ResNet implementation. Defaults to False.
logits_config (Optional[Mapping[str, Any]]) – A dictionary of keyword arguments for the logits layer.
name (Optional[str]) – Name of the module.
initial_conv_config (Optional[Mapping[str, FloatStrOrBool]]) – Keyword arguments passed to the constructor of the initial Conv2D module.
strides (Sequence[int]) – A sequence of length 4 that indicates the size of stride of convolutions for each block in each group.

VectorQuantizer#

`VectorQuantizer`(embedding_dim, ...[, dtype, ...])	Haiku module representing the VQ-VAE layer.
`VectorQuantizerEMA`(embedding_dim, ...[, ...])	Haiku module representing the VQ-VAE layer.

VectorQuantizer#

class haiku.nets.VectorQuantizer(embedding_dim, num_embeddings, commitment_cost, dtype=<class 'jax.numpy.float32'>, name=None, cross_replica_axis=None)[source]#

Haiku module representing the VQ-VAE layer.

Implements the algorithm presented in “Neural Discrete Representation Learning” by van den Oord et al. https://arxiv.org/abs/1711.00937

Input any tensor to be quantized. Last dimension will be used as space in which to quantize. All other dimensions will be flattened and will be seen as different examples to quantize.

The output tensor will have the same shape as the input.

For example a tensor with shape [16, 32, 32, 64] will be reshaped into [16384, 64] and all 16384 vectors (each of 64 dimensions) will be quantized independently.

embedding_dim#: integer representing the dimensionality of the tensors in the quantized space. Inputs to the modules must be in this format as well.

num_embeddings#: integer, the number of vectors in the quantized space.

commitment_cost#: scalar which controls the weighting of the loss terms (see equation 4 in the paper - this variable is Beta).

__init__(embedding_dim, num_embeddings, commitment_cost, dtype=<class 'jax.numpy.float32'>, name=None, cross_replica_axis=None)[source]#

Initializes a VQ-VAE module.

Parameters

embedding_dim (int) – dimensionality of the tensors in the quantized space. Inputs to the modules must be in this format as well.
num_embeddings (int) – number of vectors in the quantized space.
commitment_cost (float) – scalar which controls the weighting of the loss terms (see equation 4 in the paper - this variable is Beta).
dtype (Any) – dtype for the embeddings variable, defaults to float32.
name (Optional[str]) – name of the module.
cross_replica_axis (Optional[str]) – If not None, it should be a string representing the axis name over which this module is being run within a jax.pmap(). Supplying this argument means that perplexity is calculated across all replicas on that axis.

__call__(inputs, is_training)[source]#

Connects the module to some inputs.

Parameters

inputs – Tensor, final dimension must be equal to embedding_dim. All other leading dimensions will be flattened and treated as a large batch.
is_training – boolean, whether this connection is to training data.

Returns

Dictionary containing the following keys and values:

quantize: Tensor containing the quantized version of the input.
loss: Tensor containing the loss to optimize.
perplexity: Tensor containing the perplexity of the encodings.
encodings: Tensor containing the discrete encodings, ie which element of the quantized space each input element was mapped to.
encoding_indices: Tensor containing the discrete encoding indices, ie which element of the quantized space each input element was mapped to.

Return type

dict

quantize(encoding_indices)[source]#: Returns embedding tensor for a batch of indices.

VectorQuantizerEMA#

class haiku.nets.VectorQuantizerEMA(embedding_dim, num_embeddings, commitment_cost, decay, epsilon=1e-05, dtype=<class 'jax.numpy.float32'>, cross_replica_axis=None, name=None)[source]#

Haiku module representing the VQ-VAE layer.

Implements a slightly modified version of the algorithm presented in “Neural Discrete Representation Learning” by van den Oord et al. https://arxiv.org/abs/1711.00937

The difference between VectorQuantizerEMA and VectorQuantizer is that this module uses ExponentialMovingAverages to update the embedding vectors instead of an auxiliary loss. This has the advantage that the embedding updates are independent of the choice of optimizer (SGD, RMSProp, Adam, K-Fac, …) used for the encoder, decoder and other parts of the architecture. For most experiments the EMA version trains faster than the non-EMA version.

Input any tensor to be quantized. Last dimension will be used as space in which to quantize. All other dimensions will be flattened and will be seen as different examples to quantize.

The output tensor will have the same shape as the input.

For example a tensor with shape [16, 32, 32, 64] will be reshaped into [16384, 64] and all 16384 vectors (each of 64 dimensions) will be quantized independently.

embedding_dim#: integer representing the dimensionality of the tensors in the quantized space. Inputs to the modules must be in this format as well.

num_embeddings#: integer, the number of vectors in the quantized space.

commitment_cost#: scalar which controls the weighting of the loss terms (see equation 4 in the paper).

decay#: float, decay for the moving averages.

epsilon#: small float constant to avoid numerical instability.

__init__(embedding_dim, num_embeddings, commitment_cost, decay, epsilon=1e-05, dtype=<class 'jax.numpy.float32'>, cross_replica_axis=None, name=None)[source]#

Initializes a VQ-VAE EMA module.

Parameters

embedding_dim – integer representing the dimensionality of the tensors in the quantized space. Inputs to the modules must be in this format as well.
num_embeddings – integer, the number of vectors in the quantized space.
commitment_cost – scalar which controls the weighting of the loss terms (see equation 4 in the paper - this variable is Beta).
decay – float between 0 and 1, controls the speed of the Exponential Moving Averages.
epsilon (float) – small constant to aid numerical stability, default 1e-5.
dtype (Any) – dtype for the embeddings variable, defaults to float32.
cross_replica_axis (Optional[str]) – If not None, it should be a string representing the axis name over which this module is being run within a jax.pmap(). Supplying this argument means that cluster statistics and the perplexity are calculated across all replicas on that axis.
name (Optional[str]) – name of the module.

__call__(inputs, is_training)[source]#

Connects the module to some inputs.

Parameters

inputs – Tensor, final dimension must be equal to embedding_dim. All other leading dimensions will be flattened and treated as a large batch.
is_training – boolean, whether this connection is to training data. When this is set to False, the internal moving average statistics will not be updated.

Returns

Dictionary containing the following keys and values:

quantize: Tensor containing the quantized version of the input.
loss: Tensor containing the loss to optimize.
perplexity: Tensor containing the perplexity of the encodings.
encodings: Tensor containing the discrete encodings, ie which element of the quantized space each input element was mapped to.
encoding_indices: Tensor containing the discrete encoding indices, ie which element of the quantized space each input element was mapped to.

Return type

dict

quantize(encoding_indices)[source]#: Returns embedding tensor for a batch of indices.

JAX Fundamentals#

Control Flow#

`cond`(pred, true_fun, false_fun, *operands[, ...])	Conditionally apply `true_fun` or `false_fun`.
`fori_loop`(lower, upper, body_fun, init_val)	Equivalent to `jax.lax.fori_loop()` with Haiku state passed in/out.
`map`(f, xs)	Equivalent to `jax.lax.map()` but with Haiku state passed in/out.
`scan`(f, init, xs[, length, reverse, unroll])	Equivalent to `jax.lax.scan()` but with Haiku state passed in/out.
`switch`(index, branches, *operands)	Equivalent to `jax.lax.switch()` but with Haiku state passed in/out.
`while_loop`(cond_fun, body_fun, init_val)	Equivalent to jax.lax.while_loop with Haiku state threaded in/out.

cond#

haiku.cond(pred, true_fun, false_fun, *operands, operand=<object object>, linear=None)[source]#

Conditionally apply true_fun or false_fun.

Wraps XLA’s Conditional operator.

Provided arguments are correctly typed, cond() has equivalent semantics to this Python implementation, where pred must be a scalar type:

def cond(pred, true_fun, false_fun, *operands):
  if pred:
    return true_fun(*operands)
  else:
    return false_fun(*operands)

In contrast with jax.lax.select(), using cond indicates that only one of the two branches is executed (up to compiler rewrites and optimizations). However, when transformed with vmap() to operate over a batch of predicates, cond is converted to select().

Parameters

pred – Boolean scalar type, indicating which branch function to apply.
true_fun (Callable) – Function (A -> B), to be applied if pred is True.
false_fun (Callable) – Function (A -> B), to be applied if pred is False.
operands – Operands (A) input to either branch depending on pred. The type can be a scalar, array, or any pytree (nested Python tuple/list/dict) thereof.

Returns

Value (B) of either true_fun(*operands) or false_fun(*operands), depending on the value of pred. The type can be a scalar, array, or any pytree (nested Python tuple/list/dict) thereof.

fori_loop#

haiku.fori_loop(lower, upper, body_fun, init_val)[source]#: Equivalent to jax.lax.fori_loop() with Haiku state passed in/out.

map#

haiku.map(f, xs)[source]#: Equivalent to jax.lax.map() but with Haiku state passed in/out.

scan#

haiku.scan(f, init, xs, length=None, reverse=False, unroll=1)[source]#: Equivalent to jax.lax.scan() but with Haiku state passed in/out.

switch#

haiku.switch(index, branches, *operands)[source]#

Equivalent to jax.lax.switch() but with Haiku state passed in/out.

Note that creating parameters inside a switch branch is not supported, as such at init time we recommend you unconditionally evaluate all branches of your switch and only use the switch at apply. For example:

>>> experts = [hk.nets.MLP([300, 100, 10]) for _ in range(5)]
>>> x = jnp.ones([1, 28 * 28])
>>> if hk.running_init():
...   # During init unconditionally create params/state for all experts.
...   for expert in experts:
...     out = expert(x)
... else:
...   # During apply conditionally apply (and update) only one expert.
...   index = jax.random.randint(hk.next_rng_key(), [], 0, len(experts) - 1)
...   out = hk.switch(index, experts, x)

Parameters

index – Integer scalar type, indicating which branch function to apply.
branches – Sequence of functions (A -> B) to be applied based on index.
operands – Operands (A) input to whichever branch is applied.

Returns

Value (B) of branch(*operands) for the branch that was selected based on index.

while_loop#

haiku.while_loop(cond_fun, body_fun, init_val)[source]#: Equivalent to jax.lax.while_loop with Haiku state threaded in/out.

JAX Transforms#

`eval_shape`(fun, args, *kwargs)	Equivalent to jax.eval_shape with any changed Haiku state discarded.
`grad`(fun[, argnums, has_aux, holomorphic])	Creates a function which evaluates the gradient of `fun`.
`remat`(fun, *[, prevent_cse, policy, ...])	Equivalent to jax.checkpoint but passing Haiku state.
`value_and_grad`(fun[, argnums, has_aux, ...])	Creates a function which evaluates both `fun` and the grad of `fun`.
`vmap`(fun[, in_axes, out_axes, axis_name, ...])	Equivalent to `jax.vmap()` with module parameters/state not mapped.

eval_shape#

haiku.eval_shape(fun, *args, **kwargs)[source]#: Equivalent to jax.eval_shape with any changed Haiku state discarded.

grad#

haiku.grad(fun, argnums=0, has_aux=False, holomorphic=False)[source]#

Creates a function which evaluates the gradient of fun.

NOTE: You only need this in a very specific case that you want to take a gradient inside a transform()ed function and the function you are differentiating uses set_state(). For example:

>>> class MyModule(hk.Module):
...   def __call__(self, x):
...     hk.set_state("last", x ** 2)
...     return x ** 2

>>> def f(x):
...   m = MyModule()
...   g = hk.grad(m)(x)
...   return g

>>> f = hk.transform_with_state(f)
>>> x = jnp.array(2.)
>>> params, state = jax.jit(f.init)(None, x)
>>> print(state["my_module"]["last"])
4.0

Parameters

fun – Function to be differentiated. Its arguments at positions specified by argnums should be arrays, scalars, or standard Python containers. It should return a scalar (which includes arrays with shape () but not arrays with shape (1,) etc.)
argnums – Optional, integer or tuple of integers. Specifies which positional argument(s) to differentiate with respect to (default 0).
has_aux – Optional, bool. Indicates whether fun returns a pair where the first element is considered the output of the mathematical function to be differentiated and the second element is auxiliary data. Default False.
holomorphic – Optional, bool. Indicates whether fun is promised to be holomorphic. Default False.

Returns

A function with the same arguments as fun, that evaluates the gradient of fun. If argnums is an integer then the gradient has the same shape and type as the positional argument indicated by that integer. If argnums is a tuple of integers, the gradient is a tuple of values with the same shapes and types as the corresponding arguments. If has_aux is True then a pair of gradient, auxiliary_data is returned.

For example:

>>> grad_tanh = jax.grad(jax.numpy.tanh)
>>> print(grad_tanh(0.2))
0.96...

remat#

haiku.remat(fun, *, prevent_cse=True, policy=None, static_argnums=())[source]#

Equivalent to jax.checkpoint but passing Haiku state.

Return type: Callable

value_and_grad#

haiku.value_and_grad(fun, argnums=0, has_aux=False, holomorphic=False)[source]#

Creates a function which evaluates both fun and the grad of fun.

NOTE: You only need this in a very specific case that you want to take a gradient inside a transform()ed function and the function you are differentiating uses set_state(). For example:

>>> class MyModule(hk.Module):
...   def __call__(self, x):
...     hk.set_state("last", jnp.sum(x))
...     return x ** 2

>>> def f(x):
...   m = MyModule()
...   y, g = hk.value_and_grad(m)(x)
...   return y, g

>>> f = hk.transform_with_state(f)
>>> x = jnp.array(2.)
>>> _ = jax.jit(f.init)(None, x)

Parameters

fun – Function to be differentiated. Its arguments at positions specified by argnums should be arrays, scalars, or standard Python containers. It should return a scalar (which includes arrays with shape () but not arrays with shape (1,) etc.)
argnums – Optional, integer or tuple of integers. Specifies which positional argument(s) to differentiate with respect to (default 0).
has_aux – Optional, bool. Indicates whether fun returns a pair where the first element is considered the output of the mathematical function to be differentiated and the second element is auxiliary data. Default False.
holomorphic – Optional, bool. Indicates whether fun is promised to be holomorphic. Default False.

Returns

A function with the same arguments as fun that evaluates both fun and the gradient of fun and returns them as a pair (a two-element tuple). If argnums is an integer then the gradient has the same shape and type as the positional argument indicated by that integer. If argnums is a tuple of integers, the gradient is a tuple of values with the same shapes and types as the corresponding arguments.

vmap#

haiku.vmap(fun, in_axes=0, out_axes=0, axis_name=None, axis_size=None, *, split_rng)[source]#

Equivalent to jax.vmap() with module parameters/state not mapped.

The behaviour of Haiku random key APIs under vmap() is controlled by the split_rng argument:

>>> x = jnp.arange(2)
>>> f = hk.vmap(lambda _: hk.next_rng_key(), split_rng=False)
>>> key1, key2 = f(x)
>>> assert (key1 == key2).all()

>>> f = hk.vmap(lambda _: hk.next_rng_key(), split_rng=True)
>>> key1, key2 = f(x)
>>> assert not (key1 == key2).all()

Random numbers in Haiku are typically used for two things, firstly for initialising model parameters, and secondly for creating random samples as part of the forward pass of a neural network (e.g. for dropout). If you are using vmap() with a module that uses Haiku random keys for both (e.g. you don’t pass keys explicitly into the network), then it is quite likely that you will want to vary the value of split_rng depending on whether we are initalizing (e.g. creating model parameters) or applying the model. An easy way to do this is to set split_rng=(not hk.running_init()).

For more advanced use cases, such as mapping module parameters, we suggest users instead use lift() or transparent_lift() in combination with jax.vmap().

Parameters

fun (Callable[..., Any]) – See jax.vmap().
in_axes – See jax.vmap().
out_axes – See jax.vmap().
axis_name (Optional[str]) – See jax.vmap().
axis_size (Optional[int]) – See jax.vmap().
split_rng (bool) – Controls whether random key APIs in Haiku (e.g. next_rng_key()) return different (aka. the internal key is split before calling your mapped function) or the same (aka. the internal key is broadcast before calling your mapped fucntion) key. See the docstring for examples.

Return type

Callable[…, Any]

Returns

See jax.vmap().

Mixed Precision#

Automatic Mixed Precision#

`set_policy`(cls, policy)	Uses the given policy for all instances of the module class.
`current_policy`()	Retrieves the currently active policy in the current context.
`get_policy`(cls)	Retrieves the currently active policy for the given class.
`clear_policy`(cls)	Clears any policy assocated with the given class.
`push_policy`(cls, policy)	Sets the given policy for the given class while the context is active.

set_policy#

haiku.mixed_precision.set_policy(cls, policy)[source]#

Uses the given policy for all instances of the module class.

NOTE: Policies are only applied to modules created in the current thread.

A mixed precision policy describes how inputs, module parameters and module outputs should be cast at runtime. By applying a policy to a given type of module, you can control how all instances of that module behave in your program.

For example, you might want to try running a ResNet50 model in a mixture of float16 and float32 on GPU to get higher throughput. To do so you can apply a mixed precision policy to the ResNet50 type that will create parameters in float32, but cast them to float16 before use, along with all module inputs:

>>> policy = jmp.get_policy('params=float32,compute=float16,output=float32')
>>> hk.mixed_precision.set_policy(hk.nets.ResNet50, policy)
>>> net = hk.nets.ResNet50(4)
>>> x = jnp.ones([4, 224, 224, 3])
>>> print(net(x, is_training=True))
[[0. 0. 0. 0.]
 [0. 0. 0. 0.]
 [0. 0. 0. 0.]
 [0. 0. 0. 0.]]

For a fully worked mixed precision example see the imagenet example in Haiku’s examples directory. This example shows mixed precision on GPU offering a 2x speedup in training time with only a small impact on final top-1 accuracy.

>>> hk.mixed_precision.clear_policy(hk.nets.ResNet50)

Parameters

cls (type[hk.Module]) – A Haiku module class.
policy (jmp.Policy) – A JMP policy to apply to the module.

See also

push_policy(): Context manager for setting policies.
current_policy(): Retrieves the currently active policy (if any).
clear_policy(): Clears any policies associated with a class.
get_policy(): Gets the policy for a given class.

current_policy#

haiku.mixed_precision.current_policy()[source]#

Retrieves the currently active policy in the current context.

Return type: Optional[jmp.Policy]
Returns: The currently active mixed precision policy, or None.

See also

clear_policy(): Clears any policies associated with a class.
get_policy(): Gets the policy for a given class.
set_policy(): Sets a policy for a given class.
push_policy(): Context manager for setting policies.

get_policy#

haiku.mixed_precision.get_policy(cls)[source]#

Retrieves the currently active policy for the given class.

Note that policies applied explicitly to a top level class (e.g. ResNet) will be applied implicitly to all child modules (e.g. ConvND) called from the parent. This function only returns policies that have been applied explicitly (e.g. via set_policy()).

Parameters: cls (type[hk.Module]) – A Haiku module class.
Return type: Optional[jmp.Policy]
Returns: A JMP policy that is used for the given class, or None if one is not active.

See also

current_policy(): Retrieves the currently active policy (if any).
clear_policy(): Clears any policies associated with a class.
set_policy(): Sets a policy for a given class.
push_policy(): Context manager for setting policies.

clear_policy#

haiku.mixed_precision.clear_policy(cls)[source]#

Clears any policy assocated with the given class.

Parameters: cls (type[hk.Module]) – A Haiku module class.

See also

current_policy(): Retrieves the currently active policy (if any).
get_policy(): Gets the policy for a given class.
set_policy(): Sets a policy for a given class.
push_policy(): Context manager for setting policies.

push_policy#

haiku.mixed_precision.push_policy(cls, policy)[source]#

Sets the given policy for the given class while the context is active.

Parameters

cls (type[hk.Module]) – A Haiku module class.
policy (jmp.Policy) – A JMP policy to apply to the module.

Yields

None.

See also

clear_policy(): Clears any policies associated with a class.
get_policy(): Gets the policy for a given class.
set_policy(): Sets a policy for a given class.
current_policy(): Retrieves the currently active policy (if any).

🚧 Experimental#

Graphviz Visualisation#

abstract_to_dot(fun)

Converts a function using Haiku modules to a dot graph.

abstract_to_dot#

haiku.experimental.abstract_to_dot(fun)[source]#

Converts a function using Haiku modules to a dot graph.

Same as to_dot() but uses JAX’s abstract interpretation machinery to evaluate the function without requiring concrete inputs. Valid inputs for the wrapped function include jax.ShapeDtypeStruct.

abstract_to_dot() does not support data-dependent control-flow, because no concrete values are provided to the function.

Parameters: fun (Callable[..., Any]) – A function using Haiku modules.
Return type: Callable[…, str]
Returns: A function that returns the source code string to a graphviz graph describing the operations executed by the given function clustered by Haiku module.

See also

to_dot(): Generates a graphviz graph using concrete inputs.

Summarisation#

`tabulate`(f, *[, columns, filters, ...])	Produces a summarised view of the execution of `f`.
`eval_summary`(f)	Records module method calls performed by `f`.
`ArraySpec`(shape, dtype)	Shaped and sized specification of an array.
`MethodInvocation`(module_details, args_spec, ...)	Record of a method being invoked on a given module.
`ModuleDetails`(module, method_name, params, state)	Module and method related information.

tabulate#

haiku.experimental.tabulate(f, *, columns=('module', 'config', 'owned_params', 'input', 'output', 'params_size', 'params_bytes'), filters=('has_output',), tabulate_kwargs={'tablefmt': 'grid'})[source]#

Produces a summarised view of the execution of f.

>>> def f(x):
...   return hk.nets.MLP([300, 100, 10])(x)
>>> x = jnp.ones([8, 28 * 28])
>>> f = hk.transform(f)
>>> print(hk.experimental.tabulate(f)(x))
+-------------------------+------------------------------------------+-----------------+------------+------------+---------------+---------------+
| Module                  | Config                                   | Module params   | Input      | Output     |   Param count |   Param bytes |
+=========================+==========================================+=================+============+============+===============+===============+
| mlp (MLP)               | MLP(output_sizes=[300, 100, 10])         |                 | f32[8,784] | f32[8,10]  |       266,610 |       1.07 MB |
+-------------------------+------------------------------------------+-----------------+------------+------------+---------------+---------------+
| mlp/~/linear_0 (Linear) | Linear(output_size=300, name='linear_0') | w: f32[784,300] | f32[8,784] | f32[8,300] |       235,500 |     942.00 KB |
|  └ mlp (MLP)            |                                          | b: f32[300]     |            |            |               |               |
+-------------------------+------------------------------------------+-----------------+------------+------------+---------------+---------------+
| mlp/~/linear_1 (Linear) | Linear(output_size=100, name='linear_1') | w: f32[300,100] | f32[8,300] | f32[8,100] |        30,100 |     120.40 KB |
|  └ mlp (MLP)            |                                          | b: f32[100]     |            |            |               |               |
+-------------------------+------------------------------------------+-----------------+------------+------------+---------------+---------------+
| mlp/~/linear_2 (Linear) | Linear(output_size=10, name='linear_2')  | w: f32[100,10]  | f32[8,100] | f32[8,10]  |         1,010 |       4.04 KB |
|  └ mlp (MLP)            |                                          | b: f32[10]      |            |            |               |               |
+-------------------------+------------------------------------------+-----------------+------------+------------+---------------+---------------+

Possible values for columns:

module: Displays module and method name.
config: Displays the constructor arguments used for the module.
owned_params: Displays parameters directly owned by this module.
input: Displays module inputs.
output: Displays module output.
params_size: Displays the number of parameters
params_bytes: Displays parameter size in bytes.

Possible values for filters:

has_output: Only include methods returning a value other than None.
has_params: Removes methods from modules that do not have parameters.

Parameters

f (Union[Callable[..., Any], hk.Transformed, hk.TransformedWithState]) – A function to transform OR one of the init/apply functions from Haiku or the result of transform() or transform_with_state().
columns (Optional[Sequence[str]]) – A list of column names to enable.
filters (Optional[Sequence[str]]) – A list of filters to apply to remove certain module methods.
tabulate_kwargs – Keyword arguments to pass to tabulate.tabulate(..).

Return type

Callable[…, str]

Returns

A callable that takes the same arguments as f but returns a string summarising the modules used during the execution of f.

See also

eval_summary(): Raw data used to generate this table.

eval_summary#

haiku.experimental.eval_summary(f)[source]#

Records module method calls performed by f.

>>> f = lambda x: hk.nets.MLP([300, 100, 10])(x)
>>> x = jnp.ones([8, 28 * 28])
>>> for i in hk.experimental.eval_summary(f)(x):
...   print("mod := {:14} | in := {} out := {}".format(
...       i.module_details.module.module_name, i.args_spec[0], i.output_spec))
mod := mlp            | in := f32[8,784] out := f32[8,10]
mod := mlp/~/linear_0 | in := f32[8,784] out := f32[8,300]
mod := mlp/~/linear_1 | in := f32[8,300] out := f32[8,100]
mod := mlp/~/linear_2 | in := f32[8,100] out := f32[8,10]

Parameters: f (Union[Callable[..., Any], hk.Transformed, hk.TransformedWithState]) – A function or transformed function to trace.
Return type: Callable[…, Sequence[MethodInvocation]]
Returns: A callable taking the same arguments as the provided function, but returning a sequence of MethodInvocation instances revealing the methods called on each module when applying f.

See also

tabulate(): Pretty prints a summary of the execution of a function.

ArraySpec#

class haiku.experimental.ArraySpec(shape, dtype)[source]#

Shaped and sized specification of an array.

shape#

Shape of the array.

Type: Sequence[int]

dtype#

DType of the array.

Type: jnp.dtype

__delattr__(name)#: Implement delattr(self, name).

__eq__(other)#: Return self==value.

__hash__()#: Return hash(self).

__init__(shape, dtype)#

__setattr__(name, value)#: Implement setattr(self, name, value).

MethodInvocation#

class haiku.experimental.MethodInvocation(module_details, args_spec, kwargs_spec, output_spec, context, call_stack)[source]#

Record of a method being invoked on a given module.

module_details#

Details about which module and method were invoked.

Type: ModuleDetails

args_spec#

Positional arguments to the method invocation with arrays replaced by ArraySpec.

Type: tuple[Any, …]

kwargs_spec#

Keyword arguments to the method invocation with arrays replaced by ArraySpec.

Type: dict[str, Any]

output_spec#

Output of the method invocation with arrays replaced by ArraySpec.

Type: Any

context#

Additional context information for the method call as provided by intercept_methods().

Type: hk.MethodContext

call_stack#

Stack of modules currently active while calling this module method. For example if A calls B which calls C then the call stack for C will be [B_DETAILS, A_DETAILS].

Type: Sequence[ModuleDetails]

__delattr__(name)#: Implement delattr(self, name).

__eq__(other)#: Return self==value.

Haiku documentation

Haiku API reference

Contents

Haiku API reference#

Haiku Fundamentals#

Haiku Transforms#

transform#

transform_with_state#

multi_transform#

multi_transform_with_state#

without_apply_rng#

without_state#

with_empty_state#

Modules, Parameters and State#

Module#

to_module#

get_parameter#

get_state#

set_state#

Getters and Interceptors#

custom_creator#

custom_getter#

custom_setter#

GetterContext#

SetterContext#

intercept_methods#

MethodContext#

Random Numbers#

PRNGSequence#

next_rng_key#

next_rng_keys#

maybe_next_rng_key#

reserve_rng_keys#

with_rng#

maybe_get_rng_sequence_state#

replace_rng_sequence_state#

Type Hints#

LSTMState#

Params#

MutableParams#

State#

MutableState#

Transformed#

TransformedWithState#

MultiTransformed#

MultiTransformedWithState#

ModuleProtocol#

SupportsCall#

Flax Interop#

Haiku inside Flax#

Module#

flatten_flax_to_haiku#

Flax inside Haiku#

lift#

Advanced State Management#

Lifting#

lift#

lift_with_state#

transparent_lift#

transparent_lift_with_state#

LiftWithStateUpdater#

Layer Stack#

layer_stack#

LayerStackTransparencyMapping#

Naming#

name_scope#

current_name#

DO_NOT_STORE#

get_params#

get_current_state#

get_initial_state#

force_name#

name_like#

transparent#

Visualisation#

to_dot#

Common Modules#

Linear#

Linear#

Bias#