tfm.nlp.layers.MoeLayer

Sparse MoE layer with per-token routing.

tfm.nlp.layers.MoeLayer(
    experts: tfm.nlp.layers.FeedForwardExperts,
    router: MaskedRouter,
    *,
    train_capacity_factor: float = 1.0,
    eval_capacity_factor: float = 1.0,
    examples_per_group: float = 1.0,
    name: str = 'moe',
    **kwargs
)

In this TF implementation, all experts need to fit onto a single device allowing for batch parallelism only.

Uses Keras add_loss() and add_metric() APIs.

experts Instance of FeedForwardExperts. Needs to have the same num_experts as the router.
router Instance of MaskedRouter to route the tokens to the different experts.
train_capacity_factor Scaling factor to increase the expert token capacity during training. This factor plays an analogous, but slightly different, role depending on the routing assignment algorithm:

  • For "tokens choose" routing, the capacity factor only affects the maximum number of tokens that an expert will process. It does not affect how many experts a given token is routed to; see the num_selected_experts attributes of "tokens choose" routers.
  • For "experts choose" routing, because experts always fill their buffer, increasing the capacity factor will increase the number of tokens that an expert will process AND will indirectly increase the number of experts that a given token is routed to.
eval_capacity_factor As above, but used during evaluation.
examples_per_group Number of examples to form a group. Router then performs top_k token selection for each expert on a per group basis. E.g. when examples_per_group=4.0, tokens are assigned to experts in groups formed from 4 examples. When examples_per_group=0.5, each example is split into 2 groups. examples_per_group must divide the local batch size. A larger group size will result in slower but more accurate top-k and sorting computations, whereas a smaller group size will result in faster but more approximate (and potentially less stable) routing choices. In practice, we find that imperfect routing choices are tolerable and recommend choosing a group size on the order of 4096 tokens, although this number will vary based on model configuration and size.
name Layer name.
**kwargs Forwarded to super.

num_experts Number of experts (i.e. number of independent feed-forward blocks).
activity_regularizer Optional regularizer function for the output of this layer.
compute_dtype The dtype of the layer's computations.

This is equivalent to Layer.dtype_policy.compute_dtype. Unless mixed precision is used, this is the same as Layer.dtype, the dtype of the weights.

Layers automatically cast their inputs to the compute dtype, which causes computations and the output to be in the compute dtype as well. This is done by the base Layer class in Layer.call, so you do not have to insert these casts if implementing your own layer.

Layers often perform certain internal computations in higher precision when compute_dtype is float16 or bfloat16 for numeric stability. The output will still typically be float16 or bfloat16 in such cases.

dtype The dtype of the layer weights.

This is equivalent to Layer.dtype_policy.variable_dtype. Unless mixed precision is used, this is the same as Layer.compute_dtype, the dtype of the layer's computations.

dtype_policy The dtype policy associated with this layer.

This is an instance of a tf.keras.mixed_precision.Policy.

dynamic Whether the layer is dynamic (eager-only); set in the constructor.
input Retrieves the input tensor(s) of a layer.

Only applicable if the layer has exactly one input, i.e. if it is connected to one incoming layer.

input_spec InputSpec instance(s) describing the input format for this layer.

When you create a layer subclass, you can set self.input_spec to enable the layer to run input compatibility checks when it is called. Consider a Conv2D layer: it can only be called on a single input tensor of rank 4. As such, you can set, in __init__():

self.input_spec = tf.keras.layers.InputSpec(ndim=4)

Now, if you try to call the layer on an input that isn't rank 4 (for instance, an input of shape (2,), it will raise a nicely-formatted error:

ValueError: Input 0 of layer conv2d is incompatible with the layer:
expected ndim=4, found ndim=1. Full shape received: [2]

Input checks that can be specified via input_spec include:

  • Structure (e.g. a single input, a list of 2 inputs, etc)
  • Shape
  • Rank (ndim)
  • Dtype

For more information, see tf.keras.layers.InputSpec.

losses List of losses added using the add_loss() API.

Variable regularization tensors are created when this property is accessed, so it is eager safe: accessing losses under a tf.GradientTape will propagate gradients back to the corresponding variables.

class MyLayer(tf.keras.layers.Layer):
  def call(self, inputs):
    self.add_loss(tf.abs(tf.reduce_mean(inputs)))
    return inputs
l = MyLayer()
l(np.ones((10, 1)))
l.losses
[1.0]
 
inputs = tf.keras.Input(shape=(10,))
x = tf.keras.layers.Dense(10)(inputs)
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)
# Activity regularization.
len(model.losses)
0
model.add_loss(tf.abs(tf.reduce_mean(x)))
len(model.losses)
1
 
inputs = tf.keras.Input(shape=(10,))
d = tf.keras.layers.Dense(10, kernel_initializer='ones')
x = d(inputs)
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)
# Weight regularization.
model.add_loss(lambda: tf.reduce_mean(d.kernel))
model.losses
[<tf.Tensor: shape=(), dtype=float32, numpy=1.0>]

metrics List of metrics attached to the layer.
non_trainable_weights List of all non-trainable weights tracked by this layer.

Non-trainable weights are not updated during training. They are expected to be updated manually in call().

output Retrieves the output tensor(s) of a layer.

Only applicable if the layer has exactly one output, i.e. if it is connected to one incoming layer.

supports_masking Whether this layer supports computing a mask using compute_mask.
trainable

trainable_weights List of all trainable weights tracked by this layer.

Trainable weights are updated via gradient descent during training.

variable_dtype Alias of Layer.dtype, the dtype of the weights.
weights Returns the list of all layer variables/weights.

Methods

add_loss

add_loss(
    losses, **kwargs
)

Add loss tensor(s), potentially dependent on layer inputs.

Some losses (for instance, activity regularization losses) may be dependent on the inputs passed when calling a layer. Hence, when reusing the same layer on different inputs a and b, some entries in layer.losses may be dependent on a and some on b. This method automatically keeps track of dependencies.

This method can be used inside a subclassed layer or model's call function, in which case losses should be a Tensor or list of Tensors.

Example:

class MyLayer(tf.keras.layers.Layer):
  def call(self, inputs):
    self.add_loss(tf.abs(tf.reduce_mean(inputs)))
    return inputs

The same code works in distributed training: the input to add_loss() is treated like a regularization loss and averaged across replicas by the training loop (both built-in Model.fit() and compliant custom training loops).

The add_loss method can also be called directly on a Functional Model during construction. In this case, any loss Tensors passed to this Model must be symbolic and be able to be traced back to the model's Inputs. These losses become part of the model's topology and are tracked in get_config.

Example:

inputs = tf.keras.Input(shape=(10,))
x = tf.keras.layers.Dense(10)(inputs)
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)
# Activity regularization.
model.add_loss(tf.abs(tf.reduce_mean(x)))

If this is not the case for your loss (if, for example, your loss references a Variable of one of the model's layers), you can wrap your loss in a zero-argument lambda. These losses are not tracked as part of the model's topology since they can't be serialized.

Example:

inputs = tf.keras.Input(shape=(10,))
d = tf.keras.layers.Dense(10)
x = d(inputs)
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)
# Weight regularization.
model.add_loss(lambda: tf.reduce_mean(d.kernel))

Args
losses Loss tensor, or list/tuple of tensors. Rather than tensors, losses may also be zero-argument callables which create a loss tensor.
**kwargs Used for backwards compatibility only.

build

build(
    input_shape
)

Creates the variables of the layer (for subclass implementers).

This is a method that implementers of subclasses of Layer or Model can override if they need a state-creation step in-between layer instantiation and layer call. It is invoked automatically before the first execution of call().

This is typically used to create the weights of Layer subclasses (at the discretion of the subclass implementer).

Args
input_shape Instance of TensorShape, or list of instances of TensorShape if the layer expects a list of inputs (one instance per input).

build_from_config

build_from_config(
    config
)

Builds the layer's states with the supplied config dict.

By default, this method calls the build(config["input_shape"]) method, which creates weights based on the layer's input shape in the supplied config. If your config contains other information needed to load the layer's state, you should override this method.

Args
config Dict containing the input shape associated with this layer.

compute_mask

compute_mask(
    inputs, mask=None
)

Computes an output mask tensor.

Args
inputs Tensor or list of tensors.
mask Tensor or list of tensors.

Returns
None or a tensor (or list of tensors, one per output tensor of the layer).

compute_output_shape

compute_output_shape(
    input_shape
)

Computes the output shape of the layer.

This method will cause the layer's state to be built, if that has not happened before. This requires that the layer will later be used with inputs that match the input shape provided here.

Args
input_shape Shape tuple (tuple of integers) or tf.TensorShape, or structure of shape tuples / tf.TensorShape instances (one per output tensor of the layer). Shape tuples can include None for free dimensions, instead of an integer.

Returns
A tf.TensorShape instance or structure of tf.TensorShape instances.

count_params

count_params()

Count the total number of scalars composing the weights.

Returns
An integer count.

Raises
ValueError if the layer isn't yet built (in which case its weights aren't yet defined).

from_config

@classmethod
from_config(
    config
)

Creates a layer from its config.

This method is the reverse of get_config, capable of instantiating the same layer from the config dictionary. It does not handle layer connectivity (handled by Network), nor weights (handled by set_weights).

Args
config A Python dictionary, typically the output of get_config.

Returns
A layer instance.

get_build_config

get_build_config()

Returns a dictionary with the layer's input shape.

This method returns a config dict that can be used by build_from_config(config) to create all states (e.g. Variables and Lookup tables) needed by the layer.

By default, the config only contains the input shape that the layer was built with. If you're writing a custom layer that creates state in an unusual way, you should override this method to make sure this state is already created when Keras attempts to load its value upon model loading.

Returns
A dict containing the input shape associated with the layer.

get_config

get_config()

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Note that get_config() does not guarantee to return a fresh copy of dict every time it is called. The callers should make a copy of the returned dict if they want to modify it.

Returns
Python dictionary.

get_weights

get_weights()

Returns the current weights of the layer, as NumPy arrays.

The weights of a layer represent the state of the layer. This function returns both trainable and non-trainable weight values associated with this layer as a list of NumPy arrays, which can in turn be used to load state into similarly parameterized layers.

For example, a Dense layer returns a list of two values: the kernel matrix and the bias vector. These can be used to set the weights of another Dense layer:

layer_a = tf.keras.layers.Dense(1,
  kernel_initializer=tf.constant_initializer(1.))
a_out = layer_a(tf.convert_to_tensor([[1., 2., 3.]]))
layer_a.get_weights()
[array([[1.],
       [1.],
       [1.]], dtype=float32), array([0.], dtype=float32)]
layer_b = tf.keras.layers.Dense(1,
  kernel_initializer=tf.constant_initializer(2.))
b_out = layer_b(tf.convert_to_tensor([[10., 20., 30.]]))
layer_b.get_weights()
[array([[2.],
       [2.],
       [2.]], dtype=float32), array([0.], dtype=float32)]
layer_b.set_weights(layer_a.get_weights())
layer_b.get_weights()
[array([[1.],
       [1.],
       [1.]], dtype=float32), array([0.], dtype=float32)]

Returns
Weights values as a list of NumPy arrays.

load_own_variables

load_own_variables(
    store
)

Loads the state of the layer.

You can override this method to take full control of how the state of the layer is loaded upon calling keras.models.load_model().

Args
store Dict from which the state of the model will be loaded.

save_own_variables

save_own_variables(
    store
)

Saves the state of the layer.

You can override this method to take full control of how the state of the layer is saved upon calling model.save().

Args
store Dict where the state of the model will be saved.

set_weights

set_weights(
    weights
)

Sets the weights of the layer, from NumPy arrays.

The weights of a layer represent the state of the layer. This function sets the weight values from numpy arrays. The weight values should be passed in the order they are created by the layer. Note that the layer's weights must be instantiated before calling this function, by calling the layer.

For example, a Dense layer returns a list of two values: the kernel matrix and the bias vector. These can be used to set the weights of another Dense layer:

layer_a = tf.keras.layers.Dense(1,
  kernel_initializer=tf.constant_initializer(1.))
a_out = layer_a(tf.convert_to_tensor([[1., 2., 3.]]))
layer_a.get_weights()
[array([[1.],
       [1.],
       [1.]], dtype=float32), array([0.], dtype=float32)]
layer_b = tf.keras.layers.Dense(1,
  kernel_initializer=tf.constant_initializer(2.))
b_out = layer_b(tf.convert_to_tensor([[10., 20., 30.]]))
layer_b.get_weights()
[array([[2.],
       [2.],
       [2.]], dtype=float32), array([0.], dtype=float32)]
layer_b.set_weights(layer_a.get_weights())
layer_b.get_weights()
[array([[1.],
       [1.],
       [1.]], dtype=float32), array([0.], dtype=float32)]

Args
weights a list of NumPy arrays. The number of arrays and their shape must match number of the dimensions of the weights of the layer (i.e. it should match the output of get_weights).

Raises
ValueError If the provided weights list does not match the layer's specifications.

__call__

__call__(
    *args, **kwargs
)

Wraps call, applying pre- and post-processing steps.

Args
*args Positional arguments to be passed to self.call.
**kwargs Keyword arguments to be passed to self.call.

Returns
Output tensor(s).

Note

  • The following optional keyword arguments are reserved for specific uses:
    • training: Boolean scalar tensor of Python boolean indicating whether the call is meant for training or inference.
    • mask: Boolean input mask.
  • If the layer's call method takes a mask argument (as some Keras layers do), its default value will be set to the mask generated for inputs by the previous layer (if input did come from a layer that generated a corresponding mask, i.e. if it came from a Keras layer with masking support.
  • If the layer is not built, the method will call build.

Raises
ValueError if the layer's call method returns None (an invalid value).
RuntimeError if super().__init__() was not called in the constructor.