Recurrent Attention API Additions (#7980)

* Added support for passing external constants to RNN, which will pass them on to the cell

* Added class for allowing functional composition of RNN Cells, supporting constants

* put back accidentally commented out recurrent tests

* added basic example of functional cell

* new class AttentionRNN

* restored RNN layer

* renamed constants to attended in FunctionRNNCell, avoided duplicating outputs in wrapped model

* minor clean-up of docs

* Minor cleanup & improvments in docs, fixed PEP breaking formatting in attention test

* removed FunctionalRNNCell and AttentionRNN, added back support for constants in RNN

* fixed PEP8 violations

* fixed minor review comments

* added test case for when both inital_state and constants are passed to RNN.__call__

Author: andhus

keras/layers/recurrent.py

@@ -1,7 +1,6 @@
 # -*- coding: utf-8 -*-
 from __future__ import absolute_import
 import numpy as np
-import functools
 import warnings
 
 from .. import backend as K
@@ -200,7 +199,9 @@ class RNN(Layer):
     # Arguments
         cell: A RNN cell instance. A RNN cell is a class that has:
             - a `call(input_at_t, states_at_t)` method, returning
-                `(output_at_t, states_at_t_plus_1)`.
+                `(output_at_t, states_at_t_plus_1)`. The call method of the
+                cell can also take the optional argument `constants`, see
+                section "Note on passing external constants" below.
             - a `state_size` attribute. This can be a single integer
                 (single state) in which case it is
                 the size of the recurrent state
@@ -292,6 +293,14 @@ class RNN(Layer):
         `states` should be a numpy array or list of numpy arrays representing
         the initial state of the RNN layer.
 
+    # Note on passing external constants to RNNs
+        You can pass "external" constants to the cell using the `constants`
+        keyword argument of `RNN.__call__` (as well as `RNN.call`) method. This
+        requires that the `cell.call` method accepts the same keyword argument
+        `constants`. Such constants can be used to condition the cell
+        transformation on additional static inputs (not changing over time),
+        a.k.a. an attention mechanism.
+
     # Examples
 
     ```python
@@ -363,12 +372,10 @@ def __init__(self, cell,
 
         self.supports_masking = True
         self.input_spec = [InputSpec(ndim=3)]
-        if hasattr(self.cell.state_size, '__len__'):
-            self.state_spec = [InputSpec(shape=(None, dim))
-                               for dim in self.cell.state_size]
-        else:
-            self.state_spec = InputSpec(shape=(None, self.cell.state_size))
+        self.state_spec = None
         self._states = None
+        self.constants_spec = None
+        self._num_constants = None
 
     @property
     def states(self):
@@ -415,19 +422,46 @@ def compute_mask(self, inputs, mask):
             return output_mask
 
     def build(self, input_shape):
+        # Note input_shape will be list of shapes of initial states and
+        # constants if these are passed in __call__.
+        if self._num_constants is not None:
+            constants_shape = input_shape[-self._num_constants:]
+        else:
+            constants_shape = None
+
         if isinstance(input_shape, list):
             input_shape = input_shape[0]
 
         batch_size = input_shape[0] if self.stateful else None
         input_dim = input_shape[-1]
         self.input_spec[0] = InputSpec(shape=(batch_size, None, input_dim))
 
-        if self.stateful:
-            self.reset_states()
-
+        # allow cell (if layer) to build before we set or validate state_spec
         if isinstance(self.cell, Layer):
             step_input_shape = (input_shape[0],) + input_shape[2:]
-            self.cell.build(step_input_shape)
+            if constants_shape is not None:
+                self.cell.build([step_input_shape] + constants_shape)
+            else:
+                self.cell.build(step_input_shape)
+
+        # set or validate state_spec
+        if hasattr(self.cell.state_size, '__len__'):
+            state_size = list(self.cell.state_size)
+        else:
+            state_size = [self.cell.state_size]
+
+        if self.state_spec is not None:
+            # initial_state was passed in call, check compatibility
+            if not [spec.shape[-1] for spec in self.state_spec] == state_size:
+                raise ValueError(
+                    'an initial_state was passed that is not compatible with'
+                    ' cell.state_size, state_spec: {}, cell.state_size:'
+                    ' {}'.format(self.state_spec, self.cell.state_size))
+        else:
+            self.state_spec = [InputSpec(shape=(None, dim))
+                               for dim in state_size]
+        if self.stateful:
+            self.reset_states()
 
     def get_initial_state(self, inputs):
         # build an all-zero tensor of shape (samples, output_dim)
@@ -440,62 +474,65 @@ def get_initial_state(self, inputs):
         else:
             return [K.tile(initial_state, [1, self.cell.state_size])]
 
-    def __call__(self, inputs, initial_state=None, **kwargs):
-        # If there are multiple inputs, then
-        # they should be the main input and `initial_state`
-        # e.g. when loading model from file
-        if isinstance(inputs, (list, tuple)) and len(inputs) > 1 and initial_state is None:
-            initial_state = inputs[1:]
-            inputs = inputs[0]
+    def __call__(self, inputs, initial_state=None, constants=None, **kwargs):
+        inputs, initial_state, constants = self._standardize_args(
+            inputs, initial_state, constants)
 
-        # If `initial_state` is specified,
-        # and if it a Keras tensor,
-        # then add it to the inputs and temporarily
-        # modify the input spec to include the state.
-        if initial_state is None:
+        if initial_state is None and constants is None:
             return super(RNN, self).__call__(inputs, **kwargs)
 
-        if not isinstance(initial_state, (list, tuple)):
-            initial_state = [initial_state]
+        # If any of `initial_state` or `constants` are specified and are Keras
+        # tensors, then add them to the inputs and temporarily modify the
+        # input_spec to include them.
 
-        is_keras_tensor = hasattr(initial_state[0], '_keras_history')
-        for tensor in initial_state:
+        additional_inputs = []
+        additional_specs = []
+        if initial_state is not None:
+            kwargs['initial_state'] = initial_state
+            additional_inputs += initial_state
+            self.state_spec = [InputSpec(shape=K.int_shape(state))
+                               for state in initial_state]
+            additional_specs += self.state_spec
+        if constants is not None:
+            kwargs['constants'] = constants
+            additional_inputs += constants
+            self.constants_spec = [InputSpec(shape=K.int_shape(constant))
+                                   for constant in constants]
+            self._num_constants = len(constants)
+            additional_specs += self.constants_spec
+        # at this point additional_inputs cannot be empty
+        is_keras_tensor = hasattr(additional_inputs[0], '_keras_history')
+        for tensor in additional_inputs:
             if hasattr(tensor, '_keras_history') != is_keras_tensor:
-                raise ValueError('The initial state of an RNN layer cannot be'
-                                 ' specified with a mix of Keras tensors and'
-                                 ' non-Keras tensors')
+                raise ValueError('The initial state or constants of an RNN'
+                                 ' layer cannot be specified with a mix of'
+                                 ' Keras tensors and non-Keras tensors')
 
         if is_keras_tensor:
-            # Compute the full input spec, including state
-            input_spec = self.input_spec
-            state_spec = self.state_spec
-            if not isinstance(input_spec, list):
-                input_spec = [input_spec]
-            if not isinstance(state_spec, list):
-                state_spec = [state_spec]
-            self.input_spec = input_spec + state_spec
-
-            # Compute the full inputs, including state
-            inputs = [inputs] + list(initial_state)
-
-            # Perform the call
-            output = super(RNN, self).__call__(inputs, **kwargs)
-
-            # Restore original input spec
-            self.input_spec = input_spec
+            # Compute the full input spec, including state and constants
+            full_input = [inputs] + additional_inputs
+            full_input_spec = self.input_spec + additional_specs
+            # Perform the call with temporarily replaced input_spec
+            original_input_spec = self.input_spec
+            self.input_spec = full_input_spec
+            output = super(RNN, self).__call__(full_input, **kwargs)
+            self.input_spec = original_input_spec
             return output
         else:
-            kwargs['initial_state'] = initial_state
             return super(RNN, self).__call__(inputs, **kwargs)
 
-    def call(self, inputs, mask=None, training=None, initial_state=None):
+    def call(self,
+             inputs,
+             mask=None,
+             training=None,
+             initial_state=None,
+             constants=None):
         # input shape: `(samples, time (padded with zeros), input_dim)`
         # note that the .build() method of subclasses MUST define
         # self.input_spec and self.state_spec with complete input shapes.
         if isinstance(inputs, list):
-            initial_state = inputs[1:]
             inputs = inputs[0]
-        elif initial_state is not None:
+        if initial_state is not None:
             pass
         elif self.stateful:
             initial_state = self.states
@@ -525,13 +562,27 @@ def call(self, inputs, mask=None, training=None, initial_state=None):
                              'the time dimension by passing a `shape` '
                              'or `batch_shape` argument to your Input layer.')
 
+        kwargs = {}
         if has_arg(self.cell.call, 'training'):
-            step = functools.partial(self.cell.call, training=training)
+            kwargs['training'] = training
+
+        if constants:
+            if not has_arg(self.cell.call, 'constants'):
+                raise ValueError('RNN cell does not support constants')
+
+            def step(inputs, states):
+                constants = states[-self._num_constants:]
+                states = states[:-self._num_constants]
+                return self.cell.call(inputs, states, constants=constants,
+                                      **kwargs)
         else:
-            step = self.cell.call
+            def step(inputs, states):
+                return self.cell.call(inputs, states, **kwargs)
+
         last_output, outputs, states = K.rnn(step,
                                              inputs,
                                              initial_state,
+                                             constants=constants,
                                              go_backwards=self.go_backwards,
                                              mask=mask,
                                              unroll=self.unroll,
@@ -560,6 +611,47 @@ def call(self, inputs, mask=None, training=None, initial_state=None):
         else:
             return output
 
+    def _standardize_args(self, inputs, initial_state, constants):
+        """Brings the arguments of `__call__` that can contain input tensors to
+        standard format.
+
+        When running a model loaded from file, the input tensors
+        `initial_state` and `constants` can be passed to `RNN.__call__` as part
+        of `inputs` instead of by the dedicated keyword arguments. This method
+        makes sure the arguments are separated and that `initial_state` and
+        `constants` are lists of tensors (or None).
+
+        # Arguments
+            inputs: tensor or list/tuple of tensors
+            initial_state: tensor or list of tensors or None
+            constants: tensor or list of tensors or None
+
+        # Returns
+            inputs: tensor
+            initial_state: list of tensors or None
+            constants: list of tensors or None
+        """
+        if isinstance(inputs, list):
+            assert initial_state is None and constants is None
+            if self._num_constants is not None:
+                constants = inputs[-self._num_constants:]
+                inputs = inputs[:-self._num_constants]
+            if len(inputs) > 1:
+                initial_state = inputs[1:]
+            inputs = inputs[0]
+
+        def to_list_or_none(x):
+            if x is None or isinstance(x, list):
+                return x
+            if isinstance(x, tuple):
+                return list(x)
+            return [x]
+
+        initial_state = to_list_or_none(initial_state)
+        constants = to_list_or_none(constants)
+
+        return inputs, initial_state, constants
+
     def reset_states(self, states=None):
         if not self.stateful:
             raise AttributeError('Layer must be stateful.')
@@ -618,6 +710,9 @@ def get_config(self):
                   'go_backwards': self.go_backwards,
                   'stateful': self.stateful,
                   'unroll': self.unroll}
+        if self._num_constants is not None:
+            config['num_constants'] = self._num_constants
+
         cell_config = self.cell.get_config()
         config['cell'] = {'class_name': self.cell.__class__.__name__,
                           'config': cell_config}
@@ -629,7 +724,10 @@ def from_config(cls, config, custom_objects=None):
         from . import deserialize as deserialize_layer
         cell = deserialize_layer(config.pop('cell'),
                                  custom_objects=custom_objects)
-        return cls(cell, **config)
+        num_constants = config.pop('num_constants', None)
+        layer = cls(cell, **config)
+        layer._num_constants = num_constants
+        return layer
 
     @property
     def trainable_weights(self):