[RELNOTES] Refactor RNNs to rely on atomic cells. (#7943)

* Refactor RNNs to rely on atomic cells.

* Add RNN docstrings back.

* Fix Theano/CNTK RNN dropout

* Disable dropout in CNTK dynamic RNNs.

* Standardize input dropout masks in RNNs.

* Skip RNN dropout test for CNTK.

* Remove legacy constraints

* Increase stacked RNN test coverage

keras/backend/cntk_backend.py

@@ -1191,6 +1191,8 @@ def _static_rnn(step_function, inputs, initial_states,
     shape = int_shape(inputs)
     dims = len(shape)
 
+    uses_learning_phase = False
+
     if dims < 3:
         raise ValueError('Input should be at least 3D.')
 
@@ -1226,6 +1228,8 @@ def _static_rnn(step_function, inputs, initial_states,
 
             output, new_states = step_function(
                 current, tuple(states) + tuple(constants))
+            if getattr(output, '_uses_learning_phase', False):
+                uses_learning_phase = True
 
             if mask is not None:
                 mask_slice = C.ops.slice(mask, time_axis, i, i + 1)
@@ -1254,6 +1258,8 @@ def _static_rnn(step_function, inputs, initial_states,
 
             output, new_states = step_function(
                 current, tuple(states) + tuple(constants))
+            if getattr(output, '_uses_learning_phase', False):
+                uses_learning_phase = True
 
             if mask is not None:
                 mask_slice = C.ops.slice(mask, time_axis, i, i + 1)
@@ -1285,6 +1291,7 @@ def _static_rnn(step_function, inputs, initial_states,
         last_output = outputs[i]
         i += 1
 
+    last_output._uses_learning_phase = uses_learning_phase
     return last_output, final_output, states
 
 
@@ -1295,6 +1302,9 @@ def rnn(step_function, inputs, initial_states,
     shape = int_shape(inputs)
     dims = len(shape)
 
+    global uses_learning_phase
+    uses_learning_phase = False
+
     if dims < 3:
         raise ValueError('CNTK Backend: the input of rnn has only rank %d '
                          'Need at least rank 3 to run RNN.' % dims)
@@ -1370,6 +1380,11 @@ def _recurrence(x, states, m):
                 past_values.append(C.sequence.past_value(p, s))
             new_output, new_states = step_function(
                 x, tuple(past_values) + tuple(constants))
+
+            if getattr(new_output, '_uses_learning_phase', False):
+                global uses_learning_phase
+                uses_learning_phase = True
+
             if m is not None:
                 new_states = [C.element_select(m, n, s) for n, s in zip(new_states, past_values)]
             n_s = []
@@ -1398,6 +1413,7 @@ def _recurrence(x, states, m):
         else:
             f_stats.append(l_s)
 
+    last_output._uses_learning_phase = uses_learning_phase
     return last_output, final_output, f_stats
 
 

keras/backend/tensorflow_backend.py

@@ -2378,6 +2378,9 @@ def rnn(step_function, inputs, initial_states,
     if constants is None:
         constants = []
 
+    global uses_learning_phase
+    uses_learning_phase = False
+
     if unroll:
         if not inputs.get_shape()[0]:
             raise ValueError('Unrolling requires a '
@@ -2397,6 +2400,8 @@ def rnn(step_function, inputs, initial_states,
 
             for inp, mask_t in zip(input_list, mask_list):
                 output, new_states = step_function(inp, states + constants)
+                if getattr(output, '_uses_learning_phase', False):
+                    uses_learning_phase = True
 
                 # tf.where needs its condition tensor
                 # to be the same shape as its two
@@ -2435,6 +2440,8 @@ def rnn(step_function, inputs, initial_states,
         else:
             for inp in input_list:
                 output, states = step_function(inp, states + constants)
+                if getattr(output, '_uses_learning_phase', False):
+                    uses_learning_phase = True
                 successive_outputs.append(output)
                 successive_states.append(states)
             last_output = successive_outputs[-1]
@@ -2493,6 +2500,9 @@ def _step(time, output_ta_t, *states):
                 output, new_states = step_function(current_input,
                                                    tuple(states) +
                                                    tuple(constants))
+                if getattr(output, '_uses_learning_phase', False):
+                    global uses_learning_phase
+                    uses_learning_phase = True
                 for state, new_state in zip(states, new_states):
                     new_state.set_shape(state.get_shape())
                 tiled_mask_t = tf.tile(mask_t,
@@ -2517,6 +2527,9 @@ def _step(time, output_ta_t, *states):
                 output, new_states = step_function(current_input,
                                                    tuple(states) +
                                                    tuple(constants))
+                if getattr(output, '_uses_learning_phase', False):
+                    global uses_learning_phase
+                    uses_learning_phase = True
                 for state, new_state in zip(states, new_states):
                     new_state.set_shape(state.get_shape())
                 output_ta_t = output_ta_t.write(time, output)
@@ -2537,6 +2550,7 @@ def _step(time, output_ta_t, *states):
 
     axes = [1, 0] + list(range(2, len(outputs.get_shape())))
     outputs = tf.transpose(outputs, axes)
+    last_output._uses_learning_phase = uses_learning_phase
     return last_output, outputs, new_states
 
 

keras/backend/theano_backend.py

@@ -1307,6 +1307,9 @@ def rnn(step_function, inputs, initial_states,
     if constants is None:
         constants = []
 
+    global uses_learning_phase
+    uses_learning_phase = False
+
     if mask is not None:
         if mask.ndim == ndim - 1:
             mask = expand_dims(mask)
@@ -1323,6 +1326,8 @@ def rnn(step_function, inputs, initial_states,
             states = initial_states
             for i in indices:
                 output, new_states = step_function(inputs[i], states + constants)
+                if getattr(output, '_uses_learning_phase', False):
+                    uses_learning_phase = True
 
                 if len(successive_outputs) == 0:
                     prev_output = zeros_like(output)
@@ -1352,6 +1357,9 @@ def rnn(step_function, inputs, initial_states,
 
             def _step(inputs, mask, output_tm1, *states):
                 outputs, new_states = step_function(inputs, states)
+                if getattr(outputs, '_uses_learning_phase', False):
+                    global uses_learning_phase
+                    uses_learning_phase = True
                 # output previous output if masked.
                 outputs = T.switch(mask, outputs, output_tm1)
                 return_states = []
@@ -1384,6 +1392,8 @@ def _step(inputs, mask, output_tm1, *states):
             states = initial_states
             for i in indices:
                 outputs, states = step_function(inputs[i], states + constants)
+                if getattr(outputs, '_uses_learning_phase', False):
+                    uses_learning_phase = True
                 successive_outputs.append(outputs)
                 successive_states.append(states)
             outputs = T.stack(*successive_outputs)
@@ -1394,9 +1404,13 @@ def _step(inputs, mask, output_tm1, *states):
         else:
             def _step(inputs, *states):
                 outputs, new_states = step_function(inputs, states)
+                if getattr(outputs, '_uses_learning_phase', False):
+                    global uses_learning_phase
+                    uses_learning_phase = True
                 return [outputs] + new_states
 
-            # Theano likes to make shape==1 dimensions in the initial states (outputs_info) broadcastable
+            # Theano likes to make shape==1 dimensions
+            # in the initial states (outputs_info) broadcastable
             if len(initial_states) > 0:
                 initial_states[0] = T.unbroadcast(initial_states[0], 1)
 
@@ -1421,6 +1435,7 @@ def _step(inputs, *states):
     axes = [1, 0] + list(range(2, outputs.ndim))
     outputs = outputs.dimshuffle(axes)
     states = [T.squeeze(state[-1]) for state in states]
+    last_output._uses_learning_phase = uses_learning_phase
     return last_output, outputs, states