Address review comments

DeepSpeech.py

@@ -79,7 +79,7 @@ def dense(name, x, units, dropout_rate=None, relu=True):
 
 
 def rnn_impl_lstmblockfusedcell(x, seq_length, previous_state, reuse):
-    with tf.variable_scope('cudnn_lstm/rnn/multi_rnn_cell/cell_0'):
+    with tfv1.variable_scope('cudnn_lstm/rnn/multi_rnn_cell/cell_0'):
         fw_cell = tf.contrib.rnn.LSTMBlockFusedCell(Config.n_cell_dim,
                                                     reuse=reuse,
                                                     name='cudnn_compatible_lstm_cell')
@@ -95,14 +95,20 @@ def rnn_impl_lstmblockfusedcell(x, seq_length, previous_state, reuse):
 def rnn_impl_cudnn_rnn(x, seq_length, previous_state, _):
     assert previous_state is None # 'Passing previous state not supported with CuDNN backend'
 
-    # Forward direction cell:
+    # Hack: CudnnLSTM works similarly to Keras layers in that when you instantiate
+    # the object it creates the variables, and then you just call it several times
+    # to enable variable re-use. Because all of our code is structure in an old
+    # school TensorFlow structure where you can just call tf.get_variable again with
+    # reuse=True to reuse variables, we can't easily make use of the object oriented
+    # way CudnnLSTM is implemented, so we save a singleton instance in the function,
+    # emulating a static function variable.
     if not rnn_impl_cudnn_rnn.cell:
-        with tf.variable_scope('rnn'):
+        # Forward direction cell:
-            fw_cell = tf.contrib.cudnn_rnn.CudnnLSTM(num_layers=1,
+        fw_cell = tf.contrib.cudnn_rnn.CudnnLSTM(num_layers=1,
-                                                     num_units=Config.n_cell_dim,
+                                                 num_units=Config.n_cell_dim,
-                                                     input_mode='linear_input',
+                                                 input_mode='linear_input',
-                                                     direction='unidirectional',
+                                                 direction='unidirectional',
-                                                     dtype=tf.float32)
+                                                 dtype=tf.float32)
         rnn_impl_cudnn_rnn.cell = fw_cell
 
     output, output_state = rnn_impl_cudnn_rnn.cell(inputs=x,
@@ -110,18 +116,11 @@ def rnn_impl_cudnn_rnn(x, seq_length, previous_state, _):
 
     return output, output_state
 
-# Hack: CudnnLSTM works similarly to Keras layers in that when you instantiate
-# the object it creates the variables, and then you just call it several times
-# to enable variable re-use. Because all of our code is structure in an old
-# school TensorFlow structure where you can just call tf.get_variable again with
-# reuse=True to reuse variables, we can't easily make use of the object oriented
-# way CudnnLSTM is implemented, so we save a singleton instance in the function,
-# emulating a static function variable.
 rnn_impl_cudnn_rnn.cell = None
 
 
 def rnn_impl_static_rnn(x, seq_length, previous_state, reuse):
-    with tf.variable_scope('cudnn_lstm/rnn/multi_rnn_cell'):
+    with tfv1.variable_scope('cudnn_lstm/rnn/multi_rnn_cell'):
         # Forward direction cell:
         fw_cell = tfv1.nn.rnn_cell.LSTMCell(Config.n_cell_dim,
                                             reuse=reuse,
@@ -611,7 +610,7 @@ def create_inference_graph(batch_size=1, n_steps=16, tflite=False):
 
     if batch_size <= 0:
         # no state management since n_step is expected to be dynamic too (see below)
-        previous_states = None
+        previous_state = None
     else:
         previous_state_c = tfv1.placeholder(tf.float32, [batch_size, Config.n_cell_dim], name='previous_state_c')
         previous_state_h = tfv1.placeholder(tf.float32, [batch_size, Config.n_cell_dim], name='previous_state_h')
@@ -698,16 +697,7 @@ def export():
     output_names = ",".join(output_names_tensors + output_names_ops)
 
     # Create a saver using variables from the above newly created graph
-    # Training graph uses LSTMFusedCell, but the TFLite inference graph uses
+    saver = tfv1.train.Saver()
-    # a static RNN with a normal cell, so we need to rewrite the names to
-    # match the training weights when restoring.
-    def fixup(name):
-        if name.startswith('rnn/lstm_cell/'):
-            return name.replace('rnn/lstm_cell/', 'rnn/cudnn_compatible_lstm_cell/')
-        return name
-
-    mapping = {fixup(v.op.name): v for v in tf.global_variables()}
-    saver = tfv1.train.Saver(mapping)
 
     # Restore variables from training checkpoint
     checkpoint = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)

util/flags.py

@@ -55,7 +55,7 @@ def create_flags():
 
     f.DEFINE_integer('inter_op_parallelism_threads', 0, 'number of inter-op parallelism threads - see tf.ConfigProto for more details. USE OF THIS FLAG IS UNSUPPORTED')
     f.DEFINE_integer('intra_op_parallelism_threads', 0, 'number of intra-op parallelism threads - see tf.ConfigProto for more details. USE OF THIS FLAG IS UNSUPPORTED')
-    f.DEFINE_boolean('use_cudnn_rnn', False, 'use CuDNN RNN backend for training on GPU')
+    f.DEFINE_boolean('use_cudnn_rnn', False, 'use CuDNN RNN backend for training on GPU. Note that checkpoints created with this flag can only be used with CuDNN RNN, i.e. fine tuning on a CPU device will not work')
 
     # Sample limits