Merge pull request #2240 from mozilla/cudnnrnn

Add CuDNN RNN support

DeepSpeech.py

@@ -8,12 +8,12 @@ import sys
 LOG_LEVEL_INDEX = sys.argv.index('--log_level') + 1 if '--log_level' in sys.argv else 0
 os.environ['TF_CPP_MIN_LOG_LEVEL'] = sys.argv[LOG_LEVEL_INDEX] if 0 < LOG_LEVEL_INDEX < len(sys.argv) else '3'
 
-import time
 import numpy as np
 import progressbar
 import shutil
 import tensorflow as tf
 import tensorflow.compat.v1 as tfv1
+import time
 
 from datetime import datetime
 from ds_ctcdecoder import ctc_beam_search_decoder, Scorer
@@ -79,32 +79,64 @@ def dense(name, x, units, dropout_rate=None, relu=True):
 
 
 def rnn_impl_lstmblockfusedcell(x, seq_length, previous_state, reuse):
-    # Forward direction cell:
+    with tfv1.variable_scope('cudnn_lstm/rnn/multi_rnn_cell/cell_0'):
-    fw_cell = tf.contrib.rnn.LSTMBlockFusedCell(Config.n_cell_dim, reuse=reuse)
+        fw_cell = tf.contrib.rnn.LSTMBlockFusedCell(Config.n_cell_dim,
+                                                    reuse=reuse,
+                                                    name='cudnn_compatible_lstm_cell')
 
-    output, output_state = fw_cell(inputs=x,
+        output, output_state = fw_cell(inputs=x,
-                                   dtype=tf.float32,
+                                       dtype=tf.float32,
-                                   sequence_length=seq_length,
+                                       sequence_length=seq_length,
-                                   initial_state=previous_state)
+                                       initial_state=previous_state)
 
     return output, output_state
 
 
+def rnn_impl_cudnn_rnn(x, seq_length, previous_state, _):
+    assert previous_state is None # 'Passing previous state not supported with CuDNN backend'
+
+    # 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:
+        # Forward direction cell:
+        fw_cell = tf.contrib.cudnn_rnn.CudnnLSTM(num_layers=1,
+                                                 num_units=Config.n_cell_dim,
+                                                 input_mode='linear_input',
+                                                 direction='unidirectional',
+                                                 dtype=tf.float32)
+        rnn_impl_cudnn_rnn.cell = fw_cell
+
+    output, output_state = rnn_impl_cudnn_rnn.cell(inputs=x,
+                                                   sequence_lengths=seq_length)
+
+    return output, output_state
+
+rnn_impl_cudnn_rnn.cell = None
+
+
 def rnn_impl_static_rnn(x, seq_length, previous_state, reuse):
-    # Forward direction cell:
+    with tfv1.variable_scope('cudnn_lstm/rnn/multi_rnn_cell'):
-    fw_cell = tf.nn.rnn_cell.LSTMCell(Config.n_cell_dim, reuse=reuse)
+        # Forward direction cell:
+        fw_cell = tfv1.nn.rnn_cell.LSTMCell(Config.n_cell_dim,
+                                            reuse=reuse,
+                                            name='cudnn_compatible_lstm_cell')
 
-    # Split rank N tensor into list of rank N-1 tensors
+        # Split rank N tensor into list of rank N-1 tensors
-    x = [x[l] for l in range(x.shape[0])]
+        x = [x[l] for l in range(x.shape[0])]
 
-    # We parametrize the RNN implementation as the training and inference graph
+        output, output_state = tfv1.nn.static_rnn(cell=fw_cell,
-    # need to do different things here.
+                                                  inputs=x,
-    output, output_state = tf.nn.static_rnn(cell=fw_cell,
+                                                  sequence_length=seq_length,
-                                            inputs=x,
+                                                  initial_state=previous_state,
-                                            initial_state=previous_state,
+                                                  dtype=tf.float32,
-                                            dtype=tf.float32,
+                                                  scope='cell_0')
-                                            sequence_length=seq_length)
+
-    output = tf.concat(output, 0)
+        output = tf.concat(output, 0)
 
     return output, output_state
 
@@ -183,8 +215,13 @@ def calculate_mean_edit_distance_and_loss(iterator, dropout, reuse):
     # Obtain the next batch of data
     (batch_x, batch_seq_len), batch_y = iterator.get_next()
 
+    if FLAGS.use_cudnn_rnn:
+        rnn_impl = rnn_impl_cudnn_rnn
+    else:
+        rnn_impl = rnn_impl_lstmblockfusedcell
+
     # Calculate the logits of the batch
-    logits, _ = create_model(batch_x, batch_seq_len, dropout, reuse=reuse)
+    logits, _ = create_model(batch_x, batch_seq_len, dropout, reuse=reuse, rnn_impl=rnn_impl)
 
     # Compute the CTC loss using TensorFlow's `ctc_loss`
     total_loss = tfv1.nn.ctc_loss(labels=batch_y, inputs=logits, sequence_length=batch_seq_len)
@@ -573,7 +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_state = previous_state_c = previous_state_h = 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')
@@ -632,7 +669,7 @@ def create_inference_graph(batch_size=1, n_steps=16, tflite=False):
     }
 
     if not FLAGS.export_tflite:
-        inputs.update({'input_lengths': seq_length})
+        inputs['input_lengths'] = seq_length
 
     outputs = {
         'outputs': logits,
@@ -659,20 +696,8 @@ def export():
     output_names_ops = [op.name for op in outputs.values() if isinstance(op, Operation)]
     output_names = ",".join(output_names_tensors + output_names_ops)
 
-    mapping = None
+    # Create a saver using variables from the above newly created graph
-    if FLAGS.export_tflite:
+    saver = tfv1.train.Saver()
-        # Create a saver using variables from the above newly created graph
-        # Training graph uses LSTMFusedCell, but the TFLite inference graph uses
-        # 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/', 'lstm_fused_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

@@ -51,9 +51,11 @@ def create_flags():
 
     f.DEFINE_integer('export_batch_size', 1, 'number of elements per batch on the exported graph')
 
-    # Performance(UNSUPPORTED)
+    # Performance
-    f.DEFINE_integer('inter_op_parallelism_threads', 0, 'number of inter-op parallelism threads - see tf.ConfigProto for more details')
+
-    f.DEFINE_integer('intra_op_parallelism_threads', 0, 'number of intra-op parallelism threads - see tf.ConfigProto for more details')
+    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. 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