Update deep speech model with pure tensorflow API implementation (#4730)

* update

* update

* update

* update

* update

research/deep_speech/README.md

+# DeepSpeech2 Model
+## Overview
+This is an implementation of the [DeepSpeech2](https://arxiv.org/pdf/1512.02595.pdf) model. Current implementation is based on the code from the authors' [DeepSpeech code](https://github.com/PaddlePaddle/DeepSpeech) and the implementation in the [MLPerf Repo](https://github.com/mlperf/reference/tree/master/speech_recognition).
+
+DeepSpeech2 is an end-to-end deep neural network for automatic speech
+recognition (ASR). It consists of 2 convolutional layers, 5 bidirectional RNN
+layers and a fully connected layer. The feature in use is linear spectrogram
+extracted from audio input. The network uses Connectionist Temporal Classification [CTC](https://www.cs.toronto.edu/~graves/icml_2006.pdf) as the loss function.
+
+## Dataset
+The [OpenSLR LibriSpeech Corpus](http://www.openslr.org/12/) are used for model training and evaluation.
+
+The training data is a combination of train-clean-100 and train-clean-360 (~130k
+examples in total). The validation set is dev-clean which has 2.7K lines.
+The download script will preprocess the data into three columns: wav_filename,
+wav_filesize, transcript. data/dataset.py will parse the csv file and build a
+tf.data.Dataset object to feed data. Within each epoch (except for the
+first if sortagrad is enabled), the training data will be shuffled batch-wise.
+
+## Running Code
+
+### Configure Python path
+Add the top-level /models folder to the Python path with the command:
+```
+export PYTHONPATH="$PYTHONPATH:/path/to/models"
+```
+
+### Install dependencies
+
+First install shared dependencies before running the code. Issue the following command:
+```
+pip3 install -r requirements.txt
+```
+or
+```
+pip install -r requirements.txt
+```
+
+### Download and preprocess dataset
+To download the dataset, issue the following command:
+```
+python data/download.py
+```
+Arguments:
+  * `--data_dir`: Directory where to download and save the preprocessed data. By default, it is `/tmp/librispeech_data`.
+
+Use the `--help` or `-h` flag to get a full list of possible arguments.
+
+### Train and evaluate model
+To train and evaluate the model, issue the following command:
+```
+python deep_speech.py
+```
+Arguments:
+  * `--model_dir`: Directory to save model training checkpoints. By default, it is `/tmp/deep_speech_model/`.
+  * `--train_data_dir`: Directory of the training dataset.
+  * `--eval_data_dir`: Directory of the evaluation dataset.
+  * `--num_gpus`: Number of GPUs to use (specify -1 if you want to use all available GPUs).
+
+There are other arguments about DeepSpeech2 model and training/evaluation process. Use the `--help` or `-h` flag to get a full list of possible arguments with detailed descriptions.
+

research/deep_speech/data/dataset.py

@@ -17,13 +17,14 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import functools
-import multiprocessing
-
+import math
+import random
+# pylint: disable=g-bad-import-order
 import numpy as np
-import scipy.io.wavfile as wavfile
 from six.moves import xrange  # pylint: disable=redefined-builtin
+import soundfile
 import tensorflow as tf
+# pylint: enable=g-bad-import-order
 
 import data.featurizer as featurizer  # pylint: disable=g-bad-import-order
 
@@ -33,40 +34,37 @@ class AudioConfig(object):
 
   def __init__(self,
                sample_rate,
-               frame_length,
-               frame_step,
-               fft_length=None,
-               normalize=False,
-               spect_type="linear"):
+               window_ms,
+               stride_ms,
+               normalize=False):
     """Initialize the AudioConfig class.
 
     Args:
       sample_rate: an integer denoting the sample rate of the input waveform.
-      frame_length: an integer for the length of a spectrogram frame, in ms.
-      frame_step: an integer for the frame stride, in ms.
-      fft_length: an integer for the number of fft bins.
+      window_ms: an integer for the length of a spectrogram frame, in ms.
+      stride_ms: an integer for the frame stride, in ms.
       normalize: a boolean for whether apply normalization on the audio feature.
-      spect_type: a string for the type of spectrogram to be extracted.
     """
 
     self.sample_rate = sample_rate
-    self.frame_length = frame_length
-    self.frame_step = frame_step
-    self.fft_length = fft_length
+    self.window_ms = window_ms
+    self.stride_ms = stride_ms
     self.normalize = normalize
-    self.spect_type = spect_type
 
 
 class DatasetConfig(object):
   """Config class for generating the DeepSpeechDataset."""
 
-  def __init__(self, audio_config, data_path, vocab_file_path):
+  def __init__(self, audio_config, data_path, vocab_file_path, sortagrad):
     """Initialize the configs for deep speech dataset.
 
     Args:
       audio_config: AudioConfig object specifying the audio-related configs.
       data_path: a string denoting the full path of a manifest file.
       vocab_file_path: a string specifying the vocabulary file path.
+      sortagrad: a boolean, if set to true, audio sequences will be fed by
+                increasing length in the first training epoch, which will
+                expedite network convergence.
 
     Raises:
       RuntimeError: file path not exist.
@@ -77,6 +75,7 @@ class DatasetConfig(object):
     assert tf.gfile.Exists(vocab_file_path)
     self.data_path = data_path
     self.vocab_file_path = vocab_file_path
+    self.sortagrad = sortagrad
 
 
 def _normalize_audio_feature(audio_feature):
@@ -95,30 +94,23 @@ def _normalize_audio_feature(audio_feature):
   return normalized
 
 
-def _preprocess_audio(
-    audio_file_path, audio_sample_rate, audio_featurizer, normalize):
-  """Load the audio file in memory and compute spectrogram feature."""
-  tf.logging.info(
-      "Extracting spectrogram feature for {}".format(audio_file_path))
-  sample_rate, data = wavfile.read(audio_file_path)
-  assert sample_rate == audio_sample_rate
-  if data.dtype not in [np.float32, np.float64]:
-    data = data.astype(np.float32) / np.iinfo(data.dtype).max
+def _preprocess_audio(audio_file_path, audio_featurizer, normalize):
+  """Load the audio file and compute spectrogram feature."""
+  data, _ = soundfile.read(audio_file_path)
   feature = featurizer.compute_spectrogram_feature(
-      data, audio_featurizer.frame_length, audio_featurizer.frame_step,
-      audio_featurizer.fft_length)
+      data, audio_featurizer.sample_rate, audio_featurizer.stride_ms,
+      audio_featurizer.window_ms)
+  # Feature normalization
   if normalize:
     feature = _normalize_audio_feature(feature)
-  return feature
-
 
-def _preprocess_transcript(transcript, token_to_index):
-  """Process transcript as label features."""
-  return featurizer.compute_label_feature(transcript, token_to_index)
+  # Adding Channel dimension for conv2D input.
+  feature = np.expand_dims(feature, axis=2)
+  return feature
 
 
-def _preprocess_data(dataset_config, audio_featurizer, token_to_index):
-  """Generate a list of waveform, transcript pair.
+def _preprocess_data(file_path):
+  """Generate a list of tuples (wav_filename, wav_filesize, transcript).
 
   Each dataset file contains three columns: "wav_filename", "wav_filesize",
   and "transcript". This function parses the csv file and stores each example
@@ -127,42 +119,23 @@ def _preprocess_data(dataset_config, audio_featurizer, token_to_index):
   mini-batch have similar length.
 
   Args:
-    dataset_config: an instance of DatasetConfig.
-    audio_featurizer: an instance of AudioFeaturizer.
-    token_to_index: the mapping from character to its index
+    file_path: a string specifying the csv file path for a dataset.
 
   Returns:
-    features and labels array processed from the audio/text input.
+    A list of tuples (wav_filename, wav_filesize, transcript) sorted by
+    file_size.
   """
-
-  file_path = dataset_config.data_path
-  sample_rate = dataset_config.audio_config.sample_rate
-  normalize = dataset_config.audio_config.normalize
-
+  tf.logging.info("Loading data set {}".format(file_path))
   with tf.gfile.Open(file_path, "r") as f:
     lines = f.read().splitlines()
-  lines = [line.split("\t") for line in lines]
-  # Skip the csv header.
+  # Skip the csv header in lines[0].
   lines = lines[1:]
-  # Sort input data by the length of waveform.
+  # The metadata file is tab separated.
+  lines = [line.split("\t", 2) for line in lines]
+  # Sort input data by the length of audio sequence.
   lines.sort(key=lambda item: int(item[1]))
 
-  # Use multiprocessing for feature/label extraction
-  num_cores = multiprocessing.cpu_count()
-  pool = multiprocessing.Pool(processes=num_cores)
-
-  features = pool.map(
-      functools.partial(
-          _preprocess_audio, audio_sample_rate=sample_rate,
-          audio_featurizer=audio_featurizer, normalize=normalize),
-      [line[0] for line in lines])
-  labels = pool.map(
-      functools.partial(
-          _preprocess_transcript, token_to_index=token_to_index),
-      [line[2] for line in lines])
-
-  pool.terminate()
-  return features, labels
+  return [tuple(line) for line in lines]
 
 
 class DeepSpeechDataset(object):
@@ -178,22 +151,52 @@ class DeepSpeechDataset(object):
     # Instantiate audio feature extractor.
     self.audio_featurizer = featurizer.AudioFeaturizer(
         sample_rate=self.config.audio_config.sample_rate,
-        frame_length=self.config.audio_config.frame_length,
-        frame_step=self.config.audio_config.frame_step,
-        fft_length=self.config.audio_config.fft_length)
+        window_ms=self.config.audio_config.window_ms,
+        stride_ms=self.config.audio_config.stride_ms)
     # Instantiate text feature extractor.
     self.text_featurizer = featurizer.TextFeaturizer(
         vocab_file=self.config.vocab_file_path)
 
     self.speech_labels = self.text_featurizer.speech_labels
-    self.features, self.labels = _preprocess_data(
-        self.config,
-        self.audio_featurizer,
-        self.text_featurizer.token_to_idx
-    )
+    self.entries = _preprocess_data(self.config.data_path)
+    # The generated spectrogram will have 161 feature bins.
+    self.num_feature_bins = 161
+
 
-    self.num_feature_bins = (
-        self.features[0].shape[1] if len(self.features) else None)
+def batch_wise_dataset_shuffle(entries, epoch_index, sortagrad, batch_size):
+  """Batch-wise shuffling of the data entries.
+
+  Each data entry is in the format of (audio_file, file_size, transcript).
+  If epoch_index is 0 and sortagrad is true, we don't perform shuffling and
+  return entries in sorted file_size order. Otherwise, do batch_wise shuffling.
+
+  Args:
+    entries: a list of data entries.
+    epoch_index: an integer of epoch index
+    sortagrad: a boolean to control whether sorting the audio in the first
+      training epoch.
+    batch_size: an integer for the batch size.
+
+  Returns:
+    The shuffled data entries.
+  """
+  shuffled_entries = []
+  if epoch_index == 0 and sortagrad:
+    # No need to shuffle.
+    shuffled_entries = entries
+  else:
+    # Shuffle entries batch-wise.
+    max_buckets = int(math.floor(len(entries) / batch_size))
+    total_buckets = [i for i in xrange(max_buckets)]
+    random.shuffle(total_buckets)
+    shuffled_entries = []
+    for i in total_buckets:
+      shuffled_entries.extend(entries[i * batch_size : (i + 1) * batch_size])
+    # If the last batch doesn't contain enough batch_size examples,
+    # just append it to the shuffled_entries.
+    shuffled_entries.extend(entries[max_buckets * batch_size:])
+
+  return shuffled_entries
 
 
 def input_fn(batch_size, deep_speech_dataset, repeat=1):
@@ -207,49 +210,66 @@ def input_fn(batch_size, deep_speech_dataset, repeat=1):
   Returns:
     a tf.data.Dataset object for model to consume.
   """
-  features = deep_speech_dataset.features
-  labels = deep_speech_dataset.labels
+  # Dataset properties
+  data_entries = deep_speech_dataset.entries
   num_feature_bins = deep_speech_dataset.num_feature_bins
+  audio_featurizer = deep_speech_dataset.audio_featurizer
+  feature_normalize = deep_speech_dataset.config.audio_config.normalize
+  text_featurizer = deep_speech_dataset.text_featurizer
 
   def _gen_data():
-    for i in xrange(len(features)):
-      feature = np.expand_dims(features[i], axis=2)
-      input_length = [features[i].shape[0]]
-      label_length = [len(labels[i])]
-      yield {
-          "features": feature,
-          "labels": labels[i],
-          "input_length": input_length,
-          "label_length": label_length
-      }
+    """Dataset generator function."""
+    for audio_file, _, transcript in data_entries:
+      features = _preprocess_audio(
+          audio_file, audio_featurizer, feature_normalize)
+      labels = featurizer.compute_label_feature(
+          transcript, text_featurizer.token_to_index)
+      input_length = [features.shape[0]]
+      label_length = [len(labels)]
+      # Yield a tuple of (features, labels) where features is a dict containing
+      # all info about the actual data features.
+      yield (
+          {
+              "features": features,
+              "input_length": input_length,
+              "label_length": label_length
+          },
+          labels)
 
   dataset = tf.data.Dataset.from_generator(
       _gen_data,
-      output_types={
-          "features": tf.float32,
-          "labels": tf.int32,
-          "input_length": tf.int32,
-          "label_length": tf.int32
-      },
-      output_shapes={
-          "features": tf.TensorShape([None, num_feature_bins, 1]),
-          "labels": tf.TensorShape([None]),
-          "input_length": tf.TensorShape([1]),
-          "label_length": tf.TensorShape([1])
-      })
+      output_types=(
+          {
+              "features": tf.float32,
+              "input_length": tf.int32,
+              "label_length": tf.int32
+          },
+          tf.int32),
+      output_shapes=(
+          {
+              "features": tf.TensorShape([None, num_feature_bins, 1]),
+              "input_length": tf.TensorShape([1]),
+              "label_length": tf.TensorShape([1])
+          },
+          tf.TensorShape([None]))
+  )
 
   # Repeat and batch the dataset
   dataset = dataset.repeat(repeat)
+
   # Padding the features to its max length dimensions.
   dataset = dataset.padded_batch(
       batch_size=batch_size,
-      padded_shapes={
-          "features": tf.TensorShape([None, num_feature_bins, 1]),
-          "labels": tf.TensorShape([None]),
-          "input_length": tf.TensorShape([1]),
-          "label_length": tf.TensorShape([1])
-      })
+      padded_shapes=(
+          {
+              "features": tf.TensorShape([None, num_feature_bins, 1]),
+              "input_length": tf.TensorShape([1]),
+              "label_length": tf.TensorShape([1])
+          },
+          tf.TensorShape([None]))
+  )
 
   # Prefetch to improve speed of input pipeline.
-  dataset = dataset.prefetch(1)
+  dataset = dataset.prefetch(buffer_size=tf.contrib.data.AUTOTUNE)
   return dataset
+

research/deep_speech/data/featurizer.py

@@ -19,20 +19,51 @@ from __future__ import print_function
 
 import codecs
 import numpy as np
-from scipy import signal
 
 
-def compute_spectrogram_feature(waveform, frame_length, frame_step, fft_length):
-  """Compute the spectrograms for the input waveform."""
-  _, _, stft = signal.stft(
-      waveform,
-      nperseg=frame_length,
-      noverlap=frame_step,
-      nfft=fft_length)
+def compute_spectrogram_feature(samples, sample_rate, stride_ms=10.0,
+                                window_ms=20.0, max_freq=None, eps=1e-14):
+  """Compute the spectrograms for the input samples(waveforms).
 
-  # Perform transpose to set its shape as [time_steps, feature_num_bins]
-  spectrogram = np.transpose(np.absolute(stft), (1, 0))
-  return spectrogram
+  More about spectrogram computation, please refer to:
+  https://en.wikipedia.org/wiki/Short-time_Fourier_transform.
+  """
+  if max_freq is None:
+    max_freq = sample_rate / 2
+  if max_freq > sample_rate / 2:
+    raise ValueError("max_freq must not be greater than half of sample rate.")
+
+  if stride_ms > window_ms:
+    raise ValueError("Stride size must not be greater than window size.")
+
+  stride_size = int(0.001 * sample_rate * stride_ms)
+  window_size = int(0.001 * sample_rate * window_ms)
+
+  # Extract strided windows
+  truncate_size = (len(samples) - window_size) % stride_size
+  samples = samples[:len(samples) - truncate_size]
+  nshape = (window_size, (len(samples) - window_size) // stride_size + 1)
+  nstrides = (samples.strides[0], samples.strides[0] * stride_size)
+  windows = np.lib.stride_tricks.as_strided(
+      samples, shape=nshape, strides=nstrides)
+  assert np.all(
+      windows[:, 1] == samples[stride_size:(stride_size + window_size)])
+
+  # Window weighting, squared Fast Fourier Transform (fft), scaling
+  weighting = np.hanning(window_size)[:, None]
+  fft = np.fft.rfft(windows * weighting, axis=0)
+  fft = np.absolute(fft)
+  fft = fft**2
+  scale = np.sum(weighting**2) * sample_rate
+  fft[1:-1, :] *= (2.0 / scale)
+  fft[(0, -1), :] /= scale
+  # Prepare fft frequency list
+  freqs = float(sample_rate) / window_size * np.arange(fft.shape[0])
+
+  # Compute spectrogram feature
+  ind = np.where(freqs <= max_freq)[0][-1] + 1
+  specgram = np.log(fft[:ind, :] + eps)
+  return np.transpose(specgram, (1, 0))
 
 
 class AudioFeaturizer(object):
@@ -40,21 +71,18 @@ class AudioFeaturizer(object):
 
   def __init__(self,
                sample_rate=16000,
-               frame_length=25,
-               frame_step=10,
-               fft_length=None):
+               window_ms=20.0,
+               stride_ms=10.0):
     """Initialize the audio featurizer class according to the configs.
 
     Args:
       sample_rate: an integer specifying the sample rate of the input waveform.
-      frame_length: an integer for the length of a spectrogram frame, in ms.
-      frame_step: an integer for the frame stride, in ms.
-      fft_length: an integer for the number of fft bins.
+      window_ms: an integer for the length of a spectrogram frame, in ms.
+      stride_ms: an integer for the frame stride, in ms.
     """
-    self.frame_length = int(sample_rate * frame_length / 1e3)
-    self.frame_step = int(sample_rate * frame_step / 1e3)
-    self.fft_length = fft_length if fft_length else int(2**(np.ceil(
-        np.log2(self.frame_length))))
+    self.sample_rate = sample_rate
+    self.window_ms = window_ms
+    self.stride_ms = stride_ms
 
 
 def compute_label_feature(text, token_to_idx):
@@ -75,16 +103,16 @@ class TextFeaturizer(object):
     lines = []
     with codecs.open(vocab_file, "r", "utf-8") as fin:
       lines.extend(fin.readlines())
-    self.token_to_idx = {}
-    self.idx_to_token = {}
+    self.token_to_index = {}
+    self.index_to_token = {}
     self.speech_labels = ""
-    idx = 0
+    index = 0
     for line in lines:
       line = line[:-1]  # Strip the '\n' char.
       if line.startswith("#"):
         # Skip from reading comment line.
         continue
-      self.token_to_idx[line] = idx
-      self.idx_to_token[idx] = line
+      self.token_to_index[line] = index
+      self.index_to_token[index] = line
       self.speech_labels += line
-      idx += 1
+      index += 1

research/deep_speech/data/vocabulary.txt

 # List of alphabets (utf-8 encoded). Note that '#' starts a comment line, which
 # will be ignored by the parser.
 # begin of vocabulary
+ 
 a
 b
 c
@@ -28,6 +29,5 @@ x
 y
 z
 '
- 
 -
 # end of vocabulary

research/deep_speech/decoder.py

@@ -18,189 +18,78 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+import itertools
+
 from nltk.metrics import distance
-from six.moves import xrange
-import tensorflow as tf
+import numpy as np
 
 
 class DeepSpeechDecoder(object):
-  """Basic decoder class from which all other decoders inherit.
-
-  Implements several helper functions. Subclasses should implement the decode()
-  method.
-  """
+  """Greedy decoder implementation for Deep Speech model."""
 
-  def __init__(self, labels, blank_index=28, space_index=27):
+  def __init__(self, labels, blank_index=28):
     """Decoder initialization.
 
     Arguments:
-      labels (string): mapping from integers to characters.
-      blank_index (int, optional): index for the blank '_' character.
-        Defaults to 0.
-      space_index (int, optional): index for the space ' ' character.
+      labels: a string specifying the speech labels for the decoder to use.
+      blank_index: an integer specifying index for the blank character.
         Defaults to 28.
     """
     # e.g. labels = "[a-z]' _"
     self.labels = labels
-    self.int_to_char = dict([(i, c) for (i, c) in enumerate(labels)])
     self.blank_index = blank_index
-    self.space_index = space_index
-
-  def convert_to_strings(self, sequences, sizes=None):
-    """Given a list of numeric sequences, returns the corresponding strings."""
-    strings = []
-    for x in xrange(len(sequences)):
-      seq_len = sizes[x] if sizes is not None else len(sequences[x])
-      string = self._convert_to_string(sequences[x], seq_len)
-      strings.append(string)
-    return strings
-
-  def _convert_to_string(self, sequence, sizes):
-    return ''.join([self.int_to_char[sequence[i]] for i in range(sizes)])
-
-  def process_strings(self, sequences, remove_repetitions=False):
-    """Process strings.
+    self.int_to_char = dict([(i, c) for (i, c) in enumerate(labels)])
 
-    Given a list of strings, removes blanks and replace space character with
-    space. Option to remove repetitions (e.g. 'abbca' -> 'abca').
+  def convert_to_string(self, sequence):
+    """Convert a sequence of indexes into corresponding string."""
+    return ''.join([self.int_to_char[i] for i in sequence])
 
-    Arguments:
-      sequences: list of 1-d array of integers
-      remove_repetitions (boolean, optional): If true, repeating characters
-        are removed. Defaults to False.
-
-    Returns:
-      The processed string.
-    """
-    processed_strings = []
-    for sequence in sequences:
-      string = self.process_string(remove_repetitions, sequence).strip()
-      processed_strings.append(string)
-    return processed_strings
-
-  def process_string(self, remove_repetitions, sequence):
-    """Process each given sequence."""
-    seq_string = ''
-    for i, char in enumerate(sequence):
-      if char != self.int_to_char[self.blank_index]:
-        # if this char is a repetition and remove_repetitions=true,
-        # skip.
-        if remove_repetitions and i != 0 and char == sequence[i - 1]:
-          pass
-        elif char == self.labels[self.space_index]:
-          seq_string += ' '
-        else:
-          seq_string += char
-    return seq_string
-
-  def wer(self, output, target):
+  def wer(self, decode, target):
     """Computes the Word Error Rate (WER).
 
     WER is defined as the edit distance between the two provided sentences after
     tokenizing to words.
 
     Args:
-      output: string of the decoded output.
-      target: a string for the true transcript.
+      decode: string of the decoded output.
+      target: a string for the ground truth label.
 
     Returns:
-      A float number for the WER of the current sentence pair.
+      A float number for the WER of the current decode-target pair.
     """
     # Map each word to a new char.
-    words = set(output.split() + target.split())
+    words = set(decode.split() + target.split())
     word2char = dict(zip(words, range(len(words))))
 
-    new_output = [chr(word2char[w]) for w in output.split()]
+    new_decode = [chr(word2char[w]) for w in decode.split()]
     new_target = [chr(word2char[w]) for w in target.split()]
 
-    return distance.edit_distance(''.join(new_output), ''.join(new_target))
+    return distance.edit_distance(''.join(new_decode), ''.join(new_target))
 
-  def cer(self, output, target):
+  def cer(self, decode, target):
     """Computes the Character Error Rate (CER).
 
-    CER is  defined as the edit distance between the given strings.
+    CER is defined as the edit distance between the two given strings.
 
     Args:
-      output: a string of the decoded output.
-      target: a string for the ground truth transcript.
+      decode: a string of the decoded output.
+      target: a string for the ground truth label.
 
     Returns:
       A float number denoting the CER for the current sentence pair.
     """
-    return distance.edit_distance(output, target)
-
-  def batch_wer(self, decoded_output, targets):
-    """Compute the aggregate WER for each batch.
-
-    Args:
-      decoded_output: 2d array of integers for the decoded output of a batch.
-      targets: 2d array of integers for the labels of a batch.
-
-    Returns:
-      A float number for the aggregated WER for the current batch output.
-    """
-    # Convert numeric representation to string.
-    decoded_strings = self.convert_to_strings(decoded_output)
-    decoded_strings = self.process_strings(
-        decoded_strings, remove_repetitions=True)
-    target_strings = self.convert_to_strings(targets)
-    target_strings = self.process_strings(
-        target_strings, remove_repetitions=True)
-    wer = 0
-    for i in xrange(len(decoded_strings)):
-      wer += self.wer(decoded_strings[i], target_strings[i]) / float(
-          len(target_strings[i].split()))
-    return wer
-
-  def batch_cer(self, decoded_output, targets):
-    """Compute the aggregate CER for each batch.
-
-    Args:
-      decoded_output: 2d array of integers for the decoded output of a batch.
-      targets: 2d array of integers for the labels of a batch.
-
-    Returns:
-      A float number for the aggregated CER for the current batch output.
-    """
-    # Convert numeric representation to string.
-    decoded_strings = self.convert_to_strings(decoded_output)
-    decoded_strings = self.process_strings(
-        decoded_strings, remove_repetitions=True)
-    target_strings = self.convert_to_strings(targets)
-    target_strings = self.process_strings(
-        target_strings, remove_repetitions=True)
-    cer = 0
-    for i in xrange(len(decoded_strings)):
-      cer += self.cer(decoded_strings[i], target_strings[i]) / float(
-          len(target_strings[i]))
-    return cer
-
-  def decode(self, sequences, sizes=None):
-    """Perform sequence decoding.
-
-    Given a matrix of character probabilities, returns the decoder's best guess
-    of the transcription.
-
-    Arguments:
-      sequences: 2D array of character probabilities, where sequences[c, t] is
-        the probability of character c at time t.
-      sizes(optional): Size of each sequence in the mini-batch.
-
-    Returns:
-      string: sequence of the model's best guess for the transcription.
-    """
-    strings = self.convert_to_strings(sequences, sizes)
-    return self.process_strings(strings, remove_repetitions=True)
-
-
-class GreedyDecoder(DeepSpeechDecoder):
-  """Greedy decoder."""
-
-  def decode(self, logits, seq_len):
-    # Reshape to [max_time, batch_size, num_classes]
-    logits = tf.transpose(logits, (1, 0, 2))
-    decoded, _ = tf.nn.ctc_greedy_decoder(logits, seq_len)
-    decoded_dense = tf.Session().run(tf.sparse_to_dense(
-        decoded[0].indices, decoded[0].dense_shape, decoded[0].values))
-    result = self.convert_to_strings(decoded_dense)
-    return self.process_strings(result, remove_repetitions=True), decoded_dense
+    return distance.edit_distance(decode, target)
+
+  def decode(self, logits):
+    """Decode the best guess from logits using greedy algorithm."""
+    # Choose the class with maximimum probability.
+    best = list(np.argmax(logits, axis=1))
+    # Merge repeated chars.
+    merge = [k for k, _ in itertools.groupby(best)]
+    # Remove the blank index in the decoded sequence.
+    merge_remove_blank = []
+    for k in merge:
+      if k != self.blank_index:
+        merge_remove_blank.append(k)
+
+    return self.convert_to_string(merge_remove_blank)

research/deep_speech/deep_speech.py

@@ -41,8 +41,50 @@ _WER_KEY = "WER"
 _CER_KEY = "CER"
 
 
-def evaluate_model(
-    estimator, batch_size, speech_labels, targets, input_fn_eval):
+def compute_length_after_conv(max_time_steps, ctc_time_steps, input_length):
+  """Computes the time_steps/ctc_input_length after convolution.
+
+  Suppose that the original feature contains two parts:
+  1) Real spectrogram signals, spanning input_length steps.
+  2) Padded part with all 0s.
+  The total length of those two parts is denoted as max_time_steps, which is
+  the padded length of the current batch. After convolution layers, the time
+  steps of a spectrogram feature will be decreased. As we know the percentage
+  of its original length within the entire length, we can compute the time steps
+  for the signal after conv as follows (using ctc_input_length to denote):
+  ctc_input_length = (input_length / max_time_steps) * output_length_of_conv.
+  This length is then fed into ctc loss function to compute loss.
+
+  Args:
+    max_time_steps: max_time_steps for the batch, after padding.
+    ctc_time_steps: number of timesteps after convolution.
+    input_length: actual length of the original spectrogram, without padding.
+
+  Returns:
+    the ctc_input_length after convolution layer.
+  """
+  ctc_input_length = tf.to_float(tf.multiply(
+      input_length, ctc_time_steps))
+  return tf.to_int32(tf.floordiv(
+      ctc_input_length, tf.to_float(max_time_steps)))
+
+
+def ctc_loss(label_length, ctc_input_length, labels, logits):
+  """Computes the ctc loss for the current batch of predictions."""
+  label_length = tf.to_int32(tf.squeeze(label_length))
+  ctc_input_length = tf.to_int32(tf.squeeze(ctc_input_length))
+  sparse_labels = tf.to_int32(
+      tf.keras.backend.ctc_label_dense_to_sparse(labels, label_length))
+  y_pred = tf.log(tf.transpose(
+      logits, perm=[1, 0, 2]) + tf.keras.backend.epsilon())
+
+  return tf.expand_dims(
+      tf.nn.ctc_loss(labels=sparse_labels, inputs=y_pred,
+                     sequence_length=ctc_input_length),
+      axis=1)
+
+
+def evaluate_model(estimator, speech_labels, entries, input_fn_eval):
   """Evaluate the model performance using WER anc CER as metrics.
 
   WER: Word Error Rate
@@ -50,44 +92,34 @@ def evaluate_model(
 
   Args:
     estimator: estimator to evaluate.
-    batch_size: size of a mini-batch.
     speech_labels: a string specifying all the character in the vocabulary.
-    targets: a list of list of integers for the featurized transcript.
+    entries: a list of data entries (audio_file, file_size, transcript) for the
+      given dataset.
     input_fn_eval: data input function for evaluation.
 
   Returns:
     Evaluation result containing 'wer' and 'cer' as two metrics.
   """
   # Get predictions
-  predictions = estimator.predict(
-      input_fn=input_fn_eval, yield_single_examples=False)
+  predictions = estimator.predict(input_fn=input_fn_eval)
 
-  y_preds = []
-  input_lengths = []
-  for p in predictions:
-    y_preds.append(p["y_pred"])
-    input_lengths.append(p["ctc_input_length"])
+  # Get probabilities of each predicted class
+  probs = [pred["probabilities"] for pred in predictions]
 
-  num_of_examples = len(targets)
-  total_wer, total_cer = 0, 0
-  greedy_decoder = decoder.GreedyDecoder(speech_labels)
-  for i in range(len(y_preds)):
-    # Compute the CER and WER for the current batch,
-    # and aggregate to total_cer, total_wer.
-    y_pred_tensor = tf.convert_to_tensor(y_preds[i])
-    batch_targets = targets[i * batch_size : (i + 1) * batch_size]
-    seq_len = tf.squeeze(input_lengths[i], axis=1)
-
-    # Perform decoding
-    _, decoded_output = greedy_decoder.decode(
-        y_pred_tensor, seq_len)
+  num_of_examples = len(probs)
+  targets = [entry[2] for entry in entries]  # The ground truth transcript
 
+  total_wer, total_cer = 0, 0
+  greedy_decoder = decoder.DeepSpeechDecoder(speech_labels)
+  for i in range(num_of_examples):
+    # Decode string.
+    decoded_str = greedy_decoder.decode(probs[i])
     # Compute CER.
-    batch_cer = greedy_decoder.batch_cer(decoded_output, batch_targets)
-    total_cer += batch_cer
+    total_cer += greedy_decoder.cer(decoded_str, targets[i]) / float(
+        len(targets[i]))
     # Compute WER.
-    batch_wer = greedy_decoder.batch_wer(decoded_output, batch_targets)
-    total_wer += batch_wer
+    total_wer += greedy_decoder.wer(decoded_str, targets[i]) / float(
+        len(targets[i].split()))
 
   # Get mean value
   total_cer /= num_of_examples
@@ -103,45 +135,76 @@ def evaluate_model(
   return eval_results
 
 
-def convert_keras_to_estimator(keras_model, num_gpus):
-  """Configure and convert keras model to Estimator.
+def model_fn(features, labels, mode, params):
+  """Define model function for deep speech model.
 
   Args:
-    keras_model: A Keras model object.
-    num_gpus: An integer, the number of GPUs.
+    features: a dictionary of input_data features. It includes the data
+      input_length, label_length and the spectrogram features.
+    labels: a list of labels for the input data.
+    mode: current estimator mode; should be one of
+      `tf.estimator.ModeKeys.TRAIN`, `EVALUATE`, `PREDICT`.
+    params: a dict of hyper parameters to be passed to model_fn.
 
   Returns:
-    estimator: The converted Estimator.
+    EstimatorSpec parameterized according to the input params and the
+    current mode.
   """
-  # keras optimizer is not compatible with distribution strategy.
-  # Use tf optimizer instead
-  optimizer = tf.train.MomentumOptimizer(
-      learning_rate=flags_obj.learning_rate, momentum=flags_obj.momentum,
-      use_nesterov=True)
-
-  # ctc_loss is wrapped as a Lambda layer in the model.
-  keras_model.compile(
-      optimizer=optimizer, loss={"ctc_loss": lambda y_true, y_pred: y_pred})
-
-  distribution_strategy = distribution_utils.get_distribution_strategy(
-      num_gpus)
-  run_config = tf.estimator.RunConfig(
-      train_distribute=distribution_strategy)
-
-  estimator = tf.keras.estimator.model_to_estimator(
-      keras_model=keras_model, model_dir=flags_obj.model_dir, config=run_config)
-
-  return estimator
+  num_classes = params["num_classes"]
+  input_length = features["input_length"]
+  label_length = features["label_length"]
+  features = features["features"]
+
+  # Create DeepSpeech2 model.
+  model = deep_speech_model.DeepSpeech2(
+      flags_obj.rnn_hidden_layers, flags_obj.rnn_type,
+      flags_obj.is_bidirectional, flags_obj.rnn_hidden_size,
+      num_classes, flags_obj.use_bias)
+
+  if mode == tf.estimator.ModeKeys.PREDICT:
+    logits = model(features, training=False)
+    predictions = {
+        "classes": tf.argmax(logits, axis=2),
+        "probabilities": tf.nn.softmax(logits),
+        "logits": logits
+    }
+    return tf.estimator.EstimatorSpec(
+        mode=mode,
+        predictions=predictions)
+
+  # In training mode.
+  logits = model(features, training=True)
+  probs = tf.nn.softmax(logits)
+  ctc_input_length = compute_length_after_conv(
+      tf.shape(features)[1], tf.shape(probs)[1], input_length)
+  # Compute CTC loss
+  loss = tf.reduce_mean(ctc_loss(
+      label_length, ctc_input_length, labels, probs))
+
+  optimizer = tf.train.AdamOptimizer(learning_rate=flags_obj.learning_rate)
+  global_step = tf.train.get_or_create_global_step()
+  minimize_op = optimizer.minimize(loss, global_step=global_step)
+  update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
+  # Create the train_op that groups both minimize_ops and update_ops
+  train_op = tf.group(minimize_op, update_ops)
+
+  return tf.estimator.EstimatorSpec(
+      mode=mode,
+      loss=loss,
+      train_op=train_op)
 
 
 def generate_dataset(data_dir):
   """Generate a speech dataset."""
-  audio_conf = dataset.AudioConfig(
-      flags_obj.sample_rate, flags_obj.frame_length, flags_obj.frame_step)
+  audio_conf = dataset.AudioConfig(sample_rate=flags_obj.sample_rate,
+                                   window_ms=flags_obj.window_ms,
+                                   stride_ms=flags_obj.stride_ms,
+                                   normalize=True)
   train_data_conf = dataset.DatasetConfig(
       audio_conf,
       data_dir,
       flags_obj.vocabulary_file,
+      flags_obj.sortagrad
   )
   speech_dataset = dataset.DeepSpeechDataset(train_data_conf)
   return speech_dataset
@@ -150,30 +213,27 @@ def generate_dataset(data_dir):
 def run_deep_speech(_):
   """Run deep speech training and eval loop."""
   # Data preprocessing
-  # The file name of training and test dataset
   tf.logging.info("Data preprocessing...")
-
   train_speech_dataset = generate_dataset(flags_obj.train_data_dir)
   eval_speech_dataset = generate_dataset(flags_obj.eval_data_dir)
 
   # Number of label classes. Label string is "[a-z]' -"
   num_classes = len(train_speech_dataset.speech_labels)
 
-  # Input shape of each data example:
-  # [time_steps (T), feature_bins(F), channel(C)]
-  # Channel is set as 1 by default.
-  input_shape = (None, train_speech_dataset.num_feature_bins, 1)
-
-  # Create deep speech model and convert it to Estimator
-  tf.logging.info("Creating Estimator from Keras model...")
-  keras_model = deep_speech_model.DeepSpeech(
-      input_shape, flags_obj.rnn_hidden_layers, flags_obj.rnn_type,
-      flags_obj.is_bidirectional, flags_obj.rnn_hidden_size,
-      flags_obj.rnn_activation, num_classes, flags_obj.use_bias)
-
-  # Convert to estimator
+  # Use distribution strategy for multi-gpu training
   num_gpus = flags_core.get_num_gpus(flags_obj)
-  estimator = convert_keras_to_estimator(keras_model, num_gpus)
+  distribution_strategy = distribution_utils.get_distribution_strategy(num_gpus)
+  run_config = tf.estimator.RunConfig(
+      train_distribute=distribution_strategy)
+
+  estimator = tf.estimator.Estimator(
+      model_fn=model_fn,
+      model_dir=flags_obj.model_dir,
+      config=run_config,
+      params={
+          "num_classes": num_classes,
+      }
+  )
 
   # Benchmark logging
   run_params = {
@@ -181,7 +241,6 @@ def run_deep_speech(_):
       "train_epochs": flags_obj.train_epochs,
       "rnn_hidden_size": flags_obj.rnn_hidden_size,
       "rnn_hidden_layers": flags_obj.rnn_hidden_layers,
-      "rnn_activation": flags_obj.rnn_activation,
       "rnn_type": flags_obj.rnn_type,
       "is_bidirectional": flags_obj.is_bidirectional,
       "use_bias": flags_obj.use_bias
@@ -194,6 +253,7 @@ def run_deep_speech(_):
 
   train_hooks = hooks_helper.get_train_hooks(
       flags_obj.hooks,
+      model_dir=flags_obj.model_dir,
       batch_size=flags_obj.batch_size)
 
   per_device_batch_size = distribution_utils.per_device_batch_size(
@@ -213,14 +273,19 @@ def run_deep_speech(_):
     tf.logging.info("Starting a training cycle: %d/%d",
                     cycle_index + 1, total_training_cycle)
 
+    # Perform batch_wise dataset shuffling
+    train_speech_dataset.entries = dataset.batch_wise_dataset_shuffle(
+        train_speech_dataset.entries, cycle_index, flags_obj.sortagrad,
+        flags_obj.batch_size)
+
     estimator.train(input_fn=input_fn_train, hooks=train_hooks)
 
     # Evaluation
     tf.logging.info("Starting to evaluate...")
 
     eval_results = evaluate_model(
-        estimator, flags_obj.batch_size, eval_speech_dataset.speech_labels,
-        eval_speech_dataset.labels, input_fn_eval)
+        estimator, eval_speech_dataset.speech_labels,
+        eval_speech_dataset.entries, input_fn_eval)
 
     # Log the WER and CER results.
     benchmark_logger.log_evaluation_result(eval_results)
@@ -233,9 +298,6 @@ def run_deep_speech(_):
         flags_obj.wer_threshold, eval_results[_WER_KEY]):
       break
 
-  # Clear the session explicitly to avoid session delete error
-  tf.keras.backend.clear_session()
-
 
 def define_deep_speech_flags():
   """Add flags for run_deep_speech."""
@@ -257,8 +319,8 @@ def define_deep_speech_flags():
   flags_core.set_defaults(
       model_dir="/tmp/deep_speech_model/",
       export_dir="/tmp/deep_speech_saved_model/",
-      train_epochs=2,
-      batch_size=4,
+      train_epochs=200,
+      batch_size=128,
       hooks="")
 
   # Deep speech flags
@@ -272,16 +334,22 @@ def define_deep_speech_flags():
       default="/tmp/librispeech_data/test-clean/LibriSpeech/test-clean-20.csv",
       help=flags_core.help_wrap("The csv file path of evaluation dataset."))
 
+  flags.DEFINE_bool(
+      name="sortagrad", default=True,
+      help=flags_core.help_wrap(
+          "If true, sort examples by audio length and perform no "
+          "batch_wise shuffling for the first epoch."))
+
   flags.DEFINE_integer(
       name="sample_rate", default=16000,
       help=flags_core.help_wrap("The sample rate for audio."))
 
   flags.DEFINE_integer(
-      name="frame_length", default=25,
+      name="window_ms", default=20,
       help=flags_core.help_wrap("The frame length for spectrogram."))
 
   flags.DEFINE_integer(
-      name="frame_step", default=10,
+      name="stride_ms", default=10,
       help=flags_core.help_wrap("The frame step."))
 
   flags.DEFINE_string(
@@ -290,11 +358,11 @@ def define_deep_speech_flags():
 
   # RNN related flags
   flags.DEFINE_integer(
-      name="rnn_hidden_size", default=256,
+      name="rnn_hidden_size", default=800,
       help=flags_core.help_wrap("The hidden size of RNNs."))
 
   flags.DEFINE_integer(
-      name="rnn_hidden_layers", default=3,
+      name="rnn_hidden_layers", default=5,
       help=flags_core.help_wrap("The number of RNN layers."))
 
   flags.DEFINE_bool(
@@ -311,20 +379,11 @@ def define_deep_speech_flags():
       case_sensitive=False,
       help=flags_core.help_wrap("Type of RNN cell."))
 
-  flags.DEFINE_enum(
-      name="rnn_activation", default="tanh",
-      enum_values=["tanh", "relu"], case_sensitive=False,
-      help=flags_core.help_wrap("Type of the activation within RNN."))
-
   # Training related flags
   flags.DEFINE_float(
-      name="learning_rate", default=0.0003,
+      name="learning_rate", default=5e-4,
       help=flags_core.help_wrap("The initial learning rate."))
 
-  flags.DEFINE_float(
-      name="momentum", default=0.9,
-      help=flags_core.help_wrap("Momentum to accelerate SGD optimizer."))
-
   # Evaluation metrics threshold
   flags.DEFINE_float(
       name="wer_threshold", default=None,
@@ -345,3 +404,4 @@ if __name__ == "__main__":
   define_deep_speech_flags()
   flags_obj = flags.FLAGS
   absl_app.run(main)
+

research/deep_speech/deep_speech_model.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-"""Network structure for DeepSpeech model."""
+"""Network structure for DeepSpeech2 model."""
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
@@ -20,175 +20,166 @@ from __future__ import print_function
 from six.moves import xrange  # pylint: disable=redefined-builtin
 import tensorflow as tf
 
-# Supported rnn cells
+# Supported rnn cells.
 SUPPORTED_RNNS = {
-    "lstm": tf.keras.layers.LSTM,
-    "rnn": tf.keras.layers.SimpleRNN,
-    "gru": tf.keras.layers.GRU,
+    "lstm": tf.nn.rnn_cell.BasicLSTMCell,
+    "rnn": tf.nn.rnn_cell.RNNCell,
+    "gru": tf.nn.rnn_cell.GRUCell,
 }
 
-# Parameters for batch normalization
-_MOMENTUM = 0.1
-_EPSILON = 1e-05
+# Parameters for batch normalization.
+_BATCH_NORM_EPSILON = 1e-5
+_BATCH_NORM_DECAY = 0.997
 
+# Filters of convolution layer
+_CONV_FILTERS = 32
 
-def _conv_bn_layer(cnn_input, filters, kernel_size, strides, layer_id):
-  """2D convolution + batch normalization layer.
+
+def batch_norm(inputs, training):
+  """Batch normalization layer.
+
+  Note that the momentum to use will affect validation accuracy over time.
+  Batch norm has different behaviors during training/evaluation. With a large
+  momentum, the model takes longer to get a near-accurate estimation of the
+  moving mean/variance over the entire training dataset, which means we need
+  more iterations to see good evaluation results. If the training data is evenly
+  distributed over the feature space, we can also try setting a smaller momentum
+  (such as 0.1) to get good evaluation result sooner.
+
+  Args:
+    inputs: input data for batch norm layer.
+    training: a boolean to indicate if it is in training stage.
+
+  Returns:
+    tensor output from batch norm layer.
+  """
+  return tf.layers.batch_normalization(
+      inputs=inputs, momentum=_BATCH_NORM_DECAY, epsilon=_BATCH_NORM_EPSILON,
+      fused=True, training=training)
+
+
+def _conv_bn_layer(inputs, padding, filters, kernel_size, strides, layer_id,
+                   training):
+  """Defines 2D constitutional + batch normalization layer.
 
   Args:
-    cnn_input: input data for convolution layer.
+    inputs: input data for convolution layer.
+    padding: padding to be applied before convolution layer.
     filters: an integer, number of output filters in the convolution.
     kernel_size: a tuple specifying the height and width of the 2D convolution
       window.
     strides: a tuple specifying the stride length of the convolution.
     layer_id: an integer specifying the layer index.
+    training: a boolean to indicate which stage we are in (training/eval).
 
   Returns:
     tensor output from the current layer.
   """
-  output = tf.keras.layers.Conv2D(
-      filters=filters, kernel_size=kernel_size, strides=strides, padding="same",
-      activation="linear", name="cnn_{}".format(layer_id))(cnn_input)
-  output = tf.keras.layers.BatchNormalization(
-      momentum=_MOMENTUM, epsilon=_EPSILON)(output)
-  return output
-
-
-def _rnn_layer(input_data, rnn_cell, rnn_hidden_size, layer_id, rnn_activation,
-               is_batch_norm, is_bidirectional):
+  # Perform symmetric padding on the feature dimension of time_step
+  # This step is required to avoid issues when RNN output sequence is shorter
+  # than the label length.
+  inputs = tf.pad(
+      inputs,
+      [[0, 0], [padding[0], padding[0]], [padding[1], padding[1]], [0, 0]])
+  inputs = tf.layers.conv2d(
+      inputs=inputs, filters=filters, kernel_size=kernel_size, strides=strides,
+      padding="valid", use_bias=False, activation=tf.nn.relu6,
+      name="cnn_{}".format(layer_id))
+  return batch_norm(inputs, training)
+
+
+def _rnn_layer(inputs, rnn_cell, rnn_hidden_size, layer_id, is_batch_norm,
+               is_bidirectional, training):
   """Defines a batch normalization + rnn layer.
 
   Args:
-    input_data: input tensors for the current layer.
+    inputs: input tensors for the current layer.
     rnn_cell: RNN cell instance to use.
     rnn_hidden_size: an integer for the dimensionality of the rnn output space.
     layer_id: an integer for the index of current layer.
-    rnn_activation: activation function to use.
     is_batch_norm: a boolean specifying whether to perform batch normalization
       on input states.
     is_bidirectional: a boolean specifying whether the rnn layer is
       bi-directional.
+    training: a boolean to indicate which stage we are in (training/eval).
 
   Returns:
     tensor output for the current layer.
   """
   if is_batch_norm:
-    input_data = tf.keras.layers.BatchNormalization(
-        momentum=_MOMENTUM, epsilon=_EPSILON)(input_data)
-  rnn_layer = rnn_cell(
-      rnn_hidden_size, activation=rnn_activation, return_sequences=True,
-      name="rnn_{}".format(layer_id))
-  if is_bidirectional:
-    rnn_layer = tf.keras.layers.Bidirectional(rnn_layer, merge_mode="sum")
-
-  return rnn_layer(input_data)
-
+    inputs = batch_norm(inputs, training)
 
-def _ctc_lambda_func(args):
-  """Compute ctc loss."""
-  # py2 needs explicit tf import for keras Lambda layer
-  import tensorflow as tf
+  # Construct forward/backward RNN cells.
+  fw_cell = rnn_cell(num_units=rnn_hidden_size,
+                     name="rnn_fw_{}".format(layer_id))
+  bw_cell = rnn_cell(num_units=rnn_hidden_size,
+                     name="rnn_bw_{}".format(layer_id))
 
-  y_pred, labels, input_length, label_length = args
-  return tf.keras.backend.ctc_batch_cost(
-      labels, y_pred, input_length, label_length)
-
-
-def _calc_ctc_input_length(args):
-  """Compute the actual input length after convolution for ctc_loss function.
-
-  Basically, we need to know the scaled input_length after conv layers.
-  new_input_length = old_input_length * ctc_time_steps / max_time_steps
-
-  Args:
-    args: the input args to compute ctc input length.
-
-  Returns:
-    ctc_input_length, which is required for ctc loss calculation.
-  """
-  # py2 needs explicit tf import for keras Lambda layer
-  import tensorflow as tf
+  if is_bidirectional:
+    outputs, _ = tf.nn.bidirectional_dynamic_rnn(
+        cell_fw=fw_cell, cell_bw=bw_cell, inputs=inputs, dtype=tf.float32,
+        swap_memory=True)
+    rnn_outputs = tf.concat(outputs, -1)
+  else:
+    rnn_outputs = tf.nn.dynamic_rnn(
+        fw_cell, inputs, dtype=tf.float32, swap_memory=True)
 
-  input_length, input_data, y_pred = args
-  max_time_steps = tf.shape(input_data)[1]
-  ctc_time_steps = tf.shape(y_pred)[1]
-  ctc_input_length = tf.multiply(
-      tf.to_float(input_length), tf.to_float(ctc_time_steps))
-  ctc_input_length = tf.to_int32(tf.floordiv(
-      ctc_input_length, tf.to_float(max_time_steps)))
-  return ctc_input_length
+  return rnn_outputs
 
 
-class DeepSpeech(tf.keras.models.Model):
-  """DeepSpeech model."""
+class DeepSpeech2(object):
+  """Define DeepSpeech2 model."""
 
-  def __init__(self, input_shape, num_rnn_layers, rnn_type, is_bidirectional,
-               rnn_hidden_size, rnn_activation, num_classes, use_bias):
-    """Initialize DeepSpeech model.
+  def __init__(self, num_rnn_layers, rnn_type, is_bidirectional,
+               rnn_hidden_size, num_classes, use_bias):
+    """Initialize DeepSpeech2 model.
 
     Args:
-      input_shape: an tuple to indicate the dimension of input dataset. It has
-        the format of [time_steps(T), feature_bins(F), channel(1)]
       num_rnn_layers: an integer, the number of rnn layers. By default, it's 5.
       rnn_type: a string, one of the supported rnn cells: gru, rnn and lstm.
       is_bidirectional: a boolean to indicate if the rnn layer is bidirectional.
       rnn_hidden_size: an integer for the number of hidden states in each unit.
-      rnn_activation: a string to indicate rnn activation function. It can be
-        one of tanh and relu.
       num_classes: an integer, the number of output classes/labels.
       use_bias: a boolean specifying whether to use bias in the last fc layer.
     """
-    # Input variables
-    input_data = tf.keras.layers.Input(
-        shape=input_shape, name="features")
-
-    # Two cnn layers
-    conv_layer_1 = _conv_bn_layer(
-        input_data, filters=32, kernel_size=(41, 11), strides=(2, 2),
-        layer_id=1)
-
-    conv_layer_2 = _conv_bn_layer(
-        conv_layer_1, filters=32, kernel_size=(21, 11), strides=(2, 1),
-        layer_id=2)
+    self.num_rnn_layers = num_rnn_layers
+    self.rnn_type = rnn_type
+    self.is_bidirectional = is_bidirectional
+    self.rnn_hidden_size = rnn_hidden_size
+    self.num_classes = num_classes
+    self.use_bias = use_bias
+
+  def __call__(self, inputs, training):
+    # Two cnn layers.
+    inputs = _conv_bn_layer(
+        inputs, padding=(20, 5), filters=_CONV_FILTERS, kernel_size=(41, 11),
+        strides=(2, 2), layer_id=1, training=training)
+
+    inputs = _conv_bn_layer(
+        inputs, padding=(10, 5), filters=_CONV_FILTERS, kernel_size=(21, 11),
+        strides=(2, 1), layer_id=2, training=training)
+
     # output of conv_layer2 with the shape of
-    # [batch_size (N), times (T), features (F), channels (C)]
+    # [batch_size (N), times (T), features (F), channels (C)].
+    # Convert the conv output to rnn input.
+    batch_size = tf.shape(inputs)[0]
+    feat_size = inputs.get_shape().as_list()[2]
+    inputs = tf.reshape(
+        inputs,
+        [batch_size, -1, feat_size * _CONV_FILTERS])
 
     # RNN layers.
-    # Convert the conv output to rnn input
-    rnn_input = tf.keras.layers.TimeDistributed(tf.keras.layers.Flatten())(
-        conv_layer_2)
-
-    rnn_cell = SUPPORTED_RNNS[rnn_type]
-    for layer_counter in xrange(num_rnn_layers):
-      # No batch normalization on the first layer
+    rnn_cell = SUPPORTED_RNNS[self.rnn_type]
+    for layer_counter in xrange(self.num_rnn_layers):
+      # No batch normalization on the first layer.
       is_batch_norm = (layer_counter != 0)
-      rnn_input = _rnn_layer(
-          rnn_input, rnn_cell, rnn_hidden_size, layer_counter + 1,
-          rnn_activation, is_batch_norm, is_bidirectional)
-
-    # FC layer with batch norm
-    fc_input = tf.keras.layers.BatchNormalization(
-        momentum=_MOMENTUM, epsilon=_EPSILON)(rnn_input)
-
-    y_pred = tf.keras.layers.Dense(num_classes, activation="softmax",
-                                   use_bias=use_bias, name="y_pred")(fc_input)
-
-    # For ctc loss
-    labels = tf.keras.layers.Input(name="labels", shape=[None,], dtype="int32")
-    label_length = tf.keras.layers.Input(
-        name="label_length", shape=[1], dtype="int32")
-    input_length = tf.keras.layers.Input(
-        name="input_length", shape=[1], dtype="int32")
-    ctc_input_length = tf.keras.layers.Lambda(
-        _calc_ctc_input_length, output_shape=(1,), name="ctc_input_length")(
-            [input_length, input_data, y_pred])
-
-    # Keras doesn't currently support loss funcs with extra parameters
-    # so CTC loss is implemented in a lambda layer
-    ctc_loss = tf.keras.layers.Lambda(
-        _ctc_lambda_func, output_shape=(1,), name="ctc_loss")(
-            [y_pred, labels, ctc_input_length, label_length])
-
-    super(DeepSpeech, self).__init__(
-        inputs=[input_data, labels, input_length, label_length],
-        outputs=[ctc_input_length, ctc_loss, y_pred])
+      inputs = _rnn_layer(
+          inputs, rnn_cell, self.rnn_hidden_size, layer_counter + 1,
+          is_batch_norm, self.is_bidirectional, training)
+
+    # FC layer with batch norm.
+    inputs = batch_norm(inputs, training)
+    logits = tf.layers.dense(inputs, self.num_classes, use_bias=self.use_bias)
+
+    return logits
+

research/deep_speech/requirements.txt

+nltk>=3.3
+soundfile>=0.10.2
+sox>=1.3.3