Merge branch 'develop' into refine_seq2seq

Author: lcy-seso

.travis.yml

@@ -1,3 +1,4 @@
+group: deprecated-2017Q2
 language: cpp
 cache: ccache
 sudo: required

.travis/unittest.sh

@@ -8,8 +8,8 @@ abort(){
 
 unittest(){
     cd $1 > /dev/null
-    if [ -f "requirements.txt" ]; then
-        pip install -r requirements.txt
+    if [ -f "setup.sh" ]; then
+        sh setup.sh
     fi
     if [ $? != 0 ]; then
         exit 1

deep_speech_2/README.md

@@ -5,7 +5,7 @@
 Please replace `$PADDLE_INSTALL_DIR` with your own paddle installation directory.
 
 ```
-pip install -r requirements.txt
+sh setup.sh
 export LD_LIBRARY_PATH=$PADDLE_INSTALL_DIR/Paddle/third_party/install/warpctc/lib:$LD_LIBRARY_PATH
 ```
 

deep_speech_2/data_utils/data.py

@@ -7,6 +7,7 @@
 
 import random
 import numpy as np
+import multiprocessing
 import paddle.v2 as paddle
 from data_utils import utils
 from data_utils.augmentor.augmentation import AugmentationPipeline
@@ -44,6 +45,8 @@ class DataGenerator(object):
     :types max_freq: None|float
     :param specgram_type: Specgram feature type. Options: 'linear'.
     :type specgram_type: str
+    :param num_threads: Number of CPU threads for processing data.
+    :type num_threads: int
     :param random_seed: Random seed.
     :type random_seed: int
     """
@@ -58,6 +61,7 @@ def __init__(self,
                  window_ms=20.0,
                  max_freq=None,
                  specgram_type='linear',
+                 num_threads=multiprocessing.cpu_count(),
                  random_seed=0):
         self._max_duration = max_duration
         self._min_duration = min_duration
@@ -70,6 +74,7 @@ def __init__(self,
             stride_ms=stride_ms,
             window_ms=window_ms,
             max_freq=max_freq)
+        self._num_threads = num_threads
         self._rng = random.Random(random_seed)
         self._epoch = 0
 
@@ -207,10 +212,14 @@ def _instance_reader_creator(self, manifest):
 
         def reader():
             for instance in manifest:
-                yield self._process_utterance(instance["audio_filepath"],
-                                              instance["text"])
+                yield instance
 
-        return reader
+        def mapper(instance):
+            return self._process_utterance(instance["audio_filepath"],
+                                           instance["text"])
+
+        return paddle.reader.xmap_readers(
+            mapper, reader, self._num_threads, 1024, order=True)
 
     def _padding_batch(self, batch, padding_to=-1, flatten=False):
         """

deep_speech_2/data_utils/speech.py

@@ -94,7 +94,7 @@ def concatenate(cls, *segments):
         return cls(samples, sample_rate, transcripts)
 
     @classmethod
-    def slice_from_file(cls, filepath, start=None, end=None, transcript):
+    def slice_from_file(cls, filepath, transcript, start=None, end=None):
         """Loads a small section of an speech without having to load
         the entire file into the memory which can be incredibly wasteful.
 

deep_speech_2/infer.py

@@ -6,6 +6,7 @@
 import argparse
 import gzip
 import distutils.util
+import multiprocessing
 import paddle.v2 as paddle
 from data_utils.data import DataGenerator
 from model import deep_speech2
@@ -38,6 +39,11 @@
     default=True,
     type=distutils.util.strtobool,
     help="Use gpu or not. (default: %(default)s)")
+parser.add_argument(
+    "--num_threads_data",
+    default=multiprocessing.cpu_count(),
+    type=int,
+    help="Number of cpu threads for preprocessing data. (default: %(default)s)")
 parser.add_argument(
     "--mean_std_filepath",
     default='mean_std.npz',
@@ -67,7 +73,8 @@ def infer():
     data_generator = DataGenerator(
         vocab_filepath=args.vocab_filepath,
         mean_std_filepath=args.mean_std_filepath,
-        augmentation_config='{}')
+        augmentation_config='{}',
+        num_threads=args.num_threads_data)
 
     # create network config
     # paddle.data_type.dense_array is used for variable batch input.

deep_speech_2/requirements.txt

@@ -1,4 +1,3 @@
 SoundFile==0.9.0.post1
 wget==3.2
-scikits.samplerate==0.3.3
-scipy==0.13.0b1
+scipy==0.13.1

deep_speech_2/setup.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+
+# install python dependencies
+if [ -f 'requirements.txt' ]; then
+    pip install -r requirements.txt
+fi
+if [ $? != 0 ]; then
+    echo "Install python dependencies failed !!!"
+    exit 1
+fi
+
+# install scikits.samplerate
+curl -O "http://www.mega-nerd.com/SRC/libsamplerate-0.1.9.tar.gz"
+if [ $? != 0 ]; then
+    echo "Download libsamplerate-0.1.9.tar.gz failed !!!"
+    exit 1
+fi
+tar -xvf libsamplerate-0.1.9.tar.gz
+cd libsamplerate-0.1.9
+./configure && make && make install
+cd -
+rm -rf libsamplerate-0.1.9
+rm libsamplerate-0.1.9.tar.gz
+pip install scikits.samplerate==0.3.3
+if [ $? != 0 ]; then
+    echo "Install scikits.samplerate failed !!!"
+    exit 1
+fi
+
+echo "Install all dependencies successfully."

deep_speech_2/train.py

@@ -9,6 +9,7 @@
 import gzip
 import time
 import distutils.util
+import multiprocessing
 import paddle.v2 as paddle
 from model import deep_speech2
 from data_utils.data import DataGenerator
@@ -52,6 +53,18 @@
     default=True,
     type=distutils.util.strtobool,
     help="Use sortagrad or not. (default: %(default)s)")
+parser.add_argument(
+    "--max_duration",
+    default=100.0,
+    type=float,
+    help="Audios with duration larger than this will be discarded. "
+    "(default: %(default)s)")
+parser.add_argument(
+    "--min_duration",
+    default=0.0,
+    type=float,
+    help="Audios with duration smaller than this will be discarded. "
+    "(default: %(default)s)")
 parser.add_argument(
     "--shuffle_method",
     default='instance_shuffle',
@@ -63,6 +76,11 @@
     default=4,
     type=int,
     help="Trainer number. (default: %(default)s)")
+parser.add_argument(
+    "--num_threads_data",
+    default=multiprocessing.cpu_count(),
+    type=int,
+    help="Number of cpu threads for preprocessing data. (default: %(default)s)")
 parser.add_argument(
     "--mean_std_filepath",
     default='mean_std.npz',
@@ -107,7 +125,10 @@ def data_generator():
         return DataGenerator(
             vocab_filepath=args.vocab_filepath,
             mean_std_filepath=args.mean_std_filepath,
-            augmentation_config=args.augmentation_config)
+            augmentation_config=args.augmentation_config,
+            max_duration=args.max_duration,
+            min_duration=args.min_duration,
+            num_threads=args.num_threads_data)
 
     train_generator = data_generator()
     test_generator = data_generator()

language_model/network_conf.py

@@ -51,56 +51,41 @@ def rnn_lm(vocab_size, emb_dim, rnn_type, hidden_size, num_layer):
     return cost, output
 
 
-def ngram_lm(vocab_size, emb_dim, hidden_size, num_layer):
+def ngram_lm(vocab_size, emb_dim, hidden_size, num_layer, gram_num=4):
     """
     N-Gram language model definition.
 
     :param vocab_size: size of vocab.
     :param emb_dim: embedding vector's dimension.
     :param hidden_size: size of unit.
-    :param num_layer: layer number.
+    :param num_layer: number of hidden layers.
+    :param gram_size: gram number in n-gram method
     :return: cost and output layer of model.
     """
 
     assert emb_dim > 0 and hidden_size > 0 and vocab_size > 0 and num_layer > 0
 
-    def wordemb(inlayer):
-        wordemb = paddle.layer.table_projection(
-            input=inlayer,
-            size=emb_dim,
-            param_attr=paddle.attr.Param(
-                name="_proj", initial_std=0.001, learning_rate=1, l2_rate=0))
-        return wordemb
-
     # input layers
-    first_word = paddle.layer.data(
-        name="first_word", type=paddle.data_type.integer_value(vocab_size))
-    second_word = paddle.layer.data(
-        name="second_word", type=paddle.data_type.integer_value(vocab_size))
-    third_word = paddle.layer.data(
-        name="third_word", type=paddle.data_type.integer_value(vocab_size))
-    fourth_word = paddle.layer.data(
-        name="fourth_word", type=paddle.data_type.integer_value(vocab_size))
+    emb_layers = []
+    for i in range(gram_num):
+        word = paddle.layer.data(
+            name="__word%02d__" % (i + 1),
+            type=paddle.data_type.integer_value(vocab_size))
+        emb = paddle.layer.embedding(
+            input=word,
+            size=emb_dim,
+            param_attr=paddle.attr.Param(name="_proj", initial_std=1e-3))
+        emb_layers.append(emb)
     next_word = paddle.layer.data(
-        name="next_word", type=paddle.data_type.integer_value(vocab_size))
-
-    # embedding layer
-    first_emb = wordemb(first_word)
-    second_emb = wordemb(second_word)
-    third_emb = wordemb(third_word)
-    fourth_emb = wordemb(fourth_word)
-
-    context_emb = paddle.layer.concat(
-        input=[first_emb, second_emb, third_emb, fourth_emb])
+        name="__next_word__", type=paddle.data_type.integer_value(vocab_size))
 
     # hidden layer
-    hidden = paddle.layer.fc(
-        input=context_emb, size=hidden_size, act=paddle.activation.Relu())
-    for _ in range(num_layer - 1):
+    for i in range(num_layer):
         hidden = paddle.layer.fc(
-            input=hidden, size=hidden_size, act=paddle.activation.Relu())
+            input=hidden if i else paddle.layer.concat(input=emb_layers),
+            size=hidden_size,
+            act=paddle.activation.Relu())
 
-    # fc(full connected) and output layer
     predict_word = paddle.layer.fc(
         input=[hidden], size=vocab_size, act=paddle.activation.Softmax())