1. Fix incorrect decoder result printing.

2. Fix incorrect batch-norm usage in RNN.
3. Fix overlapping train/dev/test manfests.
4. Update README.md and requirements.txt.
5. Expose more arguments to users in argparser.
6. Update all other details.

deep_speech_2/README.md

@@ -1,9 +1,58 @@
 # Deep Speech 2 on PaddlePaddle
 
+## Quick Start
+
+### Installation
+
+Please replace `$PADDLE_INSTALL_DIR` with your paddle installation directory.
+
+```
+pip install -r requirements.txt
+export LD_LIBRARY_PATH=$PADDLE_INSTALL_DIR/Paddle/third_party/install/warpctc/lib:$LD_LIBRARY_PATH
+```
+
+For some machines, we also need to install libsndfile1. Details to be added.
+
+### Preparing Dataset(s)
+
 ```
-sh requirements.sh
 python librispeech.py
-python train.py
 ```
 
-Please add warp-ctc library path (usually $PADDLE_INSTALL_DIR/Paddle/third_party/install/warpctc/lib) to LD_LIBRARY_PATH.
+More help for arguments:
+
+```
+python librispeech.py --help
+```
+
+### Traininig
+
+For GPU Training:
+
+```
+CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --trainer_count 4
+```
+
+For CPU Training:
+
+```
+python train.py --trainer_count 8 --use_gpu False
+```
+
+More help for arguments:
+
+```
+python train.py --help
+```
+
+### Inferencing
+
+```
+python infer.py
+```
+
+More help for arguments:
+
+```
+python infer.py --help
+```

deep_speech_2/audio_data_utils.py

@@ -16,7 +16,7 @@
 
 class DataGenerator(object):
     """
-    DataGenerator provides basic audio data preprocessing pipeline, and offer
+    DataGenerator provides basic audio data preprocessing pipeline, and offers
     both instance-level and batch-level data reader interfaces.
     Normalized FFT are used as audio features here.
 

deep_speech_2/infer.py

@@ -4,9 +4,10 @@
 
 import paddle.v2 as paddle
 from itertools import groupby
+import distutils.util
 import argparse
 import gzip
-import audio_data_utils
+from audio_data_utils import DataGenerator
 from model import deep_speech2
 
 parser = argparse.ArgumentParser(
@@ -15,15 +16,42 @@
     "--num_samples",
     default=10,
     type=int,
-    help="Number of samples for inference.")
+    help="Number of samples for inference. (default: %(default)s)")
 parser.add_argument(
-    "--num_conv_layers", default=2, type=int, help="Convolution layer number.")
+    "--num_conv_layers",
+    default=2,
+    type=int,
+    help="Convolution layer number. (default: %(default)s)")
+parser.add_argument(
+    "--num_rnn_layers",
+    default=3,
+    type=int,
+    help="RNN layer number. (default: %(default)s)")
+parser.add_argument(
+    "--rnn_layer_size",
+    default=512,
+    type=int,
+    help="RNN layer cell number. (default: %(default)s)")
+parser.add_argument(
+    "--use_gpu",
+    default=True,
+    type=distutils.util.strtobool,
+    help="Use gpu or not. (default: %(default)s)")
 parser.add_argument(
-    "--num_rnn_layers", default=3, type=int, help="RNN layer number.")
+    "--normalizer_manifest_path",
+    default='./manifest.libri.train-clean-100',
+    type=str,
+    help="Manifest path for normalizer. (default: %(default)s)")
 parser.add_argument(
-    "--rnn_layer_size", default=512, type=int, help="RNN layer cell number.")
+    "--decode_manifest_path",
+    default='./manifest.libri.test-clean',
+    type=str,
+    help="Manifest path for decoding. (default: %(default)s)")
 parser.add_argument(
-    "--use_gpu", default=True, type=bool, help="Use gpu or not.")
+    "--model_filepath",
+    default='./params.tar.gz',
+    type=str,
+    help="Model filepath. (default: %(default)s)")
 args = parser.parse_args()
 
 
@@ -39,18 +67,27 @@ def remove_duplicate_and_blank(id_list, blank_id):
     return [id for id in id_list if id != blank_id]
 
 
-def max_infer():
+def best_path_decode():
     """
     Max-ctc-decoding for DeepSpeech2.
     """
+    # initialize data generator
+    data_generator = DataGenerator(
+        vocab_filepath='eng_vocab.txt',
+        normalizer_manifest_path=args.normalizer_manifest_path,
+        normalizer_num_samples=200,
+        max_duration=20.0,
+        min_duration=0.0,
+        stride_ms=10,
+        window_ms=20)
     # create network config
-    _, vocab_list = audio_data_utils.get_vocabulary()
-    dict_size = len(vocab_list)
+    dict_size = data_generator.vocabulary_size()
+    vocab_list = data_generator.vocabulary_list()
     audio_data = paddle.layer.data(
         name="audio_spectrogram",
         height=161,
-        width=1000,
-        type=paddle.data_type.dense_vector(161000))
+        width=2000,
+        type=paddle.data_type.dense_vector(322000))
     text_data = paddle.layer.data(
         name="transcript_text",
         type=paddle.data_type.integer_value_sequence(dict_size))
@@ -64,19 +101,17 @@ def max_infer():
 
     # load parameters
     parameters = paddle.parameters.Parameters.from_tar(
-        gzip.open("params.tar.gz"))
+        gzip.open(args.model_filepath))
 
     # prepare infer data
-    feeding = {
-        "audio_spectrogram": 0,
-        "transcript_text": 1,
-    }
-    test_batch_reader = audio_data_utils.padding_batch_reader(
-        paddle.batch(
-            audio_data_utils.reader_creator(
-                manifest_path="./libri.manifest.test", sort_by_duration=False),
-            batch_size=args.num_samples),
-        padding=[-1, 1000])
+    feeding = data_generator.data_name_feeding()
+    test_batch_reader = data_generator.batch_reader_creator(
+        manifest_path=args.decode_manifest_path,
+        batch_size=args.num_samples,
+        padding_to=2000,
+        flatten=True,
+        sort_by_duration=False,
+        shuffle=False)
     infer_data = test_batch_reader().next()
 
     # run max-ctc-decoding
@@ -89,7 +124,7 @@ def max_infer():
     # postprocess
     instance_length = len(max_id_results) / args.num_samples
     instance_list = [
-        max_id_results[i:i + instance_length]
+        max_id_results[i * instance_length:(i + 1) * instance_length]
         for i in xrange(0, args.num_samples)
     ]
     for i, instance in enumerate(instance_list):
@@ -102,7 +137,7 @@ def max_infer():
 
 def main():
     paddle.init(use_gpu=args.use_gpu, trainer_count=1)
-    max_infer()
+    best_path_decode()
 
 
 if __name__ == '__main__':

deep_speech_2/librispeech.py

@@ -7,46 +7,64 @@
 """
 
 import paddle.v2 as paddle
+from paddle.v2.dataset.common import md5file
 import os
 import wget
 import tarfile
 import argparse
 import soundfile
 import json
 
-DATA_HOME = os.path.expanduser('~/.cache/paddle/dataset/speech')
+DATA_HOME = os.path.expanduser('~/.cache2/paddle/dataset/speech')
 
-URL_TEST = "http://www.openslr.org/resources/12/test-clean.tar.gz"
-URL_DEV = "http://www.openslr.org/resources/12/dev-clean.tar.gz"
-URL_TRAIN = "http://www.openslr.org/resources/12/train-clean-100.tar.gz"
+URL_ROOT = "http://www.openslr.org/resources/12"
+URL_TEST_CLEAN = URL_ROOT + "/test-clean.tar.gz"
+URL_TEST_OTHER = URL_ROOT + "/test-other.tar.gz"
+URL_DEV_CLEAN = URL_ROOT + "/dev-clean.tar.gz"
+URL_DEV_OTHER = URL_ROOT + "/dev-other.tar.gz"
+URL_TRAIN_CLEAN_100 = URL_ROOT + "/train-clean-100.tar.gz"
+URL_TRAIN_CLEAN_360 = URL_ROOT + "/train-clean-360.tar.gz"
+URL_TRAIN_OTHER_500 = URL_ROOT + "/train-other-500.tar.gz"
+
+MD5_TEST_CLEAN = "32fa31d27d2e1cad72775fee3f4849a9"
+MD5_DEV_CLEAN = "42e2234ba48799c1f50f24a7926300a1"
+MD5_TRAIN_CLEAN_100 = "2a93770f6d5c6c964bc36631d331a522"
+MD5_TRAIN_CLEAN_360 = "c0e676e450a7ff2f54aeade5171606fa"
+MD5_TRAIN_CLEAN_500 = "d1a0fd59409feb2c614ce4d30c387708"
 
 parser = argparse.ArgumentParser(
     description='Downloads and prepare LibriSpeech dataset.')
 parser.add_argument(
     "--target_dir",
     default=DATA_HOME + "/Libri",
     type=str,
-    help="Directory to save the dataset.")
+    help="Directory to save the dataset. (default: %(default)s)")
 parser.add_argument(
-    "--manifest",
-    default="./libri.manifest",
+    "--manifest_prefix",
+    default="manifest.libri",
     type=str,
-    help="Filepath prefix for output manifests.")
+    help="Filepath prefix for output manifests. (default: %(default)s)")
 args = parser.parse_args()
 
 
-def download(url, target_dir):
-    if not os.path.exists(target_dir):
-        os.makedirs(target_dir)
+def download(url, md5sum, target_dir):
+    """
+    Download file from url to target_dir, and check md5sum.
+    """
+    if not os.path.exists(target_dir): os.makedirs(target_dir)
     filepath = os.path.join(target_dir, url.split("/")[-1])
-    if not os.path.exists(filepath):
+    if not (os.path.exists(filepath) and md5file(filepath) == md5sum):
         print("Downloading %s ..." % url)
         wget.download(url, target_dir)
-        print("")
+        print("\nMD5 Chesksum %s ..." % filepath)
+        assert md5file(filepath) == md5sum, "MD5 checksum failed."
     return filepath
 
 
 def unpack(filepath, target_dir):
+    """
+    Unpack the file to the target_dir.
+    """
     print("Unpacking %s ..." % filepath)
     tar = tarfile.open(filepath)
     tar.extractall(target_dir)
@@ -55,6 +73,14 @@ def unpack(filepath, target_dir):
 
 
 def create_manifest(data_dir, manifest_path):
+    """
+    Create a manifest file summarizing the dataset (list of filepath and meta
+    data).
+
+    Each line of the manifest contains one audio clip filepath, its
+    transcription text string, and its duration. Manifest file servers as a
+    unified interfance to organize data sets.
+    """
     print("Creating manifest %s ..." % manifest_path)
     json_lines = []
     for subfolder, _, filelist in os.walk(data_dir):
@@ -81,25 +107,31 @@ def create_manifest(data_dir, manifest_path):
             out_file.write(line + '\n')
 
 
-def prepare_dataset(url, target_dir, manifest_path):
-    filepath = download(url, target_dir)
+def prepare_dataset(url, md5sum, target_dir, manifest_path):
+    """
+    Download, unpack and create summmary manifest file.
+    """
+    filepath = download(url, md5sum, target_dir)
     unpacked_dir = unpack(filepath, target_dir)
     create_manifest(unpacked_dir, manifest_path)
 
 
 def main():
     prepare_dataset(
-        url=URL_TEST,
-        target_dir=os.path.join(args.target_dir),
-        manifest_path=args.manifest + ".test")
+        url=URL_TEST_CLEAN,
+        md5sum=MD5_TEST_CLEAN,
+        target_dir=os.path.join(args.target_dir, "test-clean"),
+        manifest_path=args.manifest_prefix + ".test-clean")
     prepare_dataset(
-        url=URL_DEV,
-        target_dir=os.path.join(args.target_dir),
-        manifest_path=args.manifest + ".dev")
+        url=URL_DEV_CLEAN,
+        md5sum=MD5_DEV_CLEAN,
+        target_dir=os.path.join(args.target_dir, "dev-clean"),
+        manifest_path=args.manifest_prefix + ".dev-clean")
     prepare_dataset(
-        url=URL_TRAIN,
-        target_dir=os.path.join(args.target_dir),
-        manifest_path=args.manifest + ".train")
+        url=URL_TRAIN_CLEAN_100,
+        md5sum=MD5_TRAIN_CLEAN_100,
+        target_dir=os.path.join(args.target_dir, "train-clean-100"),
+        manifest_path=args.manifest_prefix + ".train-clean-100")
 
 
 if __name__ == '__main__':