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.

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
+```

audio_data_utils.py

@@ -16,7 +16,7 @@ logger = logging.getLogger(__name__)
 
 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.
 

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 @@ parser.add_argument(
     "--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__':

librispeech.py

@@ -7,6 +7,7 @@
 """
 
 import paddle.v2 as paddle
+from paddle.v2.dataset.common import md5file
 import os
 import wget
 import tarfile
@@ -14,11 +15,22 @@ 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.')
@@ -26,27 +38,33 @@ 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__':

model.py

@@ -24,45 +24,23 @@ def conv_bn_layer(input, filter_size, num_channels_in, num_channels_out, stride,
     return paddle.layer.batch_norm(input=conv_layer, act=act)
 
 
-def bidirectonal_simple_rnn_bn_layer(name, input, size, act):
+def bidirectional_simple_rnn_bn_layer(name, input, size, act):
     """
-    Bidirectonal simple rnn layer with batch normalization.
-    The batch normalization is only performed on input-state projection
-    (sequence-wise normalization).
-
-    Question: does mean and variance statistics computed over the whole sequence
-    or just on each individual time steps?
+    Bidirectonal simple rnn layer with sequence-wise batch normalization.
+    The batch normalization is only performed on input-state weights.
     """
-
-    def __simple_rnn_step__(input):
-        last_state = paddle.layer.memory(name=name + "_state", size=size)
-        input_fc = paddle.layer.fc(
-            input=input,
-            size=size,
-            act=paddle.activation.Linear(),
-            bias_attr=False)
+    # input-hidden weights shared across bi-direcitonal rnn.
+    input_proj = paddle.layer.fc(
+        input=input, size=size, act=paddle.activation.Linear(), bias_attr=False)
     # batch norm is only performed on input-state projection 
-        input_fc_bn = paddle.layer.batch_norm(
-            input=input_fc, act=paddle.activation.Linear())
-        state_fc = paddle.layer.fc(
-            input=last_state,
-            size=size,
-            act=paddle.activation.Linear(),
-            bias_attr=False)
-        return paddle.layer.addto(
-            name=name + "_state", input=[input_fc_bn, state_fc], act=act)
-
-    forward = paddle.layer.recurrent_group(
-        step=__simple_rnn_step__, input=input)
-    return forward
-    # argument reverse is not exposed in V2 recurrent_group
-    #backward = paddle.layer.recurrent_group(
-
-
-#step=__simple_rnn_step__,
-#input=input,
-#reverse=True)
-#return paddle.layer.concat(input=[forward, backward])
+    input_proj_bn = paddle.layer.batch_norm(
+        input=input_proj, act=paddle.activation.Linear())
+    # forward and backward in time
+    forward_simple_rnn = paddle.layer.recurrent(
+        input=input_proj_bn, act=act, reverse=False)
+    backward_simple_rnn = paddle.layer.recurrent(
+        input=input_proj_bn, act=act, reverse=True)
+    return paddle.layer.concat(input=[forward_simple_rnn, backward_simple_rnn])
 
 
 def conv_group(input, num_stacks):
@@ -86,7 +64,9 @@ def conv_group(input, num_stacks):
             stride=(1, 2),
             padding=(5, 10),
             act=paddle.activation.BRelu())
-    return conv
+    output_num_channels = 32
+    output_height = 160 // pow(2, num_stacks) + 1
+    return conv, output_num_channels, output_height
 
 
 def rnn_group(input, size, num_stacks):
@@ -95,7 +75,7 @@ def rnn_group(input, size, num_stacks):
     """
     output = input
     for i in xrange(num_stacks):
-        output = bidirectonal_simple_rnn_bn_layer(
+        output = bidirectional_simple_rnn_bn_layer(
             name=str(i), input=output, size=size, act=paddle.activation.BRelu())
     return output
 
@@ -125,15 +105,16 @@ def deep_speech2(audio_data,
     :rtype: tuple of LayerOutput
     """
     # convolution group
-    conv_group_output = conv_group(input=audio_data, num_stacks=num_conv_layers)
+    conv_group_output, conv_group_num_channels, conv_group_height = conv_group(
+        input=audio_data, num_stacks=num_conv_layers)
     # convert data form convolution feature map to sequence of vectors
     conv2seq = paddle.layer.block_expand(
         input=conv_group_output,
-        num_channels=32,
+        num_channels=conv_group_num_channels,
         stride_x=1,
         stride_y=1,
         block_x=1,
-        block_y=21)
+        block_y=conv_group_height)
     # rnn group
     rnn_group_output = rnn_group(
         input=conv2seq, size=rnn_size, num_stacks=num_rnn_layers)

requirements.sh

@@ -1,5 +0,0 @@
-pip install wget
-pip install soundfile
-
-# For Ubuntu only
-apt-get install libsndfile1

requirements.txt

@@ -0,0 +1,2 @@
+SoundFile==0.9.0.post1
+wget==3.2

train.py

@@ -3,6 +3,7 @@
 """
 
 import paddle.v2 as paddle
+import distutils.util
 import argparse
 import gzip
 import time
@@ -17,21 +18,61 @@ parser = argparse.ArgumentParser(
     description='Simplified version of DeepSpeech2 trainer.')
 parser.add_argument(
     "--batch_size", default=32, type=int, help="Minibatch size.")
-parser.add_argument("--trainer", default=1, type=int, help="Trainer number.")
 parser.add_argument(
-    "--num_passes", default=20, type=int, help="Training pass number.")
+    "--num_passes",
+    default=20,
+    type=int,
+    help="Training pass number. (default: %(default)s)")
 parser.add_argument(
-    "--num_conv_layers", default=3, 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=5, type=int, help="RNN layer number.")
+    "--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.")
+    "--rnn_layer_size",
+    default=512,
+    type=int,
+    help="RNN layer cell number. (default: %(default)s)")
 parser.add_argument(
-    "--use_gpu", default=True, type=bool, help="Use gpu or not.")
+    "--adam_learning_rate",
+    default=5e-4,
+    type=float,
+    help="Learning rate for ADAM Optimizer. (default: %(default)s)")
 parser.add_argument(
-    "--use_sortagrad", default=False, type=bool, help="Use sortagrad or not.")
+    "--use_gpu",
+    default=True,
+    type=distutils.util.strtobool,
+    help="Use gpu or not. (default: %(default)s)")
 parser.add_argument(
-    "--trainer_count", default=8, type=int, help="Trainer number.")
+    "--use_sortagrad",
+    default=False,
+    type=distutils.util.strtobool,
+    help="Use sortagrad or not. (default: %(default)s)")
+parser.add_argument(
+    "--trainer_count",
+    default=4,
+    type=int,
+    help="Trainer number. (default: %(default)s)")
+parser.add_argument(
+    "--normalizer_manifest_path",
+    default='./manifest.libri.train-clean-100',
+    type=str,
+    help="Manifest path for normalizer. (default: %(default)s)")
+parser.add_argument(
+    "--train_manifest_path",
+    default='./manifest.libri.train-clean-100',
+    type=str,
+    help="Manifest path for training. (default: %(default)s)")
+parser.add_argument(
+    "--dev_manifest_path",
+    default='./manifest.libri.dev-clean',
+    type=str,
+    help="Manifest path for validation. (default: %(default)s)")
 args = parser.parse_args()
 
 
@@ -39,37 +80,15 @@ def train():
     """
     DeepSpeech2 training.
     """
-    # create data readers
+    # initialize data generator
     data_generator = DataGenerator(
         vocab_filepath='eng_vocab.txt',
-        normalizer_manifest_path='./libri.manifest.train',
+        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)
-    train_batch_reader_sortagrad = data_generator.batch_reader_creator(
-        manifest_path='./libri.manifest.dev.small',
-        batch_size=args.batch_size // args.trainer,
-        padding_to=2000,
-        flatten=True,
-        sort_by_duration=True,
-        shuffle=False)
-    train_batch_reader_nosortagrad = data_generator.batch_reader_creator(
-        manifest_path='./libri.manifest.dev.small',
-        batch_size=args.batch_size // args.trainer,
-        padding_to=2000,
-        flatten=True,
-        sort_by_duration=False,
-        shuffle=True)
-    test_batch_reader = data_generator.batch_reader_creator(
-        manifest_path='./libri.manifest.test',
-        batch_size=args.batch_size // args.trainer,
-        padding_to=2000,
-        flatten=True,
-        sort_by_duration=False,
-        shuffle=False)
-    feeding = data_generator.data_name_feeding()
 
     # create network config
     dict_size = data_generator.vocabulary_size()
@@ -92,28 +111,58 @@ def train():
     # create parameters and optimizer
     parameters = paddle.parameters.create(cost)
     optimizer = paddle.optimizer.Adam(
-        learning_rate=5e-5, gradient_clipping_threshold=400)
+        learning_rate=args.adam_learning_rate, gradient_clipping_threshold=400)
     trainer = paddle.trainer.SGD(
         cost=cost, parameters=parameters, update_equation=optimizer)
 
+    # prepare data reader
+    train_batch_reader_sortagrad = data_generator.batch_reader_creator(
+        manifest_path=args.train_manifest_path,
+        batch_size=args.batch_size // args.trainer_count,
+        padding_to=2000,
+        flatten=True,
+        sort_by_duration=True,
+        shuffle=False)
+    train_batch_reader_nosortagrad = data_generator.batch_reader_creator(
+        manifest_path=args.train_manifest_path,
+        batch_size=args.batch_size // args.trainer_count,
+        padding_to=2000,
+        flatten=True,
+        sort_by_duration=False,
+        shuffle=True)
+    test_batch_reader = data_generator.batch_reader_creator(
+        manifest_path=args.dev_manifest_path,
+        batch_size=args.batch_size // args.trainer_count,
+        padding_to=2000,
+        flatten=True,
+        sort_by_duration=False,
+        shuffle=False)
+    feeding = data_generator.data_name_feeding()
+
     # create event handler
     def event_handler(event):
         global start_time
+        global cost_sum
+        global cost_counter
         if isinstance(event, paddle.event.EndIteration):
-            if event.batch_id % 10 == 0:
+            cost_sum += event.cost
+            cost_counter += 1
+            if event.batch_id % 50 == 0:
                 print "\nPass: %d, Batch: %d, TrainCost: %f" % (
-                    event.pass_id, event.batch_id, event.cost)
+                    event.pass_id, event.batch_id, cost_sum / cost_counter)
+                cost_sum, cost_counter = 0.0, 0
+                with gzip.open("params.tar.gz", 'w') as f:
+                    parameters.to_tar(f)
             else:
                 sys.stdout.write('.')
                 sys.stdout.flush()
         if isinstance(event, paddle.event.BeginPass):
             start_time = time.time()
+            cost_sum, cost_counter = 0.0, 0
         if isinstance(event, paddle.event.EndPass):
             result = trainer.test(reader=test_batch_reader, feeding=feeding)
-            print "\n------- Time: %d,  Pass: %d, TestCost: %s" % (
+            print "\n------- Time: %d sec,  Pass: %d, ValidationCost: %s" % (
                 time.time() - start_time, event.pass_id, result.cost)
-            with gzip.open("params.tar.gz", 'w') as f:
-                parameters.to_tar(f)
 
     # run train
     # first pass with sortagrad