Various fixes and changes

README.md

@@ -1,8 +1,7 @@
 # deepspeech.pytorch
 
-* Fix WER/CER measurements
 * Add tests for dataloading
-* Script to download an4 and create manifests (streamline rather than 2 separate scripts)
+* Fix validation. Assume problems with SequenceWise module and using batch norm in 3d mode.
 * Support LibriSpeech via multi-processed scripts
 
 Implementation of [Baidu Warp-CTC](https://github.com/baidu-research/warp-ctc) using pytorch.

data/get_an4.py

@@ -3,8 +3,12 @@
 import io
 import shutil
 
+import subprocess
+
+from data.utils import create_manifest
+
 parser = argparse.ArgumentParser(description='Processes and downloads an4.')
-parser.add_argument('--an4_path', default='dataset/', help='Path to save dataset')
+parser.add_argument('--an4_path', default='an4_dataset/', help='Path to save dataset')
 parser.add_argument('--sample_rate', default=16000, type=int, help='Sample rate')
 
 
@@ -62,13 +66,18 @@ def main():
  args = parser.parse_args()
  root_path = 'an4/'
  name = 'an4'
- os.system('wget http:https://www.speech.cs.cmu.edu/databases/an4/an4_raw.bigendian.tar.gz')
- os.system('tar -xzvf an4_raw.bigendian.tar.gz')
+ subprocess.call(['wget http:https://www.speech.cs.cmu.edu/databases/an4/an4_raw.bigendian.tar.gz'], shell=True)
+ subprocess.call(['tar -xzvf an4_raw.bigendian.tar.gz'], stdout=open(os.devnull, 'wb'), shell=True)
  os.makedirs(args.an4_path)
  format_data('train', name, 'an4_clstk')
  format_data('test', name, 'an4test_clstk')
  shutil.rmtree(root_path)
  os.remove('an4_raw.bigendian.tar.gz')
+ train_path = args.an4_path + '/train/'
+ test_path = args.an4_path + '/test/'
+ print ('Creating manifests...')
+ create_manifest(train_path, 'train')
+ create_manifest(test_path, 'test')
 
 
 if __name__ == '__main__':

data/utils.py

@@ -0,0 +1,37 @@
+import argparse
+import io
+import os
+
+import subprocess
+
+parser = argparse.ArgumentParser(description='Creates training and testing manifests')
+parser.add_argument('--root_path', default='an4_dataset', help='Path to the dataset')
+
+"""
+We need to add progress bars like will did in gen audio. Just copy that code.
+We also need a call (basically the same in the dataloader, a find that gives us the total number of wav files)
+"""
+
+
+def create_manifest(data_path, tag, ordered=True):
+ manifest_path = '%s_manifest.csv' % tag
+ file_paths = []
+ with os.popen('find %s -type f -name "*.wav"' % data_path) as pipe:
+ for file_path in pipe:
+ file_paths.append(file_path.strip())
+ if ordered:
+ print("Sorting files by length...")
+
+ def func(element):
+ output = subprocess.check_output(
+ ['soxi -D %s' % element.strip()],
+ shell=True
+ )
+ return float(output)
+
+ file_paths.sort(key=func)
+ with io.FileIO(manifest_path, "w") as file:
+ for wav_path in file_paths:
+ transcript_path = wav_path.replace('/wav/', '/txt/').replace('.wav', '.txt')
+ sample = os.path.abspath(wav_path) + ',' + os.path.abspath(transcript_path) + '\n'
+ file.write(sample)

main.py

@@ -11,9 +11,9 @@
 
 parser = argparse.ArgumentParser(description='DeepSpeech pytorch params')
 parser.add_argument('--train_manifest', metavar='DIR',
- help='path to train manifest csv', default='train_manifest.csv')
+ help='path to train manifest csv', default='data/train_manifest.csv')
 parser.add_argument('--test_manifest', metavar='DIR',
- help='path to test manifest csv', default='test_manifest.csv')
+ help='path to test manifest csv', default='data/test_manifest.csv')
 parser.add_argument('--sample_rate', default=16000, type=int, help='Sample rate')
 parser.add_argument('--batch_size', default=20, type=int, help='Batch size for training')
 parser.add_argument('--num_workers', default=4, type=int, help='Number of workers used in dataloading')
@@ -64,12 +64,12 @@ def main():
 
  train_dataloader_config = dict(type="audio,transcription",
  audio=audio_config,
- manifest_filename='train_manifest.csv',
+ manifest_filename=args.train_manifest,
  alphabet=alphabet,
  normalize=True)
  test_dataloader_config = dict(type="audio,transcription",
  audio=audio_config,
- manifest_filename='test_manifest.csv',
+ manifest_filename=args.test_manifest,
  alphabet=alphabet,
  normalize=True)
  train_loader = AudioDataLoader(AudioDataset(train_dataloader_config), args.batch_size,
@@ -183,8 +183,8 @@ def main():
  target_strings = decoder.process_strings(decoder.convert_to_strings(split_targets))
  wer, cer = 0, 0
  for x in range(len(target_strings)):
- wer += decoder.wer(decoded_output[x], target_strings[x])
- cer += decoder.cer(decoded_output[x], target_strings[x])
+ wer += decoder.wer(decoded_output[x], target_strings[x]) / float(len(target_strings[x].split()))
+ cer += decoder.cer(decoded_output[x], target_strings[x]) / float(len(target_strings[x]))
  total_cer += cer
  total_wer += wer
 

model.py

@@ -34,19 +34,20 @@ def __init__(self, input_size, hidden_size, bidirectional=False, batch_norm=True
  self.batch_norm_activate = batch_norm
  self.bidirectional = bidirectional
  self.batch_norm = nn.BatchNorm1d(input_size)
- self.rnn = nn.GRU(input_size=input_size, hidden_size=hidden_size,
- bidirectional=bidirectional, bias=False)
+ self.rnn = nn.LSTM(input_size=input_size, hidden_size=hidden_size,
+  bidirectional=bidirectional, bias=False)
  self.num_directions = 2 if bidirectional else 1
 
  def forward(self, x):
+ c0 = Variable(torch.zeros(self.num_directions, x.size(1), self.hidden_size).type_as(x.data))
  h0 = Variable(torch.zeros(self.num_directions, x.size(1), self.hidden_size).type_as(x.data))
  if self.batch_norm_activate:
  t, n = x.size(0), x.size(1)
  x = x.view(n, -1, t)
  x = self.batch_norm(x)
  x = x.transpose(1, 2).transpose(0, 1)
  x = x.contiguous()
- x, _ = self.rnn(x, h0)
+ x, _ = self.rnn(x, (c0, h0))
  if self.bidirectional:
  x = x.view(x.size(0), x.size(1), 2, -1).sum(2).view(x.size(0), x.size(1), -1) # (TxNxD*2) -> (TxNxD) by sum
  return x