fix scripts

examples/wenetspeech/asr1/conf/conformer.yaml

@@ -52,6 +52,7 @@ test_manifest: data/test_meeting/data.list
 use_stream_data: True
 unit_type: 'char'
 vocab_filepath: data/lang_char/vocab.txt 
+preprocess_config: conf/preprocess.yaml
 cmvn_file: data/mean_std.json
 spm_model_prefix: ''
 feat_dim: 80
@@ -65,30 +66,17 @@ maxlen_in: 1200 # if input length(number of frames) > maxlen-in, data is automa
 minlen_out: 0
 maxlen_out: 150 # if output length(number of tokens) > maxlen-out, data is automatically removed
 resample_rate: 16000
-shuffle_size: 1500
-sort_size: 1000
+shuffle_size: 1500 # read number of 'shuffle_size' data as a chunk, shuffle the data in the chunk
+sort_size: 1000 # read number of 'sort_size' data as a chunk, sort the data in the chunk 
 num_workers: 8
 prefetch_factor: 10
 dist_sampler: True
 num_encs: 1
-augment_conf:
- max_w: 80
- w_inplace: True
- w_mode: "PIL"
- max_f: 30
- num_f_mask: 2
- f_inplace: True
- f_replace_with_zero: False
- max_t: 40
- num_t_mask: 2
- t_inplace: True
- t_replace_with_zero: False
-
 
 ###########################################
 # Training #
 ###########################################
-n_epoch: 30 
+n_epoch: 32 
 accum_grad: 32
 global_grad_clip: 5.0
 log_interval: 100

examples/wenetspeech/asr1/local/data.sh

@@ -90,8 +90,8 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
  for x in $dev_set $test_sets ${train_set}; do
  dst=$shards_dir/$x
  mkdir -p $dst
- utils/make_filted_shard_list.py --resample 16000 --num_utts_per_shard 1000 \
- --do_filter --num_node 1 --num_gpus_per_node 8 \
+ utils/make_filted_shard_list.py --num_node 1 --num_gpus_per_node 8 --num_utts_per_shard 1000 \
+ --do_filter --resample 16000  \
  --num_threads 32 --segments data/$x/segments \
  data/$x/wav.scp data/$x/text \
  $(realpath $dst) data/$x/data.list

paddlespeech/audio/streamdata/__init__.py

@@ -1,4 +1,5 @@
 # Copyright (c) 2017-2019 NVIDIA CORPORATION. All rights reserved.
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
 # See the LICENSE file for licensing terms (BSD-style).
 # Modified from https://github.com/webdataset/webdataset
 #
@@ -26,22 +27,22 @@
  pipelinefilter,
  rename,
  rename_keys,
- rsample,
+ audio_resample,
  select,
  shuffle,
  slice,
  to_tuple,
  transform_with,
  unbatched,
  xdecode,
- data_filter,
- tokenize,
- resample,
- compute_fbank,
- spec_aug,
+ audio_data_filter,
+ audio_tokenize,
+ audio_resample,
+ audio_compute_fbank,
+ audio_spec_aug,
  sort,
- padding,
- cmvn,
+ audio_padding,
+ audio_cmvn,
  placeholder,
 )
 from .handlers import (

paddlespeech/audio/streamdata/autodecode.py

@@ -291,12 +291,12 @@ def torch_video(key, data):
 
 
 ################################################################
-# paddleaudio
+# paddlespeech.audio
 ################################################################
 
 
 def paddle_audio(key, data):
- """Decode audio using the paddleaudio library.
+ """Decode audio using the paddlespeech.audio library.
 
  :param key: file name extension
  :param data: data to be decoded
@@ -305,13 +305,13 @@ def paddle_audio(key, data):
  if extension not in ["flac", "mp3", "sox", "wav", "m4a", "ogg", "wma"]:
  return None
 
- import paddleaudio
+ import paddlespeech.audio
 
  with tempfile.TemporaryDirectory() as dirname:
  fname = os.path.join(dirname, f"file.{extension}")
  with open(fname, "wb") as stream:
  stream.write(data)
- return paddleaudio.load(fname)
+ return paddlespeech.audio.load(fname)
 
 
 ################################################################

paddlespeech/audio/streamdata/compat.py

@@ -78,29 +78,29 @@ def extract_keys(self, *args, **kw):
  def xdecode(self, *args, **kw):
  return self.compose(filters.xdecode(*args, **kw))
 
- def data_filter(self, *args, **kw):
- return self.compose(filters.data_filter(*args, **kw))
+ def audio_data_filter(self, *args, **kw):
+ return self.compose(filters.audio_data_filter(*args, **kw))
 
- def tokenize(self, *args, **kw):
- return self.compose(filters.tokenize(*args, **kw))
+ def audio_tokenize(self, *args, **kw):
+ return self.compose(filters.audio_tokenize(*args, **kw))
 
  def resample(self, *args, **kw):
  return self.compose(filters.resample(*args, **kw)) 
 
- def compute_fbank(self, *args, **kw):
- return self.compose(filters.compute_fbank(*args, **kw))
+ def audio_compute_fbank(self, *args, **kw):
+ return self.compose(filters.audio_compute_fbank(*args, **kw))
 
- def spec_aug(self, *args, **kw):
- return self.compose(filters.spec_aug(*args, **kw))
+ def audio_spec_aug(self, *args, **kw):
+ return self.compose(filters.audio_spec_aug(*args, **kw))
 
  def sort(self, size=500):
  return self.compose(filters.sort(size))
 
- def padding(self):
- return self.compose(filters.padding())
+ def audio_padding(self):
+ return self.compose(filters.audio_padding())
 
- def cmvn(self, cmvn_file):
- return self.compose(filters.cmvn(cmvn_file))
+ def audio_cmvn(self, cmvn_file):
+ return self.compose(filters.audio_cmvn(cmvn_file))
 
 class WebDataset(DataPipeline, FluidInterface):
  """Small fluid-interface wrapper for DataPipeline."""

paddlespeech/audio/streamdata/filters.py

@@ -579,7 +579,7 @@ def _xdecode(
 
 
 
-def _data_filter(source,
+def _audio_data_filter(source,
  frame_shift=10,
  max_length=10240,
  min_length=10,
@@ -629,9 +629,9 @@ def _data_filter(source,
  continue
  yield sample
 
-data_filter = pipelinefilter(_data_filter)
+audio_data_filter = pipelinefilter(_audio_data_filter)
 
-def _tokenize(source,
+def _audio_tokenize(source,
  symbol_table,
  bpe_model=None,
  non_lang_syms=None,
@@ -693,9 +693,9 @@ def _tokenize(source,
  sample['label'] = label
  yield sample
 
-tokenize = pipelinefilter(_tokenize)
+audio_tokenize = pipelinefilter(_audio_tokenize)
 
-def _resample(source, resample_rate=16000):
+def _audio_resample(source, resample_rate=16000):
  """ Resample data.
  Inplace operation.
 
@@ -718,9 +718,9 @@ def _resample(source, resample_rate=16000):
  ))
  yield sample
 
-resample = pipelinefilter(_resample)
+audio_resample = pipelinefilter(_audio_resample)
 
-def _compute_fbank(source,
+def _audio_compute_fbank(source,
  num_mel_bins=80,
  frame_length=25,
  frame_shift=10,
@@ -756,9 +756,9 @@ def _compute_fbank(source,
  yield dict(fname=sample['fname'], label=sample['label'], feat=mat)
 
 
-compute_fbank = pipelinefilter(_compute_fbank)
+audio_compute_fbank = pipelinefilter(_audio_compute_fbank)
 
-def _spec_aug(source,
+def _audio_spec_aug(source,
  max_w=5, 
  w_inplace=True, 
  w_mode="PIL",
@@ -799,7 +799,7 @@ def _spec_aug(source,
  sample['feat'] = paddle.to_tensor(x, dtype=paddle.float32)
  yield sample
 
-spec_aug = pipelinefilter(_spec_aug)
+audio_spec_aug = pipelinefilter(_audio_spec_aug)
 
 
 def _sort(source, sort_size=500):
@@ -881,7 +881,7 @@ def dynamic_batched(source, max_frames_in_batch=12000):
  yield buf
 
 
-def _padding(source):
+def _audio_padding(source):
  """ Padding the data into training data
 
  Args:
@@ -914,9 +914,9 @@ def _padding(source):
  yield (sorted_keys, padded_feats, feats_lengths, padding_labels, 
  label_lengths)
 
-padding = pipelinefilter(_padding)
+audio_padding = pipelinefilter(_audio_padding)
 
-def _cmvn(source, cmvn_file):
+def _audio_cmvn(source, cmvn_file):
  global_cmvn = GlobalCMVN(cmvn_file)
  for batch in source:
  sorted_keys, padded_feats, feats_lengths, padding_labels, label_lengths = batch
@@ -926,7 +926,7 @@ def _cmvn(source, cmvn_file):
  yield (sorted_keys, padded_feats, feats_lengths, padding_labels, 
  label_lengths)
 
-cmvn = pipelinefilter(_cmvn)
+audio_cmvn = pipelinefilter(_audio_cmvn)
 
 def _placeholder(source):
  for data in source:

paddlespeech/audio/streamdata/tariterators.py

@@ -21,7 +21,7 @@
 meta_prefix = "__"
 meta_suffix = "__"
 
-from ... import audio as paddleaudio
+import paddlespeech
 import paddle
 import numpy as np
 
@@ -118,7 +118,7 @@ def tar_file_iterator(
  assert pos > 0
  prefix, postfix = name[:pos], name[pos + 1:]
  if postfix == 'wav':
- waveform, sample_rate = paddleaudio.load(stream.extractfile(tarinfo), normal=False)
+ waveform, sample_rate = paddlespeech.audio.load(stream.extractfile(tarinfo), normal=False)
  result = dict(fname=prefix, wav=waveform, sample_rate = sample_rate)
  else:
  txt = stream.extractfile(tarinfo).read().decode('utf8').strip()
@@ -167,7 +167,7 @@ def tar_file_and_group_iterator(
  if postfix == 'txt':
  example['txt'] = file_obj.read().decode('utf8').strip()
  elif postfix in AUDIO_FORMAT_SETS:
- waveform, sample_rate = paddleaudio.load(file_obj, normal=False)
+ waveform, sample_rate = paddlespeech.audio.load(file_obj, normal=False)
  waveform = paddle.to_tensor(np.expand_dims(np.array(waveform),0), dtype=paddle.float32)
 
  example['wav'] = waveform