fix ds2 scripts and bugs

deepspeech/exps/deepspeech2/bin/infer.py

-# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Inferer for DeepSpeech2 model."""
-from deepspeech.exps.deepspeech2.config import get_cfg_defaults
-from deepspeech.exps.deepspeech2.model import DeepSpeech2Tester as Tester
-from deepspeech.training.cli import default_argument_parser
-from deepspeech.utils.utility import print_arguments
-
-# TODO(hui zhang): dynamic load 
-
-
-def main_sp(config, args):
-    exp = Tester(config, args)
-    exp.setup()
-    exp.run_test()
-
-
-def main(config, args):
-    main_sp(config, args)
-
-
-if __name__ == "__main__":
-    parser = default_argument_parser()
-    args = parser.parse_args()
-    print_arguments(args, globals())
-
-    # https://yaml.org/type/float.html
-    config = get_cfg_defaults()
-    if args.config:
-        config.merge_from_file(args.config)
-    if args.opts:
-        config.merge_from_list(args.opts)
-    config.freeze()
-    print(config)
-    if args.dump_config:
-        with open(args.dump_config, 'w') as f:
-            print(config, file=f)
-
-    main(config, args)

deepspeech/exps/deepspeech2/model.py

@@ -113,6 +113,7 @@ class DeepSpeech2Trainer(Trainer):
         if self.parallel:
             model = paddle.DataParallel(model)
 
+        logger.info(f"{model}")
         layer_tools.print_params(model, logger.info)
 
         grad_clip = ClipGradByGlobalNormWithLog(
@@ -192,7 +193,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
             trans.append(''.join([chr(i) for i in ids]))
         return trans
 
-    def compute_metrics(self, audio, texts, audio_len, texts_len):
+    def compute_metrics(self, audio, audio_len, texts, texts_len):
         cfg = self.config.decoding
         errors_sum, len_refs, num_ins = 0.0, 0, 0
         errors_func = error_rate.char_errors if cfg.error_rate_type == 'cer' else error_rate.word_errors
@@ -253,7 +254,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
         msg = "Test: "
         msg += "epoch: {}, ".format(self.epoch)
         msg += "step: {}, ".format(self.iteration)
-        msg += ", Final error rate [%s] (%d/%d) = %f" % (
+        msg += "Final error rate [%s] (%d/%d) = %f" % (
             error_rate_type, num_ins, num_ins, errors_sum / len_refs)
         logger.info(msg)
 
@@ -319,8 +320,9 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
         config.defrost()
         # return raw text
 
-        config.data.manifest = config.data.test_manifest
+        config.data.keep_transcription_text = True
         config.data.augmentation_config = ""
+        config.data.manifest = config.data.test_manifest
         test_dataset = ManifestDataset.from_config(config)
 
         # return text ord id

deepspeech/frontend/normalizer.py

@@ -131,8 +131,8 @@ class FeatureNormalizer(object):
     def _read_mean_std_from_file(self, filepath, eps=1e-20):
         """Load mean and std from file."""
         mean, istd = load_cmvn(filepath, filetype='json')
-        self._mean = mean
-        self._istd = istd
+        self._mean = np.expand_dims(mean, axis=-1)
+        self._istd = np.expand_dims(istd, axis=-1)
 
     def write_to_file(self, filepath):
         """Write the mean and stddev to the file.

deepspeech/io/collator.py

@@ -60,7 +60,7 @@ class SpeechCollator():
             # else text is string, convert to unicode ord
             tokens = []
             if self._keep_transcription_text:
-                assert isinstance(text, str), type(text)
+                assert isinstance(text, str), (type(text), text)
                 tokens = [ord(t) for t in text]
             else:
                 tokens = text  # token ids

deepspeech/models/deepspeech2.py

@@ -154,9 +154,9 @@ class DeepSpeech2Model(nn.Layer):
         assert (self.encoder.output_size == rnn_size * 2)
 
         self.decoder = CTCDecoder(
-            odim=dict_size + 1,  # <blank> is append after vocab
+            odim=dict_size,  # <blank> is in  vocab
             enc_n_units=self.encoder.output_size,
-            blank_id=dict_size,  # last token is <blank>
+            blank_id=0,  # first token is <blank>
             dropout_rate=0.0,
             reduction=True,  # sum
             batch_average=True)  # sum / batch_size

deepspeech/models/u2.py

@@ -63,7 +63,7 @@ class U2BaseModel(nn.Module):
         default = CfgNode()
         # allow add new item when merge_with_file
         default.cmvn_file = ""
-        default.cmvn_file_type = "npz"
+        default.cmvn_file_type = "json"
         default.input_dim = 0
         default.output_dim = 0
         # encoder related

deepspeech/modules/mask.py

@@ -40,7 +40,8 @@ def sequence_mask(x_len, max_len=None, dtype='float32'):
         [[1., 1., 0., 0.],
          [1., 1., 1., 1.]]
     """
-    assert x_len.dim() == 1
+    # (TODO: Hui Zhang): jit not support Tenosr.dim() and Tensor.ndim
+    # assert x_len.dim() == 1, (x_len.dim(), x_len)
     max_len = max_len or x_len.max()
     x_len = paddle.unsqueeze(x_len, -1)
     row_vector = paddle.arange(max_len)

deepspeech/training/trainer.py

@@ -127,7 +127,12 @@ class Trainer():
     @mp_tools.rank_zero_only
     def save(self, tag=None, infos: dict=None):
         """Save checkpoint (model parameters and optimizer states).
+
+        Args:
+            tag (int or str, optional): None for step, else using tag, e.g epoch. Defaults to None.
+            infos (dict, optional): meta data to save. Defaults to None.
         """
+
         infos = infos if infos else dict()
         infos.update({
             "step": self.iteration,
@@ -220,7 +225,7 @@ class Trainer():
                     'epoch', {'cv_loss': cv_loss,
                               'lr': self.lr_scheduler()}, self.epoch)
 
-            self.save(infos={'val_loss': cv_loss})
+            self.save(tag=self.epoch, infos={'val_loss': cv_loss})
             self.lr_scheduler.step()
             self.new_epoch()
 

examples/aishell/s0/conf/deepspeech2.yaml

@@ -7,14 +7,20 @@ data:
   vocab_filepath: data/vocab.txt 
   augmentation_config: conf/augmentation.json
   batch_size: 64 # one gpu
-  max_duration: 27.0
-  min_duration: 0.0
+  min_input_len: 0.0
+  max_input_len: 27.0 # second
+  min_output_len: 0.0
+  max_output_len: 400.0
+  min_output_input_ratio: 0.05
+  max_output_input_ratio: 10.0
   specgram_type: linear
   target_sample_rate: 16000
   max_freq: None
   n_fft: None
   stride_ms: 10.0
   window_ms: 20.0
+  delta_delta: False
+  dither: 1.0
   use_dB_normalization: True
   target_dB: -20
   random_seed: 0
@@ -36,6 +42,7 @@ training:
   lr_decay: 0.83
   weight_decay: 1e-06
   global_grad_clip: 5.0
+  log_interval: 100
   
 decoding:
   batch_size: 128

examples/aishell/s0/local/avg.sh

+#! /usr/bin/env bash
+
+if [ $# != 2 ];then
+    echo "usage: ${0} ckpt_dir avg_num"
+    exit -1
+fi
+
+ckpt_dir=${1}
+average_num=${2}
+decode_checkpoint=${ckpt_dir}/avg_${average_num}.pdparams
+
+python3 -u ${MAIN_ROOT}/utils/avg_model.py \
+--dst_model ${decode_checkpoint} \
+--ckpt_dir ${ckpt_dir}  \
+--num ${average_num} \
+--val_best
+            
+if [ $? -ne 0 ]; then
+    echo "Failed in avg ckpt!"
+    exit 1
+fi
+
+exit 0
\ No newline at end of file

examples/aishell/s0/local/data.sh

@@ -43,17 +43,17 @@ fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     # compute mean and stddev for normalizer
+    num_workers=$(nproc)
     python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
     --manifest_path="data/manifest.train.raw" \
-    --specgram_type="fbank" \
-    --feat_dim=80 \
+    --specgram_type="linear" \
     --delta_delta=false \
     --stride_ms=10.0 \
-    --window_ms=25.0 \
+    --window_ms=20.0 \
     --sample_rate=16000 \
     --use_dB_normalization=False \
     --num_samples=-1 \
-    --num_workers=16 \
+    --num_workers=${num_workers} \
     --output_path="data/mean_std.json"
 
     if [ $? -ne 0 ]; then

examples/aishell/s0/local/export.sh

 #! /usr/bin/env bash
 
-if [ $# != 2 ];then
-    echo "usage: export ckpt_path jit_model_path"
+if [ $# != 3 ];then
+    echo "usage: $0 config_path ckpt_prefix jit_model_path"
     exit -1
 fi
 
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+echo "using $ngpu gpus..."
+
+config_path=$1
+ckpt_path_prefix=$2
+jit_model_export_path=$3
+
+device=gpu
+if [ ngpu == 0 ];then
+    device=cpu
+fi
+
 python3 -u ${BIN_DIR}/export.py \
---config conf/deepspeech2.yaml \
---checkpoint_path ${1} \
---export_path ${2} 
+--device ${device} \
+--nproc ${ngpu} \
+--config ${config_path} \
+--checkpoint_path ${ckpt_path_prefix} \
+--export_path ${jit_model_export_path}
 
 
 if [ $? -ne 0 ]; then
-    echo "Failed in evaluation!"
+    echo "Failed in export!"
     exit 1
 fi
 

examples/aishell/s0/local/test.sh

 #! /usr/bin/env bash
 
-if [[ $# != 1 ]]; then
-    echo "usage: $0 ckpt-path"
+if [ $# != 2 ];then
+    echo "usage: ${0} config_path ckpt_path_prefix"
     exit -1
 fi
 
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+echo "using $ngpu gpus..."
+
+device=gpu
+if [ ngpu == 0 ];then
+    device=cpu
+fi
+config_path=$1
+ckpt_prefix=$2
+
 # download language model
 bash local/download_lm_ch.sh
 if [ $? -ne 0 ]; then
    exit 1
 fi
 
-python3 -u ${BIN_DIR}/infer.py \
---device 'gpu' \
+python3 -u ${BIN_DIR}/test.py \
+--device ${device} \
 --nproc 1 \
---config conf/deepspeech2.yaml \
---checkpoint_path ${1} 
-
+--config ${config_path} \
+--result_file ${ckpt_prefix}.rsl \
+--checkpoint_path ${ckpt_prefix}
 
 if [ $? -ne 0 ]; then
-    echo "Failed in inference!"
+    echo "Failed in evaluation!"
     exit 1
 fi
 

examples/aishell/s0/local/train.sh

 #! /usr/bin/env bash
 
-# train model
-# if you wish to resume from an exists model, uncomment --init_from_pretrained_model
-#export FLAGS_sync_nccl_allreduce=0
+if [ $# != 2 ];then
+    echo "usage: CUDA_VISIBLE_DEVICES=0 ${0} config_path ckpt_name"
+    exit -1
+fi
 
-ngpu=$(echo ${CUDA_VISIBLE_DEVICES} | python -c 'import sys; a = sys.stdin.read(); print(len(a.split(",")));')
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
+config_path=$1
+ckpt_name=$2
+
+device=gpu
+if [ ngpu == 0 ];then
+    device=cpu
+fi
+
+mkdir -p exp
+
 python3 -u ${BIN_DIR}/train.py \
---device 'gpu' \
+--device ${device} \
 --nproc ${ngpu} \
---config conf/deepspeech2.yaml \
---output ckpt-${1}
-
+--config ${config_path} \
+--output exp/${ckpt_name}
 
 if [ $? -ne 0 ]; then
     echo "Failed in training!"
     exit 1
 fi
 
-
 exit 0

examples/aishell/s0/run.sh

 #!/bin/bash
-
+set -e
 source path.sh
-# only demos
 
-# prepare data
-bash ./local/data.sh
+gpus=0
+stage=0
+stop_stage=100
+conf_path=conf/deepspeech2.yaml
+ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}')
+avg_num=1
+avg_ckpt=avg_${avg_num}
+
+source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    # prepare data
+    bash ./local/data.sh || exit -1
+fi
 
-# train model
-CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./local/train.sh baseline
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    # train model, all `ckpt` under `exp` dir
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path}  ${ckpt}
+fi
 
-# test model
-CUDA_VISIBLE_DEVICES=0 bash ./local/test.sh
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    # avg n best model
+    ./local/avg.sh exp/${ckpt}/checkpoints ${avg_num}
+fi
 
-# infer model
-CUDA_VISIBLE_DEVICES=0 bash ./local/infer.sh ckpt/checkpoints/step-3284
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+    # test ckpt avg_n
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+fi
 
-# export model
-bash ./local/export.sh ckpt/checkpoints/step-3284 jit.model
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    # export ckpt avg_n
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
+fi

examples/aishell/s1/local/data.sh

-../../s0/local/data.sh
\ No newline at end of file

examples/aishell/s1/local/data.sh

+#! /usr/bin/env bash
+
+stage=-1
+stop_stage=100
+
+source ${MAIN_ROOT}/utils/parse_options.sh
+
+mkdir -p data
+TARGET_DIR=${MAIN_ROOT}/examples/dataset
+mkdir -p ${TARGET_DIR}
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+    # download data, generate manifests
+    python3 ${TARGET_DIR}/aishell/aishell.py \
+    --manifest_prefix="data/manifest" \
+    --target_dir="${TARGET_DIR}/aishell"
+    
+    if [ $? -ne 0 ]; then
+        echo "Prepare Aishell failed. Terminated."
+        exit 1
+    fi
+
+    for dataset in train dev test; do
+        mv data/manifest.${dataset} data/manifest.${dataset}.raw
+    done
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    # download data, generate manifests
+    # build vocabulary
+    python3 ${MAIN_ROOT}/utils/build_vocab.py \
+    --unit_type="char" \
+    --count_threshold=0 \
+    --vocab_path="data/vocab.txt" \
+    --manifest_paths "data/manifest.train.raw"
+    
+    if [ $? -ne 0 ]; then
+        echo "Build vocabulary failed. Terminated."
+        exit 1
+    fi
+fi
+
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    # compute mean and stddev for normalizer
+    num_workers=$(nproc)
+    python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
+    --manifest_path="data/manifest.train.raw" \
+    --specgram_type="fbank" \
+    --feat_dim=80 \
+    --delta_delta=false \
+    --stride_ms=10.0 \
+    --window_ms=25.0 \
+    --sample_rate=16000 \
+    --use_dB_normalization=False \
+    --num_samples=-1 \
+    --num_workers=${num_workers} \
+    --output_path="data/mean_std.json"
+    
+    if [ $? -ne 0 ]; then
+        echo "Compute mean and stddev failed. Terminated."
+        exit 1
+    fi
+fi
+
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    # format manifest with tokenids, vocab size
+    for dataset in train dev test; do
+        python3 ${MAIN_ROOT}/utils/format_data.py \
+        --feat_type "raw" \
+        --cmvn_path "data/mean_std.json" \
+        --unit_type "char" \
+        --vocab_path="data/vocab.txt" \
+        --manifest_path="data/manifest.${dataset}.raw" \
+        --output_path="data/manifest.${dataset}"
+    done
+    
+    if [ $? -ne 0 ]; then
+        echo "Formt mnaifest failed. Terminated."
+        exit 1
+    fi
+fi
+
+echo "Aishell data preparation done."
+exit 0

examples/dataset/librispeech/librispeech.py

@@ -24,6 +24,7 @@ import distutils.util
 import io
 import json
 import os
+from multiprocessing.pool import Pool
 
 import soundfile
 
@@ -122,42 +123,36 @@ def main():
     if args.target_dir.startswith('~'):
         args.target_dir = os.path.expanduser(args.target_dir)
 
-    prepare_dataset(
-        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_CLEAN,
-        md5sum=MD5_DEV_CLEAN,
-        target_dir=os.path.join(args.target_dir, "dev-clean"),
-        manifest_path=args.manifest_prefix + ".dev-clean")
+    tasks = [
+        (URL_TEST_CLEAN, MD5_TEST_CLEAN, os.path.join(args.target_dir,
+                                                      "test-clean"),
+         args.manifest_prefix + ".test-clean"),
+        (URL_DEV_CLEAN, MD5_DEV_CLEAN, os.path.join(
+            args.target_dir, "dev-clean"), args.manifest_prefix + ".dev-clean"),
+    ]
     if args.full_download:
-        prepare_dataset(
-            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")
-        prepare_dataset(
-            url=URL_TEST_OTHER,
-            md5sum=MD5_TEST_OTHER,
-            target_dir=os.path.join(args.target_dir, "test-other"),
-            manifest_path=args.manifest_prefix + ".test-other")
-        prepare_dataset(
-            url=URL_DEV_OTHER,
-            md5sum=MD5_DEV_OTHER,
-            target_dir=os.path.join(args.target_dir, "dev-other"),
-            manifest_path=args.manifest_prefix + ".dev-other")
-        prepare_dataset(
-            url=URL_TRAIN_CLEAN_360,
-            md5sum=MD5_TRAIN_CLEAN_360,
-            target_dir=os.path.join(args.target_dir, "train-clean-360"),
-            manifest_path=args.manifest_prefix + ".train-clean-360")
-        prepare_dataset(
-            url=URL_TRAIN_OTHER_500,
-            md5sum=MD5_TRAIN_OTHER_500,
-            target_dir=os.path.join(args.target_dir, "train-other-500"),
-            manifest_path=args.manifest_prefix + ".train-other-500")
+        tasks.extend([
+            (URL_TRAIN_CLEAN_100, MD5_TRAIN_CLEAN_100,
+             os.path.join(args.target_dir, "train-clean-100"),
+             args.manifest_prefix + ".train-clean-100"),
+            (URL_TEST_OTHER, MD5_TEST_OTHER, os.path.join(args.target_dir,
+                                                          "test-other"),
+             args.manifest_prefix + ".test-other"),
+            (URL_DEV_OTHER, MD5_DEV_OTHER, os.path.join(args.target_dir,
+                                                        "dev-other"),
+             args.manifest_prefix + ".dev-other"),
+            (URL_TRAIN_CLEAN_360, MD5_TRAIN_CLEAN_360,
+             os.path.join(args.target_dir, "train-clean-360"),
+             args.manifest_prefix + ".train-clean-360"),
+            (URL_TRAIN_OTHER_500, MD5_TRAIN_OTHER_500,
+             os.path.join(args.target_dir, "train-other-500"),
+             args.manifest_prefix + ".train-other-500"),
+        ])
+
+    with Pool(7) as pool:
+        pool.starmap(prepare_dataset, tasks)
+
+    print("Data download and manifest prepare done!")
 
 
 if __name__ == '__main__':

examples/dataset/mini_librispeech/mini_librispeech.py

@@ -23,6 +23,7 @@ import codecs
 import io
 import json
 import os
+from multiprocessing.pool import Pool
 
 import soundfile
 
@@ -103,16 +104,18 @@ def main():
     if args.target_dir.startswith('~'):
         args.target_dir = os.path.expanduser(args.target_dir)
 
-    prepare_dataset(
-        url=URL_TRAIN_CLEAN,
-        md5sum=MD5_TRAIN_CLEAN,
-        target_dir=os.path.join(args.target_dir, "train-clean"),
-        manifest_path=args.manifest_prefix + ".train-clean")
-    prepare_dataset(
-        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")
+    tasks = [
+        (URL_TRAIN_CLEAN, MD5_TRAIN_CLEAN,
+         os.path.join(args.target_dir, "train-clean"),
+         args.manifest_prefix + ".train-clean"),
+        (URL_DEV_CLEAN, MD5_DEV_CLEAN, os.path.join(
+            args.target_dir, "dev-clean"), args.manifest_prefix + ".dev-clean"),
+    ]
+
+    with Pool(2) as pool:
+        pool.starmap(prepare_dataset, tasks)
+
+    print("Data download and manifest prepare done!")
 
 
 if __name__ == '__main__':

examples/librispeech/s0/conf/deepspeech2.yaml

@@ -7,14 +7,20 @@ data:
   vocab_filepath: data/vocab.txt 
   augmentation_config: conf/augmentation.json
   batch_size: 20
-  max_duration: 27.0
-  min_duration: 0.0
+  min_input_len: 0.0
+  max_input_len: 27.0 # second
+  min_output_len: 0.0
+  max_output_len: 400.0
+  min_output_input_ratio: 0.05
+  max_output_input_ratio: 10.0
   specgram_type: linear
   target_sample_rate: 16000
   max_freq: None
   n_fft: None
   stride_ms: 10.0
   window_ms: 20.0
+  delta_delta: False
+  dither: 1.0
   use_dB_normalization: True
   target_dB: -20
   random_seed: 0
@@ -22,18 +28,22 @@ data:
   sortagrad: True 
   shuffle_method: batch_shuffle
   num_workers: 0
+
 model:
   num_conv_layers: 2
   num_rnn_layers: 3
   rnn_layer_size: 2048
   use_gru: False 
   share_rnn_weights: True
+
 training:
   n_epoch: 50
   lr: 1e-3
   lr_decay: 0.83
   weight_decay: 1e-06
   global_grad_clip: 5.0
+  log_interval: 100
+
 decoding:
   batch_size: 128
   error_rate_type: wer

examples/librispeech/s0/local/avg.sh

+#! /usr/bin/env bash
+
+if [ $# != 2 ];then
+    echo "usage: ${0} ckpt_dir avg_num"
+    exit -1
+fi
+
+ckpt_dir=${1}
+average_num=${2}
+decode_checkpoint=${ckpt_dir}/avg_${average_num}.pdparams
+
+python3 -u ${MAIN_ROOT}/utils/avg_model.py \
+--dst_model ${decode_checkpoint} \
+--ckpt_dir ${ckpt_dir}  \
+--num ${average_num} \
+--val_best
+            
+if [ $? -ne 0 ]; then
+    echo "Failed in avg ckpt!"
+    exit 1
+fi
+
+exit 0
\ No newline at end of file

examples/librispeech/s0/local/data.sh

 #! /usr/bin/env bash
 
+stage=-1
+stop_stage=100
+
+unit_type=char
+
+source ${MAIN_ROOT}/utils/parse_options.sh
+
 mkdir -p data
 TARGET_DIR=${MAIN_ROOT}/examples/dataset
 mkdir -p ${TARGET_DIR}
 
-# download data, generate manifests
-PYTHONPATH=.:$PYTHONPATH python3 ${TARGET_DIR}/librispeech/librispeech.py \
---manifest_prefix="data/manifest" \
---target_dir="${TARGET_DIR}/librispeech" \
---full_download="True"
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+    # download data, generate manifests
+    python3 ${TARGET_DIR}/librispeech/librispeech.py \
+    --manifest_prefix="data/manifest" \
+    --target_dir="${TARGET_DIR}/librispeech" \
+    --full_download="True"
     
-if [ $? -ne 0 ]; then
+    if [ $? -ne 0 ]; then
         echo "Prepare LibriSpeech failed. Terminated."
         exit 1
-fi
+    fi
+    
+    for set in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+        mv data/manifest.${set} data/manifest.${set}.raw
+    done
+
+    for set in train-clean-100 train-clean-360 train-other-500; do
+        cat data/manifest.${set}.raw >> data/manifest.train.raw
+    done
+
+    for set in dev-clean dev-other; do
+        cat data/manifest.${set}.raw >> data/manifest.dev.raw
+    done
 
-cat data/manifest.train-* | shuf > data/manifest.train
+    for set in test-clean test-other; do
+        cat data/manifest.${set}.raw >> data/manifest.test.raw
+    done
+fi
 
-# build vocabulary
-python3 ${MAIN_ROOT}/utils/build_vocab.py \
---count_threshold=0 \
---vocab_path="data/vocab.txt" \
---manifest_paths="data/manifest.train"
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    # build vocabulary
+    python3 ${MAIN_ROOT}/utils/build_vocab.py \
+    --unit_type ${unit_type} \
+    --count_threshold=0 \
+    --vocab_path="data/vocab.txt" \
+    --manifest_paths="data/manifest.train.raw"
     
-if [ $? -ne 0 ]; then
+    if [ $? -ne 0 ]; then
         echo "Build vocabulary failed. Terminated."
         exit 1
+    fi
 fi
 
-# compute mean and stddev for normalizer
-python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
---manifest_path="data/manifest.train" \
---num_samples=2000 \
---specgram_type="linear" \
---output_path="data/mean_std.npz"
 
-if [ $? -ne 0 ]; then
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    # compute mean and stddev for normalizer
+    num_workers=$(nproc)
+    python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
+    --manifest_path="data/manifest.train.raw" \
+    --num_samples=-1 \
+    --specgram_type="linear" \
+    --delta_delta=false \
+    --sample_rate=16000 \
+    --stride_ms=10.0 \
+    --window_ms=20.0 \
+    --use_dB_normalization=False \
+    --num_workers=${num_workers} \
+    --output_path="data/mean_std.json"
+    
+    if [ $? -ne 0 ]; then
         echo "Compute mean and stddev failed. Terminated."
         exit 1
+    fi
+fi
+
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    # format manifest with tokenids, vocab size
+    for set in train dev test dev-clean dev-other test-clean test-other; do
+    {
+        python3 ${MAIN_ROOT}/utils/format_data.py \
+        --feat_type "raw" \
+        --cmvn_path "data/mean_std.json" \
+        --unit_type ${unit_type} \
+        --vocab_path="data/vocab.txt" \
+        --manifest_path="data/manifest.${set}.raw" \
+        --output_path="data/manifest.${set}"
+
+        if [ $? -ne 0 ]; then
+            echo "Formt mnaifest.${set} failed. Terminated."
+            exit 1
+        fi
+    }&
+    done
+    wait
 fi
 
 echo "LibriSpeech Data preparation done."

examples/tiny/s0/.gitignore

+data
+exp
+*log

examples/tiny/s0/conf/deepspeech2.yaml

@@ -3,18 +3,24 @@ data:
   train_manifest: data/manifest.tiny
   dev_manifest: data/manifest.tiny
   test_manifest: data/manifest.tiny
-  mean_std_filepath: data/mean_std.npz
+  mean_std_filepath: data/mean_std.json
   vocab_filepath: data/vocab.txt 
   augmentation_config: conf/augmentation.json
   batch_size: 4
-  max_duration: 27.0
-  min_duration: 0.0
+  min_input_len: 0.0
+  max_input_len: 27.0
+  min_output_len: 0.0
+  max_output_len: 400.0
+  min_output_input_ratio: 0.05
+  max_output_input_ratio: 10.0
   specgram_type: linear
   target_sample_rate: 16000
   max_freq: None
   n_fft: None
   stride_ms: 10.0
   window_ms: 20.0
+  delta_delta: False
+  dither: 1.0
   use_dB_normalization: True
   target_dB: -20
   random_seed: 0
@@ -22,18 +28,22 @@ data:
   sortagrad: True 
   shuffle_method: batch_shuffle
   num_workers: 0
+  
 model:
   num_conv_layers: 2
   num_rnn_layers: 3
   rnn_layer_size: 2048
   use_gru: False 
   share_rnn_weights: True 
+
 training:
   n_epoch: 20
   lr: 1e-5 
   lr_decay: 1.0 
   weight_decay: 1e-06
   global_grad_clip: 5.0
+  log_interval: 1
+
 decoding:
   batch_size: 128
   error_rate_type: wer

examples/tiny/s0/local/avg.sh

+#! /usr/bin/env bash
+
+if [ $# != 2 ];then
+    echo "usage: ${0} ckpt_dir avg_num"
+    exit -1
+fi
+
+ckpt_dir=${1}
+average_num=${2}
+decode_checkpoint=${ckpt_dir}/avg_${average_num}.pdparams
+
+python3 -u ${MAIN_ROOT}/utils/avg_model.py \
+--dst_model ${decode_checkpoint} \
+--ckpt_dir ${ckpt_dir}  \
+--num ${average_num} \
+--val_best
+            
+if [ $? -ne 0 ]; then
+    echo "Failed in avg ckpt!"
+    exit 1
+fi
+
+exit 0
\ No newline at end of file

examples/tiny/s0/local/data.sh

@@ -3,10 +3,7 @@
 stage=-1
 stop_stage=100
 
-# bpemode (unigram or bpe)
-nbpe=200
-bpemode=unigram
-bpeprefix="data/bpe_${bpemode}_${nbpe}"
+unit_type=char
 
 source ${MAIN_ROOT}/utils/parse_options.sh
 
@@ -32,10 +29,8 @@ fi
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     # build vocabulary
     python3 ${MAIN_ROOT}/utils/build_vocab.py \
-    --unit_type "spm" \
-    --spm_vocab_size=${nbpe} \
-    --spm_mode ${bpemode} \
-    --spm_model_prefix ${bpeprefix} \
+    --unit_type ${unit_type} \
+    --count_threshold=0 \
     --vocab_path="data/vocab.txt" \
     --manifest_paths="data/manifest.tiny.raw"
     
@@ -51,12 +46,11 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
     --manifest_path="data/manifest.tiny.raw" \
     --num_samples=64 \
-    --specgram_type="fbank" \
-    --feat_dim=80 \
+    --specgram_type="linear" \
     --delta_delta=false \
     --sample_rate=16000 \
     --stride_ms=10.0 \
-    --window_ms=25.0 \
+    --window_ms=20.0 \
     --use_dB_normalization=False \
     --num_workers=2 \
     --output_path="data/mean_std.json"
@@ -73,8 +67,7 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     python3 ${MAIN_ROOT}/utils/format_data.py \
     --feat_type "raw" \
     --cmvn_path "data/mean_std.json" \
-    --unit_type "spm" \
-    --spm_model_prefix ${bpeprefix} \
+    --unit_type ${unit_type} \
     --vocab_path="data/vocab.txt" \
     --manifest_path="data/manifest.tiny.raw" \
     --output_path="data/manifest.tiny"

examples/tiny/s0/run.sh

@@ -2,6 +2,7 @@
 set -e
 source path.sh
 
+gpus=0
 stage=0
 stop_stage=100
 conf_path=conf/deepspeech2.yaml
@@ -18,7 +19,7 @@ fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     # train model, all `ckpt` under `exp` dir
-    CUDA_VISIBLE_DEVICES=0 ./local/train.sh ${conf_path}  ${ckpt}
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path}  ${ckpt}
 fi
 
 if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
@@ -28,10 +29,10 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
     # export ckpt avg_n
-    CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
 fi

examples/tiny/s1/local/data.sh

-../../s0/local/data.sh
\ No newline at end of file

examples/tiny/s1/local/data.sh

+#! /usr/bin/env bash
+
+stage=-1
+stop_stage=100
+
+# bpemode (unigram or bpe)
+nbpe=200
+bpemode=unigram
+bpeprefix="data/bpe_${bpemode}_${nbpe}"
+
+source ${MAIN_ROOT}/utils/parse_options.sh
+
+mkdir -p data
+TARGET_DIR=${MAIN_ROOT}/examples/dataset
+mkdir -p ${TARGET_DIR}
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+    # download data, generate manifests
+    python3 ${TARGET_DIR}/librispeech/librispeech.py \
+    --manifest_prefix="data/manifest" \
+    --target_dir="${TARGET_DIR}/librispeech" \
+    --full_download="False"
+    
+    if [ $? -ne 0 ]; then
+        echo "Prepare LibriSpeech failed. Terminated."
+        exit 1
+    fi
+    
+    head -n 64 data/manifest.dev-clean  > data/manifest.tiny.raw
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    # build vocabulary
+    python3 ${MAIN_ROOT}/utils/build_vocab.py \
+    --unit_type "spm" \
+    --spm_vocab_size=${nbpe} \
+    --spm_mode ${bpemode} \
+    --spm_model_prefix ${bpeprefix} \
+    --vocab_path="data/vocab.txt" \
+    --manifest_paths="data/manifest.tiny.raw"
+    
+    if [ $? -ne 0 ]; then
+        echo "Build vocabulary failed. Terminated."
+        exit 1
+    fi
+fi
+
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    # compute mean and stddev for normalizer
+    python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
+    --manifest_path="data/manifest.tiny.raw" \
+    --num_samples=64 \
+    --specgram_type="fbank" \
+    --feat_dim=80 \
+    --delta_delta=false \
+    --sample_rate=16000 \
+    --stride_ms=10.0 \
+    --window_ms=25.0 \
+    --use_dB_normalization=False \
+    --num_workers=2 \
+    --output_path="data/mean_std.json"
+    
+    if [ $? -ne 0 ]; then
+        echo "Compute mean and stddev failed. Terminated."
+        exit 1
+    fi
+fi
+
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    # format manifest with tokenids, vocab size
+    python3 ${MAIN_ROOT}/utils/format_data.py \
+    --feat_type "raw" \
+    --cmvn_path "data/mean_std.json" \
+    --unit_type "spm" \
+    --spm_model_prefix ${bpeprefix} \
+    --vocab_path="data/vocab.txt" \
+    --manifest_path="data/manifest.tiny.raw" \
+    --output_path="data/manifest.tiny"
+    
+    
+    if [ $? -ne 0 ]; then
+        echo "Formt mnaifest failed. Terminated."
+        exit 1
+    fi
+fi
+
+echo "LibriSpeech Data preparation done."
+exit 0