Merge pull request #851 from Jackwaterveg/release_model

Optimize the 1xt2x code

deepspeech/frontend/featurizer/text_featurizer.py

@@ -199,7 +199,6 @@ class TextFeaturizer():
         """Load vocabulary from file."""
         vocab_list = load_dict(vocab_filepath, maskctc)
         assert vocab_list is not None
-        assert SPACE in vocab_list
 
         id2token = dict(
             [(idx, token) for (idx, token) in enumerate(vocab_list)])

examples/1xt2x/aishell/.gitignore

@@ -2,3 +2,4 @@ exp
 data
 *log
 tmp
+nohup*

examples/1xt2x/aishell/local/data.sh

@@ -1,4 +1,10 @@
 #!/bin/bash
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
 
 stage=-1
 stop_stage=100
@@ -9,17 +15,16 @@ mkdir -p data
 TARGET_DIR=${MAIN_ROOT}/examples/dataset
 mkdir -p ${TARGET_DIR}
 
-bash local/download_model.sh
+bash local/download_model.sh ${ckpt_dir}
 if [ $? -ne 0 ]; then
    exit 1
 fi
 
+cd ${ckpt_dir}
 tar xzvf aishell_model_v1.8_to_v2.x.tar.gz
-mv aishell_v1.8.pdparams exp/deepspeech2/checkpoints/
-mv README.md exp/deepspeech2/
-mv mean_std.npz data/
-mv vocab.txt data/
-rm aishell_model_v1.8_to_v2.x.tar.gz -f
+cd -
+mv ${ckpt_dir}/mean_std.npz data/
+mv ${ckpt_dir}/vocab.txt data/
 
 
 if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
@@ -39,27 +44,6 @@ if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
 fi
 
 
-if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
-    # compute mean and stddev for normalizer
-    num_workers=$(nproc)
-    python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
-    --manifest_path="data/manifest.train.raw" \
-    --specgram_type="linear" \
-    --delta_delta=false \
-    --stride_ms=10.0 \
-    --window_ms=20.0 \
-    --sample_rate=16000 \
-    --use_dB_normalization=True \
-    --num_samples=2000 \
-    --num_workers=${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
@@ -67,7 +51,7 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     {
         python3 ${MAIN_ROOT}/utils/format_data.py \
                 --feat_type "raw" \
-                --cmvn_path "data/mean_std.json" \
+                --cmvn_path "data/mean_std.npz" \
                 --unit_type "char" \
                 --vocab_path="data/vocab.txt" \
                 --manifest_path="data/manifest.${dataset}.raw" \

examples/1xt2x/aishell/local/download_model.sh

@@ -1,10 +1,17 @@
 #! /usr/bin/env bash
 
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
+
 . ${MAIN_ROOT}/utils/utility.sh
 
 URL='https://deepspeech.bj.bcebos.com/mandarin_models/aishell_model_v1.8_to_v2.x.tar.gz'
-MD5=4ade113c69ea291b8ce5ec6a03296659
-TARGET=./aishell_model_v1.8_to_v2.x.tar.gz
+MD5=87e7577d4bea737dbf3e8daab37aa808
+TARGET=${ckpt_dir}/aishell_model_v1.8_to_v2.x.tar.gz
 
 
 echo "Download Aishell model ..."
@@ -13,7 +20,6 @@ if [ $? -ne 0 ]; then
     echo "Fail to download Aishell model!"
     exit 1
 fi
-tar -zxvf $TARGET
 
 
 exit 0

examples/1xt2x/aishell/path.sh

@@ -12,5 +12,5 @@ export PYTHONPATH=${LOCAL_DEEPSPEECH2}:${PYTHONPATH}
 export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib/
 
 MODEL=deepspeech2
-export BIN_DIR=${LOCAL_DEEPSPEECH2}/deepspeech2x/bin
+export BIN_DIR=${LOCAL_DEEPSPEECH2}/src_deepspeech2x/bin
 echo "BIN_DIR "${BIN_DIR}

examples/1xt2x/aishell/run.sh

@@ -7,6 +7,7 @@ stop_stage=100
 conf_path=conf/deepspeech2.yaml
 avg_num=1
 model_type=offline
+gpus=2
 
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 
@@ -17,11 +18,11 @@ echo "checkpoint name ${ckpt}"
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     # prepare data
     mkdir -p exp/${ckpt}/checkpoints
-    bash ./local/data.sh || exit -1
+    bash ./local/data.sh exp/${ckpt}/checkpoints || exit -1
 fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=1 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
 fi
 

examples/1xt2x/baidu_en8k/.gitignore

@@ -2,3 +2,4 @@ exp
 data
 *log
 tmp
+nohup*

examples/1xt2x/baidu_en8k/local/data.sh

@@ -1,7 +1,14 @@
 #!/bin/bash
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
 
 stage=-1
 stop_stage=100
+unit_type=char
 
 source ${MAIN_ROOT}/utils/parse_options.sh
 
@@ -10,17 +17,17 @@ TARGET_DIR=${MAIN_ROOT}/examples/dataset
 mkdir -p ${TARGET_DIR}
 
 
-bash local/download_model.sh
+bash local/download_model.sh ${ckpt_dir}
 if [ $? -ne 0 ]; then
    exit 1
 fi
 
+cd ${ckpt_dir}
 tar xzvf baidu_en8k_v1.8_to_v2.x.tar.gz
-mv baidu_en8k_v1.8.pdparams exp/deepspeech2/checkpoints/
-mv README.md exp/deepspeech2/
-mv mean_std.npz data/
-mv vocab.txt data/
-rm baidu_en8k_v1.8_to_v2.x.tar.gz -f
+cd -
+mv ${ckpt_dir}/mean_std.npz data/
+mv ${ckpt_dir}/vocab.txt data/
+
 
 if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
     # download data, generate manifests
@@ -53,35 +60,13 @@ if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
 fi
 
 
-if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
-    # compute mean and stddev for normalizer
-    num_workers=$(nproc)
-    python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
-    --manifest_path="data/manifest.train.raw" \
-    --num_samples=2000 \
-    --specgram_type="linear" \
-    --delta_delta=false \
-    --sample_rate=16000 \
-    --stride_ms=10.0 \
-    --window_ms=20.0 \
-    --use_dB_normalization=True \
-    --num_workers=${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" \
+        --cmvn_path "data/mean_std.npz" \
         --unit_type ${unit_type} \
         --vocab_path="data/vocab.txt" \
         --manifest_path="data/manifest.${set}.raw" \

examples/1xt2x/baidu_en8k/local/download_model.sh

@@ -1,10 +1,17 @@
 #! /usr/bin/env bash
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
+
 
 . ${MAIN_ROOT}/utils/utility.sh
 
 URL='https://deepspeech.bj.bcebos.com/eng_models/baidu_en8k_v1.8_to_v2.x.tar.gz'
-MD5=fdabeb6c96963ac85d9188f0275c6a1b
-TARGET=./baidu_en8k_v1.8_to_v2.x.tar.gz
+MD5=c1676be8505cee436e6f312823e9008c
+TARGET=${ckpt_dir}/baidu_en8k_v1.8_to_v2.x.tar.gz
 
 
 echo "Download BaiduEn8k model ..."
@@ -13,7 +20,6 @@ if [ $? -ne 0 ]; then
     echo "Fail to download BaiduEn8k model!"
     exit 1
 fi
-tar -zxvf $TARGET
 
 
 exit 0

examples/1xt2x/baidu_en8k/path.sh

@@ -12,5 +12,5 @@ export PYTHONPATH=${LOCAL_DEEPSPEECH2}:${PYTHONPATH}
 export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib/
 
 MODEL=deepspeech2
-export BIN_DIR=${LOCAL_DEEPSPEECH2}/deepspeech2x/bin
+export BIN_DIR=${LOCAL_DEEPSPEECH2}/src_deepspeech2x/bin
 echo "BIN_DIR "${BIN_DIR}

examples/1xt2x/baidu_en8k/run.sh

@@ -7,6 +7,7 @@ stop_stage=100
 conf_path=conf/deepspeech2.yaml
 avg_num=1
 model_type=offline
+gpus=0
 
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 
@@ -17,11 +18,11 @@ echo "checkpoint name ${ckpt}"
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     # prepare data
     mkdir -p exp/${ckpt}/checkpoints
-    bash ./local/data.sh || exit -1
+    bash ./local/data.sh exp/${ckpt}/checkpoints || exit -1
 fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
 fi
 

examples/1xt2x/librispeech/.gitignore

@@ -2,3 +2,4 @@ exp
 data
 *log
 tmp
+nohup*

examples/1xt2x/librispeech/local/data.sh

@@ -1,7 +1,15 @@
 #!/bin/bash
 
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
+
 stage=-1
 stop_stage=100
+unit_type=char
 
 source ${MAIN_ROOT}/utils/parse_options.sh
 
@@ -9,18 +17,16 @@ mkdir -p data
 TARGET_DIR=${MAIN_ROOT}/examples/dataset
 mkdir -p ${TARGET_DIR}
 
-
-bash local/download_model.sh
+bash local/download_model.sh ${ckpt_dir}
 if [ $? -ne 0 ]; then
    exit 1
 fi
 
+cd ${ckpt_dir}
 tar xzvf librispeech_v1.8_to_v2.x.tar.gz
-mv librispeech_v1.8.pdparams exp/deepspeech2/checkpoints/
-mv README.md exp/deepspeech2/
-mv mean_std.npz data/
-mv vocab.txt data/
-rm librispeech_v1.8_to_v2.x.tar.gz -f
+cd -
+mv ${ckpt_dir}/mean_std.npz data/
+mv ${ckpt_dir}/vocab.txt data/
 
 if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
     # download data, generate manifests
@@ -52,36 +58,13 @@ if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
     done
 fi
 
-
-if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
-    # compute mean and stddev for normalizer
-    num_workers=$(nproc)
-    python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
-    --manifest_path="data/manifest.train.raw" \
-    --num_samples=2000 \
-    --specgram_type="linear" \
-    --delta_delta=false \
-    --sample_rate=16000 \
-    --stride_ms=10.0 \
-    --window_ms=20.0 \
-    --use_dB_normalization=True \
-    --num_workers=${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" \
+        --cmvn_path "data/mean_std.npz" \
         --unit_type ${unit_type} \
         --vocab_path="data/vocab.txt" \
         --manifest_path="data/manifest.${set}.raw" \

examples/1xt2x/librispeech/local/download_model.sh

@@ -1,10 +1,17 @@
 #! /usr/bin/env bash
 
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
+
 . ${MAIN_ROOT}/utils/utility.sh
 
 URL='https://deepspeech.bj.bcebos.com/eng_models/librispeech_v1.8_to_v2.x.tar.gz'
-MD5=7b0f582fe2f5a840b840e7ee52246bc5
-TARGET=./librispeech_v1.8_to_v2.x.tar.gz
+MD5=a06d9aadb560ea113984dc98d67232c8
+TARGET=${ckpt_dir}/librispeech_v1.8_to_v2.x.tar.gz
 
 
 echo "Download LibriSpeech model ..."
@@ -13,7 +20,6 @@ if [ $? -ne 0 ]; then
     echo "Fail to download LibriSpeech model!"
     exit 1
 fi
-tar -zxvf $TARGET
 
 
 exit 0

examples/1xt2x/librispeech/path.sh

@@ -12,5 +12,4 @@ export PYTHONPATH=${LOCAL_DEEPSPEECH2}:${PYTHONPATH}
 export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib/
 
 MODEL=deepspeech2
-export BIN_DIR=${LOCAL_DEEPSPEECH2}/deepspeech2x/bin
-echo "BIN_DIR "${BIN_DIR}
+export BIN_DIR=${LOCAL_DEEPSPEECH2}/src_deepspeech2x/bin

examples/1xt2x/librispeech/run.sh

@@ -7,6 +7,7 @@ stop_stage=100
 conf_path=conf/deepspeech2.yaml
 avg_num=1
 model_type=offline
+gpus=1
 
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 
@@ -17,10 +18,10 @@ echo "checkpoint name ${ckpt}"
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     # prepare data
     mkdir -p exp/${ckpt}/checkpoints
-    bash ./local/data.sh || exit -1
+    bash ./local/data.sh exp/${ckpt}/checkpoints || exit -1
 fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
 fi

examples/1xt2x/src_deepspeech2x/bin/test.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Evaluation for DeepSpeech2 model."""
-from deepspeech2x.model import DeepSpeech2Tester as Tester
+from src_deepspeech2x.test_model import DeepSpeech2Tester as Tester
 
 from deepspeech.exps.deepspeech2.config import get_cfg_defaults
 from deepspeech.training.cli import default_argument_parser

examples/1xt2x/src_deepspeech2x/models/ds2/deepspeech2.py

@@ -15,8 +15,8 @@
 from typing import Optional
 
 import paddle
-from deepspeech2x.models.ds2.rnn import RNNStack
 from paddle import nn
+from src_deepspeech2x.models.ds2.rnn import RNNStack
 from yacs.config import CfgNode
 
 from deepspeech.models.ds2.conv import ConvStack

examples/1xt2x/src_deepspeech2x/test_model.py

@@ -20,12 +20,13 @@ from typing import Optional
 
 import numpy as np
 import paddle
-from deepspeech2x.models.ds2 import DeepSpeech2InferModel
-from deepspeech2x.models.ds2 import DeepSpeech2Model
 from paddle import distributed as dist
 from paddle.io import DataLoader
+from src_deepspeech2x.models.ds2 import DeepSpeech2InferModel
+from src_deepspeech2x.models.ds2 import DeepSpeech2Model
 from yacs.config import CfgNode
 
+from deepspeech.frontend.featurizer.text_featurizer import TextFeaturizer
 from deepspeech.io.collator import SpeechCollator
 from deepspeech.io.dataset import ManifestDataset
 from deepspeech.io.sampler import SortagradBatchSampler
@@ -38,7 +39,6 @@ from deepspeech.utils import error_rate
 from deepspeech.utils import layer_tools
 from deepspeech.utils import mp_tools
 from deepspeech.utils.log import Log
-#from deepspeech.utils.log import Autolog
 
 logger = Log(__name__).getlog()
 
@@ -272,6 +272,9 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
         return default
 
     def __init__(self, config, args):
+
+        self._text_featurizer = TextFeaturizer(
+            unit_type=config.collator.unit_type, vocab_filepath=None)
         super().__init__(config, args)
 
     def ordid2token(self, texts, texts_len):
@@ -296,9 +299,6 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
         error_rate_func = error_rate.cer if cfg.error_rate_type == 'cer' else error_rate.wer
 
         vocab_list = self.test_loader.collate_fn.vocab_list
-        if "" in vocab_list:
-            space_id = vocab_list.index("")
-            vocab_list[space_id] = " "
 
         target_transcripts = self.ordid2token(texts, texts_len)
 
@@ -337,6 +337,10 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
             cutoff_prob=cfg.cutoff_prob,
             cutoff_top_n=cfg.cutoff_top_n,
             num_processes=cfg.num_proc_bsearch)
+        result_transcripts = [
+            self._text_featurizer.detokenize(item)
+            for item in result_transcripts
+        ]
         return result_transcripts
 
     @mp_tools.rank_zero_only
@@ -367,8 +371,6 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
             error_rate_type, num_ins, num_ins, errors_sum / len_refs)
         logger.info(msg)
 
-    # self.autolog.report()
-
     def run_test(self):
         self.resume_or_scratch()
         try:

utils/format_data.py

@@ -53,7 +53,8 @@ def main():
     fout = open(args.output_path, 'w', encoding='utf-8')
 
     # get feat dim
-    mean, std = load_cmvn(args.cmvn_path, filetype='json')
+    filetype = args.cmvn_path.split(".")[-1]
+    mean, istd = load_cmvn(args.cmvn_path, filetype=filetype)
     feat_dim = mean.shape[0] #(D)
     print(f"Feature dim: {feat_dim}")