From c83c9800cc1d2a71328b0f930ab83acf46d52927 Mon Sep 17 00:00:00 2001 From: Hui Zhang Date: Fri, 21 Oct 2022 09:51:35 +0000 Subject: [PATCH] quant with wav scp --- examples/wenetspeech/asr1/local/quant.sh | 11 +- paddlespeech/s2t/exps/u2/bin/quant.py | 245 +++++++++++------------ 2 files changed, 121 insertions(+), 135 deletions(-) diff --git a/examples/wenetspeech/asr1/local/quant.sh b/examples/wenetspeech/asr1/local/quant.sh index 9dfea9045..6a2a4c72b 100755 --- a/examples/wenetspeech/asr1/local/quant.sh +++ b/examples/wenetspeech/asr1/local/quant.sh @@ -1,7 +1,7 @@ #!/bin/bash if [ $# != 4 ];then - echo "usage: ${0} config_path decode_config_path ckpt_path_prefix audio_file" + echo "usage: ${0} config_path decode_config_path ckpt_path_prefix audio_scp" exit -1 fi @@ -11,16 +11,15 @@ echo "using $ngpu gpus..." config_path=$1 decode_config_path=$2 ckpt_prefix=$3 -audio_file=$4 +audio_scp=$4 mkdir -p data -wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/ if [ $? -ne 0 ]; then exit 1 fi -if [ ! -f ${audio_file} ]; then - echo "Plase input the right audio_file path" +if [ ! -f ${audio_scp} ]; then + echo "Plase input the right audio_scp path" exit 1 fi @@ -49,7 +48,7 @@ for type in attention_rescoring; do --checkpoint_path ${ckpt_prefix} \ --opts decode.decoding_method ${type} \ --opts decode.decode_batch_size ${batch_size} \ - --audio_file ${audio_file} + --audio_scp ${audio_scp} if [ $? -ne 0 ]; then echo "Failed in evaluation!" diff --git a/paddlespeech/s2t/exps/u2/bin/quant.py b/paddlespeech/s2t/exps/u2/bin/quant.py index c38134c57..a76bb52f0 100644 --- a/paddlespeech/s2t/exps/u2/bin/quant.py +++ b/paddlespeech/s2t/exps/u2/bin/quant.py @@ -20,6 +20,7 @@ import paddle import soundfile from paddleslim import PTQ from yacs.config import CfgNode +from kaldiio import ReadHelper from paddlespeech.audio.transform.transformation import Transformation from paddlespeech.s2t.frontend.featurizer.text_featurizer import TextFeaturizer @@ -34,7 +35,7 @@ class U2Infer(): def __init__(self, config, args): self.args = args self.config = config - self.audio_file = args.audio_file + self.audio_scp = args.audio_scp self.preprocess_conf = config.preprocess_config self.preprocess_args = {"train": False} @@ -63,133 +64,117 @@ class U2Infer(): self.model.set_state_dict(model_dict) def run(self): - check(args.audio_file) - - with paddle.no_grad(): - # read - audio, sample_rate = soundfile.read( - self.audio_file, dtype="int16", always_2d=True) - audio = audio[:, 0] - logger.info(f"audio shape: {audio.shape}") - - # fbank - feat = self.preprocessing(audio, **self.preprocess_args) - logger.info(f"feat shape: {feat.shape}") - - ilen = paddle.to_tensor(feat.shape[0]) - xs = paddle.to_tensor(feat, dtype='float32').unsqueeze(0) - decode_config = self.config.decode - logger.info(f"decode cfg: {decode_config}") - reverse_weight = getattr(decode_config, 'reverse_weight', 0.0) - result_transcripts = self.model.decode( - xs, - ilen, - text_feature=self.text_feature, - decoding_method=decode_config.decoding_method, - beam_size=decode_config.beam_size, - ctc_weight=decode_config.ctc_weight, - decoding_chunk_size=decode_config.decoding_chunk_size, - num_decoding_left_chunks=decode_config.num_decoding_left_chunks, - simulate_streaming=decode_config.simulate_streaming, - reverse_weight=reverse_weight) - rsl = result_transcripts[0][0] - utt = Path(self.audio_file).name - logger.info(f"hyp: {utt} {rsl}") - # print(self.model) - # print(self.model.forward_encoder_chunk) - - logger.info("-------------start quant ----------------------") - batch_size = 1 - feat_dim = 80 - model_size = 512 - num_left_chunks = -1 - reverse_weight = 0.3 - logger.info( - f"U2 Export Model Params: batch_size {batch_size}, feat_dim {feat_dim}, model_size {model_size}, num_left_chunks {num_left_chunks}, reverse_weight {reverse_weight}" - ) - - # ######################## self.model.forward_encoder_chunk ############ - # input_spec = [ - # # (T,), int16 - # paddle.static.InputSpec(shape=[None], dtype='int16'), - # ] - # self.model.forward_feature = paddle.jit.to_static( - # self.model.forward_feature, input_spec=input_spec) - - ######################### self.model.forward_encoder_chunk ############ - input_spec = [ - # xs, (B, T, D) - paddle.static.InputSpec( - shape=[batch_size, None, feat_dim], dtype='float32'), - # offset, int, but need be tensor - paddle.static.InputSpec(shape=[1], dtype='int32'), - # required_cache_size, int - num_left_chunks, - # att_cache - paddle.static.InputSpec( - shape=[None, None, None, None], dtype='float32'), - # cnn_cache - paddle.static.InputSpec( - shape=[None, None, None, None], dtype='float32') - ] - self.model.forward_encoder_chunk = paddle.jit.to_static( - self.model.forward_encoder_chunk, input_spec=input_spec) - - ######################### self.model.ctc_activation ######################## - input_spec = [ - # encoder_out, (B,T,D) - paddle.static.InputSpec( - shape=[batch_size, None, model_size], dtype='float32') - ] - self.model.ctc_activation = paddle.jit.to_static( - self.model.ctc_activation, input_spec=input_spec) - - ######################### self.model.forward_attention_decoder ######################## - input_spec = [ - # hyps, (B, U) - paddle.static.InputSpec(shape=[None, None], dtype='int64'), - # hyps_lens, (B,) - paddle.static.InputSpec(shape=[None], dtype='int64'), - # encoder_out, (B,T,D) - paddle.static.InputSpec( - shape=[batch_size, None, model_size], dtype='float32'), - reverse_weight - ] - self.model.forward_attention_decoder = paddle.jit.to_static( - self.model.forward_attention_decoder, input_spec=input_spec) - ################################################################################ - - # jit save - logger.info(f"export save: {self.args.export_path}") - config = { - 'is_static': True, - 'combine_params': True, - 'skip_forward': True - } - self.ptq.save_quantized_model(self.model, self.args.export_path) - # paddle.jit.save( - # self.model, - # self.args.export_path, - # combine_params=True, - # skip_forward=True) - - -def check(audio_file): - if not os.path.isfile(audio_file): - print("Please input the right audio file path") - sys.exit(-1) - - logger.info("checking the audio file format......") - try: - sig, sample_rate = soundfile.read(audio_file) - except Exception as e: - logger.error(str(e)) - logger.error( - "can not open the wav file, please check the audio file format") - sys.exit(-1) - logger.info("The sample rate is %d" % sample_rate) - assert (sample_rate == 16000) - logger.info("The audio file format is right") + cnt = 0 + with ReadHelper(f"scp:{self.audio_scp}") as reader: + for key, (rate, audio) in reader: + assert rate == 16000 + cnt += 1 + if cnt > args.num_utts: + break + + with paddle.no_grad(): + logger.info(f"audio shape: {audio.shape}") + + # fbank + feat = self.preprocessing(audio, **self.preprocess_args) + logger.info(f"feat shape: {feat.shape}") + + ilen = paddle.to_tensor(feat.shape[0]) + xs = paddle.to_tensor(feat, dtype='float32').unsqueeze(0) + decode_config = self.config.decode + logger.info(f"decode cfg: {decode_config}") + result_transcripts = self.model.decode( + xs, + ilen, + text_feature=self.text_feature, + decoding_method=decode_config.decoding_method, + beam_size=decode_config.beam_size, + ctc_weight=decode_config.ctc_weight, + decoding_chunk_size=decode_config.decoding_chunk_size, + num_decoding_left_chunks=decode_config.num_decoding_left_chunks, + simulate_streaming=decode_config.simulate_streaming, + reverse_weight=decode_config.reverse_weight) + rsl = result_transcripts[0][0] + utt = key + logger.info(f"hyp: {utt} {rsl}") + # print(self.model) + # print(self.model.forward_encoder_chunk) + + + logger.info("-------------start quant ----------------------") + batch_size = 1 + feat_dim = 80 + model_size = 512 + num_left_chunks = -1 + reverse_weight = 0.3 + logger.info( + f"U2 Export Model Params: batch_size {batch_size}, feat_dim {feat_dim}, model_size {model_size}, num_left_chunks {num_left_chunks}, reverse_weight {reverse_weight}" + ) + + # ######################## self.model.forward_encoder_chunk ############ + # input_spec = [ + # # (T,), int16 + # paddle.static.InputSpec(shape=[None], dtype='int16'), + # ] + # self.model.forward_feature = paddle.jit.to_static( + # self.model.forward_feature, input_spec=input_spec) + + ######################### self.model.forward_encoder_chunk ############ + input_spec = [ + # xs, (B, T, D) + paddle.static.InputSpec( + shape=[batch_size, None, feat_dim], dtype='float32'), + # offset, int, but need be tensor + paddle.static.InputSpec(shape=[1], dtype='int32'), + # required_cache_size, int + num_left_chunks, + # att_cache + paddle.static.InputSpec( + shape=[None, None, None, None], dtype='float32'), + # cnn_cache + paddle.static.InputSpec( + shape=[None, None, None, None], dtype='float32') + ] + self.model.forward_encoder_chunk = paddle.jit.to_static( + self.model.forward_encoder_chunk, input_spec=input_spec) + + ######################### self.model.ctc_activation ######################## + input_spec = [ + # encoder_out, (B,T,D) + paddle.static.InputSpec( + shape=[batch_size, None, model_size], dtype='float32') + ] + self.model.ctc_activation = paddle.jit.to_static( + self.model.ctc_activation, input_spec=input_spec) + + ######################### self.model.forward_attention_decoder ######################## + input_spec = [ + # hyps, (B, U) + paddle.static.InputSpec(shape=[None, None], dtype='int64'), + # hyps_lens, (B,) + paddle.static.InputSpec(shape=[None], dtype='int64'), + # encoder_out, (B,T,D) + paddle.static.InputSpec( + shape=[batch_size, None, model_size], dtype='float32'), + reverse_weight + ] + self.model.forward_attention_decoder = paddle.jit.to_static( + self.model.forward_attention_decoder, input_spec=input_spec) + ################################################################################ + + # jit save + logger.info(f"export save: {self.args.export_path}") + config = { + 'is_static': True, + 'combine_params': True, + 'skip_forward': True + } + self.ptq.save_quantized_model(self.model, self.args.export_path) + # paddle.jit.save( + # self.model, + # self.args.export_path, + # combine_params=True, + # skip_forward=True) def main(config, args): @@ -202,7 +187,9 @@ if __name__ == "__main__": parser.add_argument( "--result_file", type=str, help="path of save the asr result") parser.add_argument( - "--audio_file", type=str, help="path of the input audio file") + "--audio_scp", type=str, help="path of the input audio file") + parser.add_argument( + "--num_utts", type=int, default=200, help="num utts for quant calibrition.") parser.add_argument( "--export_path", type=str,