From 5bbe6e9897f7112fec0d06b08714fc26bde20ec5 Mon Sep 17 00:00:00 2001 From: tianhao zhang <15600919271@163.com> Date: Thu, 29 Sep 2022 13:41:16 +0000 Subject: [PATCH] support u2pp cli and server, optimiz code of u2pp decode, test=asr --- .../conf/application.yaml | 2 +- docs/source/released_model.md | 1 + paddlespeech/cli/asr/infer.py | 4 +- paddlespeech/resource/model_alias.py | 2 + paddlespeech/resource/pretrained_models.py | 40 +++++++++++++++++++ paddlespeech/s2t/exps/u2/bin/test_wav.py | 4 +- paddlespeech/s2t/exps/u2/model.py | 4 +- paddlespeech/s2t/models/u2/u2.py | 33 +++++++-------- .../server/conf/ws_conformer_application.yaml | 2 +- .../engine/asr/online/python/asr_engine.py | 23 +++++++++-- 10 files changed, 83 insertions(+), 32 deletions(-) diff --git a/demos/streaming_asr_server/conf/application.yaml b/demos/streaming_asr_server/conf/application.yaml index a89d312a..d446e13b 100644 --- a/demos/streaming_asr_server/conf/application.yaml +++ b/demos/streaming_asr_server/conf/application.yaml @@ -21,7 +21,7 @@ engine_list: ['asr_online'] ################################### ASR ######################################### ################### speech task: asr; engine_type: online ####################### asr_online: - model_type: 'conformer_online_wenetspeech' + model_type: 'conformer_u2pp_online_wenetspeech' am_model: # the pdmodel file of am static model [optional] am_params: # the pdiparams file of am static model [optional] lang: 'zh' diff --git a/docs/source/released_model.md b/docs/source/released_model.md index d6691812..bdac2c5b 100644 --- a/docs/source/released_model.md +++ b/docs/source/released_model.md @@ -9,6 +9,7 @@ Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | [Ds2 Online Aishell ASR0 Model](https://paddlespeech.bj.bcebos.com/s2t/aishell/asr0/asr0_deepspeech2_online_aishell_fbank161_ckpt_0.2.1.model.tar.gz) | Aishell Dataset | Char-based | 491 MB | 2 Conv + 5 LSTM layers | 0.0666 |-| 151 h | [D2 Online Aishell ASR0](../../examples/aishell/asr0) | onnx/inference/python | [Ds2 Offline Aishell ASR0 Model](https://paddlespeech.bj.bcebos.com/s2t/aishell/asr0/asr0_deepspeech2_offline_aishell_ckpt_1.0.1.model.tar.gz)| Aishell Dataset | Char-based | 1.4 GB | 2 Conv + 5 bidirectional LSTM layers| 0.0554 |-| 151 h | [Ds2 Offline Aishell ASR0](../../examples/aishell/asr0) | inference/python | [Conformer Online Wenetspeech ASR1 Model](https://paddlespeech.bj.bcebos.com/s2t/wenetspeech/asr1/asr1_chunk_conformer_wenetspeech_ckpt_1.0.0a.model.tar.gz) | WenetSpeech Dataset | Char-based | 457 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention rescoring| 0.11 (test\_net) 0.1879 (test\_meeting) |-| 10000 h |- | python | +[Conformer U2PP Online Wenetspeech ASR1 Model](https://paddlespeech.bj.bcebos.com/s2t/wenetspeech/asr1/asr1_chunk_conformer_u2pp_wenetspeech_ckpt_1.1.1.model.tar.gz) | WenetSpeech Dataset | Char-based | 476 MB | Encoder:Conformer, Decoder:BiTransformer, Decoding method: Attention rescoring| 0.047198 (aishell test\_-1) 0.059212 (aishell test\_16) |-| 10000 h |- | python | [Conformer Online Aishell ASR1 Model](https://paddlespeech.bj.bcebos.com/s2t/aishell/asr1/asr1_chunk_conformer_aishell_ckpt_0.2.0.model.tar.gz) | Aishell Dataset | Char-based | 189 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention rescoring| 0.0544 |-| 151 h | [Conformer Online Aishell ASR1](../../examples/aishell/asr1) | python | [Conformer Offline Aishell ASR1 Model](https://paddlespeech.bj.bcebos.com/s2t/aishell/asr1/asr1_conformer_aishell_ckpt_1.0.1.model.tar.gz) | Aishell Dataset | Char-based | 189 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention rescoring | 0.0460 |-| 151 h | [Conformer Offline Aishell ASR1](../../examples/aishell/asr1) | python | [Transformer Aishell ASR1 Model](https://paddlespeech.bj.bcebos.com/s2t/aishell/asr1/asr1_transformer_aishell_ckpt_0.1.1.model.tar.gz) | Aishell Dataset | Char-based | 128 MB | Encoder:Transformer, Decoder:Transformer, Decoding method: Attention rescoring | 0.0523 || 151 h | [Transformer Aishell ASR1](../../examples/aishell/asr1) | python | diff --git a/paddlespeech/cli/asr/infer.py b/paddlespeech/cli/asr/infer.py index 7296776f..4a7feaf0 100644 --- a/paddlespeech/cli/asr/infer.py +++ b/paddlespeech/cli/asr/infer.py @@ -51,7 +51,7 @@ class ASRExecutor(BaseExecutor): self.parser.add_argument( '--model', type=str, - default='conformer_wenetspeech', + default='conformer_u2pp_wenetspeech', choices=[ tag[:tag.index('-')] for tag in self.task_resource.pretrained_models.keys() @@ -465,7 +465,7 @@ class ASRExecutor(BaseExecutor): @stats_wrapper def __call__(self, audio_file: os.PathLike, - model: str='conformer_wenetspeech', + model: str='conformer_u2pp_wenetspeech', lang: str='zh', sample_rate: int=16000, config: os.PathLike=None, diff --git a/paddlespeech/resource/model_alias.py b/paddlespeech/resource/model_alias.py index 9c76dd4b..3f36f11f 100644 --- a/paddlespeech/resource/model_alias.py +++ b/paddlespeech/resource/model_alias.py @@ -25,6 +25,8 @@ model_alias = { "deepspeech2online": ["paddlespeech.s2t.models.ds2:DeepSpeech2Model"], "conformer": ["paddlespeech.s2t.models.u2:U2Model"], "conformer_online": ["paddlespeech.s2t.models.u2:U2Model"], + "conformer_u2pp": ["paddlespeech.s2t.models.u2:U2Model"], + "conformer_u2pp_online": ["paddlespeech.s2t.models.u2:U2Model"], "transformer": ["paddlespeech.s2t.models.u2:U2Model"], "wenetspeech": ["paddlespeech.s2t.models.u2:U2Model"], diff --git a/paddlespeech/resource/pretrained_models.py b/paddlespeech/resource/pretrained_models.py index f049879a..eecf2176 100644 --- a/paddlespeech/resource/pretrained_models.py +++ b/paddlespeech/resource/pretrained_models.py @@ -68,6 +68,46 @@ asr_dynamic_pretrained_models = { '', }, }, + "conformer_u2pp_wenetspeech-zh-16k": { + '1.0': { + 'url': + 'https://paddlespeech.bj.bcebos.com/s2t/wenetspeech/asr1/asr1_chunk_conformer_u2pp_wenetspeech_ckpt_1.1.1.model.tar.gz', + 'md5': + 'eae678c04ed3b3f89672052fdc0c5e10', + 'cfg_path': + 'model.yaml', + 'ckpt_path': + 'exp/chunk_conformer_u2pp/checkpoints/avg_10', + 'model': + 'exp/chunk_conformer_u2pp/checkpoints/avg_10.pdparams', + 'params': + 'exp/chunk_conformer_u2pp/checkpoints/avg_10.pdparams', + 'lm_url': + '', + 'lm_md5': + '', + }, + }, + "conformer_u2pp_online_wenetspeech-zh-16k": { + '1.0': { + 'url': + 'https://paddlespeech.bj.bcebos.com/s2t/wenetspeech/asr1/asr1_chunk_conformer_u2pp_wenetspeech_ckpt_1.1.2.model.tar.gz', + 'md5': + '925d047e9188dea7f421a718230c9ae3', + 'cfg_path': + 'model.yaml', + 'ckpt_path': + 'exp/chunk_conformer_u2pp/checkpoints/avg_10', + 'model': + 'exp/chunk_conformer_u2pp/checkpoints/avg_10.pdparams', + 'params': + 'exp/chunk_conformer_u2pp/checkpoints/avg_10.pdparams', + 'lm_url': + '', + 'lm_md5': + '', + }, + }, "conformer_online_multicn-zh-16k": { '1.0': { 'url': diff --git a/paddlespeech/s2t/exps/u2/bin/test_wav.py b/paddlespeech/s2t/exps/u2/bin/test_wav.py index 4588def0..46925fae 100644 --- a/paddlespeech/s2t/exps/u2/bin/test_wav.py +++ b/paddlespeech/s2t/exps/u2/bin/test_wav.py @@ -40,7 +40,6 @@ class U2Infer(): self.preprocess_conf = config.preprocess_config self.preprocess_args = {"train": False} self.preprocessing = Transformation(self.preprocess_conf) - self.reverse_weight = getattr(config.model_conf, 'reverse_weight', 0.0) self.text_feature = TextFeaturizer( unit_type=config.unit_type, vocab=config.vocab_filepath, @@ -89,8 +88,7 @@ class U2Infer(): 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=self.reverse_weight) + simulate_streaming=decode_config.simulate_streaming) rsl = result_transcripts[0][0] utt = Path(self.audio_file).name logger.info(f"hyp: {utt} {result_transcripts[0][0]}") diff --git a/paddlespeech/s2t/exps/u2/model.py b/paddlespeech/s2t/exps/u2/model.py index a13a6385..a6197d07 100644 --- a/paddlespeech/s2t/exps/u2/model.py +++ b/paddlespeech/s2t/exps/u2/model.py @@ -316,7 +316,6 @@ class U2Tester(U2Trainer): vocab=self.config.vocab_filepath, spm_model_prefix=self.config.spm_model_prefix) self.vocab_list = self.text_feature.vocab_list - self.reverse_weight = getattr(config.model_conf, 'reverse_weight', 0.0) def id2token(self, texts, texts_len, text_feature): """ ord() id to chr() chr """ @@ -351,8 +350,7 @@ class U2Tester(U2Trainer): 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=self.reverse_weight) + simulate_streaming=decode_config.simulate_streaming) decode_time = time.time() - start_time for utt, target, result, rec_tids in zip( diff --git a/paddlespeech/s2t/models/u2/u2.py b/paddlespeech/s2t/models/u2/u2.py index 0a3e03b7..53c3bf55 100644 --- a/paddlespeech/s2t/models/u2/u2.py +++ b/paddlespeech/s2t/models/u2/u2.py @@ -507,16 +507,14 @@ class U2BaseModel(ASRInterface, nn.Layer): num_decoding_left_chunks, simulate_streaming) return hyps[0][0] - def attention_rescoring( - self, - speech: paddle.Tensor, - speech_lengths: paddle.Tensor, - beam_size: int, - decoding_chunk_size: int=-1, - num_decoding_left_chunks: int=-1, - ctc_weight: float=0.0, - simulate_streaming: bool=False, - reverse_weight: float=0.0, ) -> List[int]: + def attention_rescoring(self, + speech: paddle.Tensor, + speech_lengths: paddle.Tensor, + beam_size: int, + decoding_chunk_size: int=-1, + num_decoding_left_chunks: int=-1, + ctc_weight: float=0.0, + simulate_streaming: bool=False) -> List[int]: """ Apply attention rescoring decoding, CTC prefix beam search is applied first to get nbest, then we resoring the nbest on attention decoder with corresponding encoder out @@ -536,7 +534,7 @@ class U2BaseModel(ASRInterface, nn.Layer): """ assert speech.shape[0] == speech_lengths.shape[0] assert decoding_chunk_size != 0 - if reverse_weight > 0.0: + if self.reverse_weight > 0.0: # decoder should be a bitransformer decoder if reverse_weight > 0.0 assert hasattr(self.decoder, 'right_decoder') device = speech.place @@ -574,7 +572,7 @@ class U2BaseModel(ASRInterface, nn.Layer): self.eos) decoder_out, r_decoder_out, _ = self.decoder( encoder_out, encoder_mask, hyps_pad, hyps_lens, r_hyps_pad, - reverse_weight) # (beam_size, max_hyps_len, vocab_size) + self.reverse_weight) # (beam_size, max_hyps_len, vocab_size) # ctc score in ln domain decoder_out = paddle.nn.functional.log_softmax(decoder_out, axis=-1) decoder_out = decoder_out.numpy() @@ -594,12 +592,13 @@ class U2BaseModel(ASRInterface, nn.Layer): score += decoder_out[i][j][w] # last decoder output token is `eos`, for laste decoder input token. score += decoder_out[i][len(hyp[0])][self.eos] - if reverse_weight > 0: + if self.reverse_weight > 0: r_score = 0.0 for j, w in enumerate(hyp[0]): r_score += r_decoder_out[i][len(hyp[0]) - j - 1][w] r_score += r_decoder_out[i][len(hyp[0])][self.eos] - score = score * (1 - reverse_weight) + r_score * reverse_weight + score = score * (1 - self.reverse_weight + ) + r_score * self.reverse_weight # add ctc score (which in ln domain) score += hyp[1] * ctc_weight if score > best_score: @@ -748,8 +747,7 @@ class U2BaseModel(ASRInterface, nn.Layer): ctc_weight: float=0.0, decoding_chunk_size: int=-1, num_decoding_left_chunks: int=-1, - simulate_streaming: bool=False, - reverse_weight: float=0.0): + simulate_streaming: bool=False): """u2 decoding. Args: @@ -821,8 +819,7 @@ class U2BaseModel(ASRInterface, nn.Layer): decoding_chunk_size=decoding_chunk_size, num_decoding_left_chunks=num_decoding_left_chunks, ctc_weight=ctc_weight, - simulate_streaming=simulate_streaming, - reverse_weight=reverse_weight) + simulate_streaming=simulate_streaming) hyps = [hyp] else: raise ValueError(f"Not support decoding method: {decoding_method}") diff --git a/paddlespeech/server/conf/ws_conformer_application.yaml b/paddlespeech/server/conf/ws_conformer_application.yaml index d72eb237..b6128118 100644 --- a/paddlespeech/server/conf/ws_conformer_application.yaml +++ b/paddlespeech/server/conf/ws_conformer_application.yaml @@ -30,7 +30,7 @@ asr_online: decode_method: num_decoding_left_chunks: -1 force_yes: True - device: # cpu or gpu:id + device: gpu # cpu or gpu:id continuous_decoding: True # enable continue decoding when endpoint detected am_predictor_conf: diff --git a/paddlespeech/server/engine/asr/online/python/asr_engine.py b/paddlespeech/server/engine/asr/online/python/asr_engine.py index 4c7c4b37..740f5270 100644 --- a/paddlespeech/server/engine/asr/online/python/asr_engine.py +++ b/paddlespeech/server/engine/asr/online/python/asr_engine.py @@ -22,6 +22,7 @@ from numpy import float32 from yacs.config import CfgNode from paddlespeech.audio.transform.transformation import Transformation +from paddlespeech.audio.utils.tensor_utils import st_reverse_pad_list from paddlespeech.cli.asr.infer import ASRExecutor from paddlespeech.cli.log import logger from paddlespeech.resource import CommonTaskResource @@ -603,24 +604,31 @@ class PaddleASRConnectionHanddler: hyps_pad = pad_sequence( hyp_list, batch_first=True, padding_value=self.model.ignore_id) + ori_hyps_pad = hyps_pad hyps_lens = paddle.to_tensor( [len(hyp[0]) for hyp in hyps], place=self.device, dtype=paddle.long) # (beam_size,) hyps_pad, _ = add_sos_eos(hyps_pad, self.model.sos, self.model.eos, self.model.ignore_id) hyps_lens = hyps_lens + 1 # Add at begining - encoder_out = self.encoder_out.repeat(beam_size, 1, 1) encoder_mask = paddle.ones( (beam_size, 1, encoder_out.shape[1]), dtype=paddle.bool) - decoder_out, _, _ = self.model.decoder( - encoder_out, encoder_mask, hyps_pad, - hyps_lens) # (beam_size, max_hyps_len, vocab_size) + r_hyps_pad = st_reverse_pad_list(ori_hyps_pad, hyps_lens - 1, + self.model.sos, self.model.eos) + decoder_out, r_decoder_out, _ = self.model.decoder( + encoder_out, encoder_mask, hyps_pad, hyps_lens, r_hyps_pad, + self.model.reverse_weight) # (beam_size, max_hyps_len, vocab_size) # ctc score in ln domain decoder_out = paddle.nn.functional.log_softmax(decoder_out, axis=-1) decoder_out = decoder_out.numpy() + # r_decoder_out will be 0.0, if reverse_weight is 0.0 or decoder is a + # conventional transformer decoder. + r_decoder_out = paddle.nn.functional.log_softmax(r_decoder_out, axis=-1) + r_decoder_out = r_decoder_out.numpy() + # Only use decoder score for rescoring best_score = -float('inf') best_index = 0 @@ -632,6 +640,13 @@ class PaddleASRConnectionHanddler: # last decoder output token is `eos`, for laste decoder input token. score += decoder_out[i][len(hyp[0])][self.model.eos] + if self.model.reverse_weight > 0: + r_score = 0.0 + for j, w in enumerate(hyp[0]): + r_score += r_decoder_out[i][len(hyp[0]) - j - 1][w] + r_score += r_decoder_out[i][len(hyp[0])][self.model.eos] + score = score * (1 - self.model.reverse_weight + ) + r_score * self.model.reverse_weight # add ctc score (which in ln domain) score += hyp[1] * self.ctc_decode_config.ctc_weight