Merge branch 'develop' into u2pp_export

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'

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 |

examples/wenetspeech/asr1/conf/tuning/chunk_decode.yaml

 beam_size: 10
-decode_batch_size: 128
-error_rate_type: cer 
 decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
 ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+reverse_weight: 0.3 # reverse weight for attention rescoring decode mode.
 decoding_chunk_size: 16 # decoding chunk size. Defaults to -1.
     # <0: for decoding, use full chunk.
     # >0: for decoding, use fixed chunk size as set.
     # 0: used for training, it's prohibited here. 
 num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
 simulate_streaming: True  # simulate streaming inference. Defaults to False.
+decode_batch_size: 128
+error_rate_type: cer 

examples/wenetspeech/asr1/conf/tuning/decode.yaml

-decode_batch_size: 128
-error_rate_type: cer 
-decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
 beam_size: 10
+decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
 ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+reverse_weight: 0.3 # reverse weight for attention rescoring decode mode.
 decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
     # <0: for decoding, use full chunk.
     # >0: for decoding, use fixed chunk size as set.
     # 0: used for training, it's prohibited here. 
 num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-simulate_streaming: False  # simulate streaming inference. Defaults to False.
\ No newline at end of file
+simulate_streaming: False  # simulate streaming inference. Defaults to False.
+decode_batch_size: 128
+error_rate_type: cer 

paddlespeech/cli/asr/infer.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import argparse
+import io
 import os
 import sys
 import time
@@ -51,7 +52,7 @@ 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()
@@ -229,6 +230,8 @@ class ASRExecutor(BaseExecutor):
         audio_file = input
         if isinstance(audio_file, (str, os.PathLike)):
             logger.debug("Preprocess audio_file:" + audio_file)
+        elif isinstance(audio_file, io.BytesIO):
+            audio_file.seek(0)
 
         # Get the object for feature extraction
         if "deepspeech2" in model_type or "conformer" in model_type or "transformer" in model_type:
@@ -352,6 +355,8 @@ class ASRExecutor(BaseExecutor):
             if not os.path.isfile(audio_file):
                 logger.error("Please input the right audio file path")
                 return False
+        elif isinstance(audio_file, io.BytesIO):
+            audio_file.seek(0)
 
         logger.debug("checking the audio file format......")
         try:
@@ -465,7 +470,7 @@ 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,

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"],
 

paddlespeech/resource/pretrained_models.py

@@ -68,6 +68,46 @@ asr_dynamic_pretrained_models = {
             '',
         },
     },
+    "conformer_u2pp_wenetspeech-zh-16k": {
+        '1.1': {
+            '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": {
+        '1.1': {
+            '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':

paddlespeech/s2t/exps/u2/bin/quant.py

@@ -62,7 +62,6 @@ class U2Infer():
         params_path = self.args.checkpoint_path + ".pdparams"
         model_dict = paddle.load(params_path)
         self.model.set_state_dict(model_dict)
-        logger.info(f"model_dict: {model_dict.keys()}")
 
     def run(self):
         check(args.audio_file)
@@ -91,22 +90,22 @@ 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
+                reverse_weight=decode_config.reverse_weight)
             rsl = result_transcripts[0][0]
             utt = Path(self.audio_file).name
-            logger.info(f"hyp: {utt} {result_transcripts[0][0]}")
+            logger.info(f"hyp: {utt} {rsl}")
             # print(self.model)
             # print(self.model.forward_encoder_chunk)
-            # return rsl
             
-            logger.info("-------------start export ----------------------")
+            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}"
+                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 ############
@@ -146,7 +145,6 @@ class U2Infer():
                 self.model.ctc_activation, input_spec=input_spec)
 
             ######################### self.model.forward_attention_decoder ########################
-            reverse_weight = 0.3
             input_spec = [
                 # hyps, (B, U)
                 paddle.static.InputSpec(shape=[None, None], dtype='int64'),

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,
@@ -90,7 +89,7 @@ class U2Infer():
                 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)
+                reverse_weight=decode_config.reverse_weight)
             rsl = result_transcripts[0][0]
             utt = Path(self.audio_file).name
             logger.info(f"hyp: {utt} {result_transcripts[0][0]}")

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 +350,8 @@ 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
+            reverse_weight=decode_config.reverse_weight)
         decode_time = time.time() - start_time
 
         for utt, target, result, rec_tids in zip(

paddlespeech/s2t/models/u2/u2.py

@@ -131,7 +131,7 @@ class U2BaseModel(ASRInterface, nn.Layer):
         if self.ctc_weight != 1.0:
             start = time.time()
             loss_att, acc_att = self._calc_att_loss(encoder_out, encoder_mask,
-                                                    text, text_lengths)
+                                                    text, text_lengths, self.reverse_weight)
             decoder_time = time.time() - start
             #logger.debug(f"decoder time: {decoder_time}")
 
@@ -157,7 +157,8 @@ class U2BaseModel(ASRInterface, nn.Layer):
             encoder_out: paddle.Tensor,
             encoder_mask: paddle.Tensor,
             ys_pad: paddle.Tensor,
-            ys_pad_lens: paddle.Tensor, ) -> Tuple[paddle.Tensor, float]:
+            ys_pad_lens: paddle.Tensor, 
+            reverse_weight: float) -> Tuple[paddle.Tensor, float]:
         """Calc attention loss.
 
         Args:
@@ -165,6 +166,7 @@ class U2BaseModel(ASRInterface, nn.Layer):
             encoder_mask (paddle.Tensor): [B, 1, Tmax]
             ys_pad (paddle.Tensor): [B, Umax]
             ys_pad_lens (paddle.Tensor): [B]
+            reverse_weight (float): reverse decoder weight.
 
         Returns:
             Tuple[paddle.Tensor, float]: attention_loss, accuracy rate
@@ -179,15 +181,15 @@ class U2BaseModel(ASRInterface, nn.Layer):
         # 1. Forward decoder
         decoder_out, r_decoder_out, _ = self.decoder(
             encoder_out, encoder_mask, ys_in_pad, ys_in_lens, r_ys_in_pad,
-            self.reverse_weight)
+            reverse_weight)
 
         # 2. Compute attention loss
         loss_att = self.criterion_att(decoder_out, ys_out_pad)
         r_loss_att = paddle.to_tensor(0.0)
-        if self.reverse_weight > 0.0:
+        if reverse_weight > 0.0:
             r_loss_att = self.criterion_att(r_decoder_out, r_ys_out_pad)
-        loss_att = loss_att * (1 - self.reverse_weight
-                               ) + r_loss_att * self.reverse_weight
+        loss_att = loss_att * (1 - reverse_weight
+                               ) + r_loss_att * reverse_weight
         acc_att = th_accuracy(
             decoder_out.view(-1, self.vocab_size),
             ys_out_pad,
@@ -501,16 +503,15 @@ 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,
+                            reverse_weight: float=0.0) -> 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
@@ -525,6 +526,7 @@ class U2BaseModel(ASRInterface, nn.Layer):
                 0: used for training, it's prohibited here
             simulate_streaming (bool): whether do encoder forward in a
                 streaming fashion
+            reverse_weight (float): reverse deocder weight.
         Returns:
             List[int]: Attention rescoring result
         """
@@ -554,14 +556,13 @@ class U2BaseModel(ASRInterface, nn.Layer):
                 hyp_content, place=device, dtype=paddle.long)
             hyp_list.append(hyp_content)
         hyps_pad = pad_sequence(hyp_list, True, self.ignore_id)
-        ori_hyps_pad = hyps_pad
         hyps_lens = paddle.to_tensor(
             [len(hyp[0]) for hyp in hyps], place=device,
             dtype=paddle.long)  # (beam_size,)
         hyps_pad, _ = add_sos_eos(hyps_pad, self.sos, self.eos, self.ignore_id)
+        hyps_lens = hyps_lens + 1  # Add <sos> at begining
         logger.debug(
             f"hyps pad: {hyps_pad} {self.sos} {self.eos} {self.ignore_id}")
-        hyps_lens = hyps_lens + 1  # Add <sos> at begining
 
         # ctc score in ln domain
         # (beam_size, max_hyps_len, vocab_size)
@@ -598,8 +599,8 @@ class U2BaseModel(ASRInterface, nn.Layer):
                     f"hyp {i} len {len(hyp[0])} r2l score: {r_score} ctc_score: {hyp[1]} reverse_weight: {reverse_weight}"
                 )
 
-                score = score * (1 - reverse_weight) + r_score * reverse_weight
-
+                score = score * (1 - reverse_weight
+                                 ) + r_score * reverse_weight
             # add ctc score (which in ln domain)
             score += hyp[1] * ctc_weight
             if score > best_score:
@@ -769,6 +770,7 @@ class U2BaseModel(ASRInterface, nn.Layer):
             num_decoding_left_chunks (int, optional):
                     number of left chunks for decoding. Defaults to -1.
             simulate_streaming (bool, optional): simulate streaming inference. Defaults to False.
+            reverse_weight (float, optional): reverse decoder weight, used by `attention_rescoring`.
 
         Raises:
             ValueError: when not support decoding_method.

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: cpu # cpu or gpu:id
     continuous_decoding: True # enable continue decoding when endpoint detected
 
     am_predictor_conf:

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
@@ -602,6 +603,7 @@ 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,)
@@ -609,16 +611,15 @@ class PaddleASRConnectionHanddler:
                                   self.model.ignore_id)
         hyps_lens = hyps_lens + 1  # Add <sos> 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)
         # ctc score in ln domain
-        decoder_out = paddle.nn.functional.log_softmax(decoder_out, axis=-1)
+        # (beam_size, max_hyps_len, vocab_size)
+        decoder_out, r_decoder_out = self.model.forward_attention_decoder(
+            hyps_pad, hyps_lens, self.encoder_out, self.model.reverse_weight)
+
         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 = r_decoder_out.numpy()
 
         # Only use decoder score for rescoring
         best_score = -float('inf')
@@ -631,6 +632,13 @@ 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
 

paddlespeech/server/engine/text/python/text_engine.py

@@ -107,11 +107,14 @@ class PaddleTextConnectionHandler:
             assert len(tokens) == len(labels)
 
             text = ''
+            is_fast_model = 'fast' in self.text_engine.config.model_type
             for t, l in zip(tokens, labels):
                 text += t
                 if l != 0:  # Non punc.
-                    text += self._punc_list[l]
-
+                    if is_fast_model:
+                        text += self._punc_list[l - 1]
+                    else:
+                        text += self._punc_list[l]
             return text
         else:
             raise NotImplementedError
@@ -160,14 +163,23 @@ class TextEngine(BaseEngine):
             return False
 
         self.executor = TextServerExecutor()
-        self.executor._init_from_path(
-            task=config.task,
-            model_type=config.model_type,
-            lang=config.lang,
-            cfg_path=config.cfg_path,
-            ckpt_path=config.ckpt_path,
-            vocab_file=config.vocab_file)
-
+        if 'fast' in config.model_type:
+            self.executor._init_from_path_new(
+                task=config.task,
+                model_type=config.model_type,
+                lang=config.lang,
+                cfg_path=config.cfg_path,
+                ckpt_path=config.ckpt_path,
+                vocab_file=config.vocab_file)
+        else:
+            self.executor._init_from_path(
+                task=config.task,
+                model_type=config.model_type,
+                lang=config.lang,
+                cfg_path=config.cfg_path,
+                ckpt_path=config.ckpt_path,
+                vocab_file=config.vocab_file)
+        logger.info("Using model: %s." % (config.model_type))
         logger.info("Initialize Text server engine successfully on device: %s."
                     % (self.device))
         return True