Merge branch 'develop' into rsl

deepspeech/exps/deepspeech2/bin/deploy/runtime.py

@@ -18,8 +18,10 @@ import numpy as np
 import paddle
 from paddle.inference import Config
 from paddle.inference import create_predictor
+from paddle.io import DataLoader
 
 from deepspeech.exps.deepspeech2.config import get_cfg_defaults
+from deepspeech.io.collator import SpeechCollator
 from deepspeech.io.dataset import ManifestDataset
 from deepspeech.models.deepspeech2 import DeepSpeech2Model
 from deepspeech.training.cli import default_argument_parser
@@ -78,26 +80,31 @@ def inference(config, args):
 def start_server(config, args):
     """Start the ASR server"""
     config.defrost()
-    config.data.manfiest = config.data.test_manifest
-    config.data.augmentation_config = ""
-    config.data.keep_transcription_text = True
+    config.data.manifest = config.data.test_manifest
     dataset = ManifestDataset.from_config(config)
 
-    model = DeepSpeech2Model.from_pretrained(dataset, config,
+    config.collator.augmentation_config = ""
+    config.collator.keep_transcription_text = True
+    config.collator.batch_size = 1
+    config.collator.num_workers = 0
+    collate_fn = SpeechCollator.from_config(config)
+    test_loader = DataLoader(dataset, collate_fn=collate_fn, num_workers=0)
+
+    model = DeepSpeech2Model.from_pretrained(test_loader, config,
                                              args.checkpoint_path)
     model.eval()
 
     # prepare ASR inference handler
     def file_to_transcript(filename):
-        feature = dataset.process_utterance(filename, "")
-        audio = np.array([feature[0]]).astype('float32')  #[1, D, T]
-        audio_len = feature[0].shape[1]
+        feature = test_loader.collate_fn.process_utterance(filename, "")
+        audio = np.array([feature[0]]).astype('float32')  #[1, T, D]
+        audio_len = feature[0].shape[0]
         audio_len = np.array([audio_len]).astype('int64')  # [1]
 
         result_transcript = model.decode(
             paddle.to_tensor(audio),
             paddle.to_tensor(audio_len),
-            vocab_list=dataset.vocab_list,
+            vocab_list=test_loader.collate_fn.vocab_list,
             decoding_method=config.decoding.decoding_method,
             lang_model_path=config.decoding.lang_model_path,
             beam_alpha=config.decoding.alpha,
@@ -138,7 +145,7 @@ if __name__ == "__main__":
     add_arg('host_ip',          str,
             'localhost',
             "Server's IP address.")
-    add_arg('host_port',        int,    8086,    "Server's IP port.")
+    add_arg('host_port',        int,    8089,    "Server's IP port.")
     add_arg('speech_save_dir',  str,
             'demo_cache',
             "Directory to save demo audios.")

deepspeech/exps/deepspeech2/bin/deploy/server.py

@@ -16,8 +16,10 @@ import functools
 
 import numpy as np
 import paddle
+from paddle.io import DataLoader
 
 from deepspeech.exps.deepspeech2.config import get_cfg_defaults
+from deepspeech.io.collator import SpeechCollator
 from deepspeech.io.dataset import ManifestDataset
 from deepspeech.models.deepspeech2 import DeepSpeech2Model
 from deepspeech.training.cli import default_argument_parser
@@ -31,26 +33,35 @@ from deepspeech.utils.utility import print_arguments
 def start_server(config, args):
     """Start the ASR server"""
     config.defrost()
-    config.data.manfiest = config.data.test_manifest
-    config.data.augmentation_config = ""
-    config.data.keep_transcription_text = True
+    config.data.manifest = config.data.test_manifest
     dataset = ManifestDataset.from_config(config)
 
-    model = DeepSpeech2Model.from_pretrained(dataset, config,
+    config.collator.augmentation_config = ""
+    config.collator.keep_transcription_text = True
+    config.collator.batch_size = 1
+    config.collator.num_workers = 0
+    collate_fn = SpeechCollator.from_config(config)
+    test_loader = DataLoader(dataset, collate_fn=collate_fn, num_workers=0)
+
+    model = DeepSpeech2Model.from_pretrained(test_loader, config,
                                              args.checkpoint_path)
     model.eval()
 
     # prepare ASR inference handler
     def file_to_transcript(filename):
-        feature = dataset.process_utterance(filename, "")
-        audio = np.array([feature[0]]).astype('float32')  #[1, D, T]
-        audio_len = feature[0].shape[1]
+        feature = test_loader.collate_fn.process_utterance(filename, "")
+        audio = np.array([feature[0]]).astype('float32')  #[1, T, D]
+        # audio = audio.swapaxes(1,2)
+        print('---file_to_transcript feature----')
+        print(audio.shape)
+        audio_len = feature[0].shape[0]
+        print(audio_len)
         audio_len = np.array([audio_len]).astype('int64')  # [1]
 
         result_transcript = model.decode(
             paddle.to_tensor(audio),
             paddle.to_tensor(audio_len),
-            vocab_list=dataset.vocab_list,
+            vocab_list=test_loader.collate_fn.vocab_list,
             decoding_method=config.decoding.decoding_method,
             lang_model_path=config.decoding.lang_model_path,
             beam_alpha=config.decoding.alpha,
@@ -91,7 +102,7 @@ if __name__ == "__main__":
     add_arg('host_ip',          str,
             'localhost',
             "Server's IP address.")
-    add_arg('host_port',        int,    8086,    "Server's IP port.")
+    add_arg('host_port',        int,    8088,    "Server's IP port.")
     add_arg('speech_save_dir',  str,
             'demo_cache',
             "Directory to save demo audios.")

deepspeech/exps/deepspeech2/bin/tune.py

@@ -47,7 +47,7 @@ def tune(config, args):
         drop_last=False,
         collate_fn=SpeechCollator(keep_transcription_text=True))
 
-    model = DeepSpeech2Model.from_pretrained(dev_dataset, config,
+    model = DeepSpeech2Model.from_pretrained(valid_loader, config,
                                              args.checkpoint_path)
     model.eval()
 

deepspeech/exps/deepspeech2/model.py

@@ -318,7 +318,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
 
     def export(self):
         infer_model = DeepSpeech2InferModel.from_pretrained(
-            self.test_loader.dataset, self.config, self.args.checkpoint_path)
+            self.test_loader, self.config, self.args.checkpoint_path)
         infer_model.eval()
         feat_dim = self.test_loader.collate_fn.feature_size
         static_model = paddle.jit.to_static(

deepspeech/exps/u2/model.py

@@ -506,7 +506,7 @@ class U2Tester(U2Trainer):
             List[paddle.static.InputSpec]: input spec.
         """
         from deepspeech.models.u2 import U2InferModel
-        infer_model = U2InferModel.from_pretrained(self.test_loader.dataset,
+        infer_model = U2InferModel.from_pretrained(self.test_loader,
                                                    self.config.model.clone(),
                                                    self.args.checkpoint_path)
         feat_dim = self.test_loader.collate_fn.feature_size

deepspeech/io/collator.py

@@ -242,6 +242,7 @@ class SpeechCollator():
 
         # specgram augment
         specgram = self._augmentation_pipeline.transform_feature(specgram)
+        specgram = specgram.transpose([1, 0])
         return specgram, transcript_part
 
     def __call__(self, batch):
@@ -269,7 +270,7 @@ class SpeechCollator():
             #utt
             utts.append(utt)
             # audio
-            audios.append(audio.T)  # [T, D]
+            audios.append(audio)  # [T, D]
             audio_lens.append(audio.shape[1])
             # text
             # for training, text is token ids

deepspeech/models/deepspeech2.py

@@ -198,11 +198,11 @@ class DeepSpeech2Model(nn.Layer):
             cutoff_top_n, num_processes)
 
     @classmethod
-    def from_pretrained(cls, dataset, config, checkpoint_path):
+    def from_pretrained(cls, dataloader, config, checkpoint_path):
         """Build a DeepSpeech2Model model from a pretrained model.
         Parameters
         ----------
-        dataset: paddle.io.Dataset
+        dataloader: paddle.io.DataLoader
 
         config: yacs.config.CfgNode
             model configs
@@ -215,8 +215,8 @@ class DeepSpeech2Model(nn.Layer):
         DeepSpeech2Model
             The model built from pretrained result.
         """
-        model = cls(feat_size=dataset.feature_size,
-                    dict_size=dataset.vocab_size,
+        model = cls(feat_size=dataloader.collate_fn.feature_size,
+                    dict_size=dataloader.collate_fn.vocab_size,
                     num_conv_layers=config.model.num_conv_layers,
                     num_rnn_layers=config.model.num_rnn_layers,
                     rnn_size=config.model.rnn_layer_size,

deepspeech/models/u2.py

@@ -876,11 +876,11 @@ class U2Model(U2BaseModel):
         return model
 
     @classmethod
-    def from_pretrained(cls, dataset, config, checkpoint_path):
+    def from_pretrained(cls, dataloader, config, checkpoint_path):
         """Build a DeepSpeech2Model model from a pretrained model.
 
         Args:
-            dataset (paddle.io.Dataset): not used.
+            dataloader (paddle.io.DataLoader): not used.
             config (yacs.config.CfgNode):  model configs
             checkpoint_path (Path or str): the path of pretrained model checkpoint, without extension name
 
@@ -888,8 +888,8 @@ class U2Model(U2BaseModel):
             DeepSpeech2Model: The model built from pretrained result.
         """
         config.defrost()
-        config.input_dim = dataset.feature_size
-        config.output_dim = dataset.vocab_size
+        config.input_dim = dataloader.collate_fn.feature_size
+        config.output_dim = dataloader.collate_fn.vocab_size
         config.freeze()
         model = cls.from_config(config)
 

deepspeech/utils/socket_server.py

@@ -48,9 +48,9 @@ def warm_up_test(audio_process_handler,
     rng = random.Random(random_seed)
     samples = rng.sample(manifest, num_test_cases)
     for idx, sample in enumerate(samples):
-        print("Warm-up Test Case %d: %s", idx, sample['audio_filepath'])
+        print("Warm-up Test Case %d: %s" % (idx, sample['feat']))
         start_time = time.time()
-        transcript = audio_process_handler(sample['audio_filepath'])
+        transcript = audio_process_handler(sample['feat'])
         finish_time = time.time()
         print("Response Time: %f, Transcript: %s" %
               (finish_time - start_time, transcript))

examples/aishell/s0/README.md

@@ -2,10 +2,10 @@
 
 ## Deepspeech2
 
-| Model | release | Config | Test set | Loss | CER |  
-| --- | --- | --- | --- | --- | --- |  
-| DeepSpeech2 58.4M | 2.1.0 | conf/deepspeech2.yaml + spec aug + new datapipe | test | 6.396368026733398 | 0.068382 ~ 0.073507|  
-| DeepSpeech2 58.4M | 2.1.0 | conf/deepspeech2.yaml + spec aug | test | 7.483316898345947 | 0.077860 |  
-| DeepSpeech2 58.4M | 2.1.0 | conf/deepspeech2.yaml | test | 7.299022197723389 | 0.078671 |
-| DeepSpeech2 58.4M | 2.0.0 | conf/deepspeech2.yaml | test | - | 0.078977 |  
-| DeepSpeech2 58.4M | 1.8.5 | - | test | - | 0.080447 |  
+| Model | Params | Release | Config | Test set | Loss | CER |  
+| --- | --- | --- | --- | --- | --- | --- |  
+| DeepSpeech2 | 58.4M | 2.2.0 | conf/deepspeech2.yaml + spec aug + new datapipe | test | 6.396368026733398 | 0.068382,0.073507 |  
+| DeepSpeech2 | 58.4M | 2.1.0 | conf/deepspeech2.yaml + spec aug | test | 7.483316898345947 | 0.077860 |  
+| DeepSpeech2 | 58.4M | 2.1.0 | conf/deepspeech2.yaml | test | 7.299022197723389 | 0.078671 |
+| DeepSpeech2 | 58.4M | 2.0.0 | conf/deepspeech2.yaml | test | - | 0.078977 |  
+| DeepSpeech2 | 58.4M | 1.8.5 | - | test | - | 0.080447 |  

examples/aishell/s0/conf/deepspeech2.yaml

@@ -10,8 +10,8 @@ data:
   min_output_input_ratio: 0.00
   max_output_input_ratio: .inf
 
-
 collator:
+  batch_size: 64 # one gpu
   mean_std_filepath: data/mean_std.json
   unit_type: char
   vocab_filepath: data/vocab.txt 
@@ -33,7 +33,6 @@ collator:
   sortagrad: True
   shuffle_method: batch_shuffle
   num_workers: 0
-  batch_size: 64 # one gpu
 
 model:
   num_conv_layers: 2

examples/aishell/s0/run.sh

@@ -31,10 +31,10 @@ fi
 
 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
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
     # export ckpt avg_n
-    CUDA_VISIBLE_DEVICES=${gpus} ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
+    CUDA_VISIBLE_DEVICES=0 ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
 fi

examples/aishell/s1/README.md

@@ -9,6 +9,7 @@
 | conformer | 47.07M  | conf/conformer.yaml | spec_aug + shift | test | ctc_prefix_beam_search | - | 0.062196 |  
 | conformer | 47.07M  | conf/conformer.yaml | spec_aug + shift | test | attention_rescoring | - | 0.054694 |  
 
+
 ## Chunk Conformer
 
 | Model | Params | Config | Augmentation| Test set | Decode method | Chunk | Loss | WER |  

examples/librispeech/s0/conf/deepspeech2.yaml

@@ -3,16 +3,21 @@ data:
   train_manifest: data/manifest.train
   dev_manifest: data/manifest.dev-clean
   test_manifest: data/manifest.test-clean
-  mean_std_filepath: data/mean_std.json
-  vocab_filepath: data/vocab.txt 
-  augmentation_config: conf/augmentation.json
-  batch_size: 20
   min_input_len: 0.0
   max_input_len: 27.0 # second
   min_output_len: 0.0
   max_output_len: .inf
   min_output_input_ratio: 0.00
   max_output_input_ratio: .inf
+
+collator:
+  batch_size: 20
+  mean_std_filepath: data/mean_std.json
+  unit_type: char
+  vocab_filepath: data/vocab.txt 
+  augmentation_config: conf/augmentation.json
+  random_seed: 0
+  spm_model_prefix: 
   specgram_type: linear
   target_sample_rate: 16000
   max_freq: None