funasr1.0

docs/benchmark/benchmark_libtorch.md

@@ -1 +0,0 @@
-../../funasr/runtime/docs/benchmark_libtorch.md

docs/benchmark/benchmark_onnx.md

@@ -1 +0,0 @@
-../../funasr/runtime/docs/benchmark_onnx.md

docs/benchmark/benchmark_onnx_cpp.md

@@ -1 +0,0 @@
-../../funasr/runtime/docs/benchmark_onnx_cpp.md

examples/industrial_data_pretraining/paraformer/README_zh.md

@@ -0,0 +1,42 @@
+(简体中文|[English](./README.md))
+
+# 语音识别
+
+> **注意**:
+> pipeline 支持 [modelscope模型仓库](https://alibaba-damo-academy.github.io/FunASR/en/model_zoo/modelscope_models.html#pretrained-models-on-modelscope) 中的所有模型进行推理和微调。这里我们以典型模型作为示例来演示使用方法。
+
+## 推理
+
+### 快速使用
+#### [Paraformer 模型](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary)
+```python
+from funasr import AutoModel
+
+model = AutoModel(model="/Users/zhifu/Downloads/modelscope_models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch")
+
+res = model(input="/Users/zhifu/Downloads/modelscope_models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/example/asr_example.wav")
+print(res)
+```
+
+### API接口说明
+#### AutoModel 定义
+- `model`: [模型仓库](https://alibaba-damo-academy.github.io/FunASR/en/model_zoo/modelscope_models.html#pretrained-models-on-modelscope) 中的模型名称，或本地磁盘中的模型路径
+- `device`: `cuda`（默认），使用 GPU 进行推理。如果为`cpu`，则使用 CPU 进行推理
+- `ncpu`: `None` （默认），设置用于 CPU 内部操作并行性的线程数
+- `output_dir`: `None` （默认），如果设置，输出结果的输出路径
+- `batch_size`: `1` （默认），解码时的批处理大小
+#### AutoModel 推理
+- `input`: 要解码的输入，可以是：
+  - wav文件路径, 例如: asr_example.wav
+  - pcm文件路径, 例如: asr_example.pcm，此时需要指定音频采样率fs（默认为16000）
+  - 音频字节数流，例如：麦克风的字节数数据
+  - wav.scp，kaldi 样式的 wav 列表 (`wav_id \t wav_path`), 例如:
+  ```text
+  asr_example1  ./audios/asr_example1.wav
+  asr_example2  ./audios/asr_example2.wav
+  ```
+  在这种输入 `wav.scp` 的情况下，必须设置 `output_dir` 以保存输出结果
+  - 音频采样点，例如：`audio, rate = soundfile.read("asr_example_zh.wav")`, 数据类型为 numpy.ndarray。支持batch输入，类型为list：
+  ```[audio_sample1, audio_sample2, ..., audio_sampleN]```
+  - fbank输入，支持组batch。shape为[batch, frames, dim]，类型为torch.Tensor，例如
+- `output_dir`: None （默认），如果设置，输出结果的输出路径

examples/industrial_data_pretraining/paraformer/demo.py

@@ -8,4 +8,15 @@ from funasr import AutoModel
 model = AutoModel(model="/Users/zhifu/Downloads/modelscope_models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch")
 
 res = model(input="/Users/zhifu/Downloads/modelscope_models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/example/asr_example.wav")
-print(res)
+print(res)
+
+
+from funasr import AutoFrontend
+
+frontend = AutoFrontend(model="/Users/zhifu/Downloads/modelscope_models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch")
+
+fbanks = frontend(input="/Users/zhifu/funasr_github/test_local/wav.scp", batch_size=2)
+
+for batch_idx, fbank_dict in enumerate(fbanks):
+	res = model(**fbank_dict)
+	print(res)

funasr/__init__.py

@@ -30,4 +30,4 @@ def import_submodules(package, recursive=True):
 
 import_submodules(__name__)
 
-from funasr.bin.inference import AutoModel
+from funasr.bin.inference import AutoModel, AutoFrontend

funasr/bin/inference.py

@@ -16,11 +16,12 @@ import time
 import random
 import string
 from funasr.register import tables
-from funasr.datasets.audio_datasets.load_audio_extract_fbank import load_audio
+
+from funasr.datasets.audio_datasets.load_audio_extract_fbank import load_audio, extract_fbank
 from funasr.utils.vad_utils import slice_padding_audio_samples
 from funasr.utils.timestamp_tools import time_stamp_sentence
 
-def build_iter_for_infer(data_in, input_len=None, data_type="sound"):
+def build_iter_for_infer(data_in, input_len=None, data_type="sound", key=None):
 	"""
 	
 	:param input:
@@ -63,7 +64,8 @@ def build_iter_for_infer(data_in, input_len=None, data_type="sound"):
 	else: # raw text; audio sample point, fbank; bytes
 		if isinstance(data_in, bytes): # audio bytes
 			data_in = load_bytes(data_in)
-		key = "rand_key_" + ''.join(random.choice(chars) for _ in range(13))
+		if key is None:
+			key = "rand_key_" + ''.join(random.choice(chars) for _ in range(13))
 		data_list = [data_in]
 		key_list = [key]
 	
@@ -121,11 +123,14 @@ class AutoModel:
 		set_all_random_seed(kwargs.get("seed", 0))
 		
 		device = kwargs.get("device", "cuda")
-		if not torch.cuda.is_available() or kwargs.get("ngpu", 1):
+		if not torch.cuda.is_available() or kwargs.get("ngpu", 0):
 			device = "cpu"
 			kwargs["batch_size"] = 1
 		kwargs["device"] = device
 		
+		if kwargs.get("ncpu", None):
+			torch.set_num_threads(kwargs.get("ncpu"))
+		
 		# build tokenizer
 		tokenizer = kwargs.get("tokenizer", None)
 		if tokenizer is not None:
@@ -169,17 +174,18 @@ class AutoModel:
 		else:
 			return self.generate_with_vad(input, input_len=input_len, **cfg)
 		
-	def generate(self, input, input_len=None, model=None, kwargs=None, **cfg):
+	def generate(self, input, input_len=None, model=None, kwargs=None, key=None, **cfg):
+		# import pdb; pdb.set_trace()
 		kwargs = self.kwargs if kwargs is None else kwargs
 		kwargs.update(cfg)
 		model = self.model if model is None else model
 		
 		data_type = kwargs.get("data_type", "sound")
 		batch_size = kwargs.get("batch_size", 1)
-		# if kwargs.get("device", "cpu") == "cpu":
-		# 	batch_size = 1
+		if kwargs.get("device", "cpu") == "cpu":
+			batch_size = 1
 		
-		key_list, data_list = build_iter_for_infer(input, input_len=input_len, data_type=data_type)
+		key_list, data_list = build_iter_for_infer(input, input_len=input_len, data_type=data_type, key=key)
 		
 		speed_stats = {}
 		asr_result_list = []
@@ -193,7 +199,7 @@ class AutoModel:
 			key_batch = key_list[beg_idx:end_idx]
 			batch = {"data_in": data_batch, "key": key_batch}
 			if (end_idx - beg_idx) == 1 and isinstance(data_batch[0], torch.Tensor): # fbank
-				batch["data_batch"] = data_batch[0]
+				batch["data_in"] = data_batch[0]
 				batch["data_lengths"] = input_len
 		
 			time1 = time.perf_counter()
@@ -348,6 +354,74 @@ class AutoModel:
 		                     f"time_speech_total_all_samples: {time_speech_total_all_samples: 0.3f}, "
 		                     f"time_escape_total_all_samples: {time_escape_total_all_samples:0.3f}")
 		return results_ret_list
+
+
+class AutoFrontend:
+	def __init__(self, **kwargs):
+		assert "model" in kwargs
+		if "model_conf" not in kwargs:
+			logging.info("download models from model hub: {}".format(kwargs.get("model_hub", "ms")))
+			kwargs = download_model(**kwargs)
+		
+		# build frontend
+		frontend = kwargs.get("frontend", None)
+		if frontend is not None:
+			frontend_class = tables.frontend_classes.get(frontend.lower())
+			frontend = frontend_class(**kwargs["frontend_conf"])
+
+		self.frontend = frontend
+		self.kwargs = kwargs
 	
+	def __call__(self, input, input_len=None, kwargs=None, **cfg):
+		
+		kwargs = self.kwargs if kwargs is None else kwargs
+		kwargs.update(cfg)
+
+
+		key_list, data_list = build_iter_for_infer(input, input_len=input_len)
+		batch_size = kwargs.get("batch_size", 1)
+		device = kwargs.get("device", "cpu")
+		if device == "cpu":
+			batch_size = 1
+		
+		meta_data = {}
+		
+		result_list = []
+		num_samples = len(data_list)
+		pbar = tqdm(colour="blue", total=num_samples + 1, dynamic_ncols=True)
+		
+		time0 = time.perf_counter()
+		for beg_idx in range(0, num_samples, batch_size):
+			end_idx = min(num_samples, beg_idx + batch_size)
+			data_batch = data_list[beg_idx:end_idx]
+			key_batch = key_list[beg_idx:end_idx]
+
+			# extract fbank feats
+			time1 = time.perf_counter()
+			audio_sample_list = load_audio(data_batch, fs=self.frontend.fs, audio_fs=kwargs.get("fs", 16000))
+			time2 = time.perf_counter()
+			meta_data["load_data"] = f"{time2 - time1:0.3f}"
+			speech, speech_lengths = extract_fbank(audio_sample_list, data_type=kwargs.get("data_type", "sound"),
+			                                       frontend=self.frontend)
+			time3 = time.perf_counter()
+			meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
+			meta_data["batch_data_time"] = speech_lengths.sum().item() * self.frontend.frame_shift * self.frontend.lfr_n / 1000
+			
+			speech.to(device=device), speech_lengths.to(device=device)
+			batch = {"input": speech, "input_len": speech_lengths, "key": key_batch}
+			result_list.append(batch)
+			
+			pbar.update(1)
+			description = (
+				f"{meta_data}, "
+			)
+			pbar.set_description(description)
+		
+		time_end = time.perf_counter()
+		pbar.set_description(f"time escaped total: {time_end - time0:0.3f}")
+		
+		return result_list
+
+
 if __name__ == '__main__':
 	main_hydra()

funasr/models/paraformer/model.py

@@ -495,6 +495,8 @@ class Paraformer(nn.Module):
 
 		results = []
 		b, n, d = decoder_out.size()
+		if isinstance(key[0], (list, tuple)):
+			key = key[0]
 		for i in range(b):
 			x = encoder_out[i, :encoder_out_lens[i], :]
 			am_scores = decoder_out[i, :pre_token_length[i], :]
@@ -535,6 +537,7 @@ class Paraformer(nn.Module):
 					text = tokenizer.tokens2text(token)
 					
 					text_postprocessed, _ = postprocess_utils.sentence_postprocess(token)
+					
 					result_i = {"key": key[i], "text": text_postprocessed}