speech_recognition/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR synced 2025-09-15 14:48:36 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge pull request #317 from xiaowan0322/feat/cuda

Browse Source 
[Export] support gpu inference
This commit is contained in:
zhifu gao 2023-03-30 18:54:20 +08:00 committed by GitHub
commit 2e5cd36e0f
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 26 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
funasr/export/export_model.py
View File

@ -19,6 +19,7 @@ class ModelExport:
self, self,
cache_dir: Union[Path, str] = None, cache_dir: Union[Path, str] = None,
onnx: bool = True, onnx: bool = True,
device: str = "cpu",
quant: bool = True, quant: bool = True,
fallback_num: int = 0, fallback_num: int = 0,
audio_in: str = None, audio_in: str = None,
@ -36,6 +37,7 @@ class ModelExport:
) )
print("output dir: {}".format(self.cache_dir)) print("output dir: {}".format(self.cache_dir))
self.onnx = onnx self.onnx = onnx
self.device = device
self.quant = quant self.quant = quant
self.fallback_num = fallback_num self.fallback_num = fallback_num
self.frontend = None self.frontend = None
@ -112,6 +114,10 @@ class ModelExport:
else: else:
dummy_input = model.get_dummy_inputs() dummy_input = model.get_dummy_inputs()
if self.device == 'cuda':
model = model.cuda()
dummy_input = tuple([i.cuda() for i in dummy_input])
# model_script = torch.jit.script(model) # model_script = torch.jit.script(model)
model_script = torch.jit.trace(model, dummy_input) model_script = torch.jit.trace(model, dummy_input)
model_script.save(os.path.join(path, f'{model.model_name}.torchscripts')) model_script.save(os.path.join(path, f'{model.model_name}.torchscripts'))
@ -234,6 +240,7 @@ if __name__ == '__main__':
parser.add_argument('--model-name', type=str, required=True) parser.add_argument('--model-name', type=str, required=True)
parser.add_argument('--export-dir', type=str, required=True) parser.add_argument('--export-dir', type=str, required=True)
parser.add_argument('--type', type=str, default='onnx', help='["onnx", "torch"]') parser.add_argument('--type', type=str, default='onnx', help='["onnx", "torch"]')
parser.add_argument('--device', type=str, default='cpu', help='["cpu", "cuda"]')
parser.add_argument('--quantize', type=str2bool, default=False, help='export quantized model') parser.add_argument('--quantize', type=str2bool, default=False, help='export quantized model')
parser.add_argument('--fallback-num', type=int, default=0, help='amp fallback number') parser.add_argument('--fallback-num', type=int, default=0, help='amp fallback number')
parser.add_argument('--audio_in', type=str, default=None, help='["wav", "wav.scp"]') parser.add_argument('--audio_in', type=str, default=None, help='["wav", "wav.scp"]')
@ -243,6 +250,7 @@ if __name__ == '__main__':
export_model = ModelExport( export_model = ModelExport(
cache_dir=args.export_dir, cache_dir=args.export_dir,
onnx=args.type == 'onnx', onnx=args.type == 'onnx',
device=args.device,
quant=args.quantize, quant=args.quantize,
fallback_num=args.fallback_num, fallback_num=args.fallback_num,
audio_in=args.audio_in, audio_in=args.audio_in,

15
funasr/runtime/python/libtorch/demo.py
View File

@ -1,20 +1,15 @@
from funasr_torch import Paraformer from funasr_torch import Paraformer
#model_dir = "/Users/shixian/code/funasr/export/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
model_dir = "/Users/shixian/code/funasr/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
model = Paraformer(model_dir, batch_size=2)
# when using paraformer-tiny-commandword_asr_nat-zh-cn-16k-vocab544-pytorch, you should set pred_bias=0 model_dir = "/nfs/zhifu.gzf/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
# plot_timestamp_to works only when using speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch
# model_dir = "/Users/shixian/code/funasr/export/damo/speech_paraformer-tiny-commandword_asr_nat-zh-cn-16k-vocab544-pytorch" model = Paraformer(model_dir, batch_size=1) # cpu
# model = Paraformer(model_dir, batch_size=2, pred_bias=0) # model = Paraformer(model_dir, batch_size=1, device_id=0) # gpu
# when using paraformer-large-vad-punc model, you can set plot_timestamp_to="./xx.png" to get figure of alignment besides timestamps # when using paraformer-large-vad-punc model, you can set plot_timestamp_to="./xx.png" to get figure of alignment besides timestamps
# model_dir = "/Users/shixian/code/funasr/export/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
# model = Paraformer(model_dir, batch_size=1)
# model = Paraformer(model_dir, batch_size=1, plot_timestamp_to="test.png") # model = Paraformer(model_dir, batch_size=1, plot_timestamp_to="test.png")
wav_path = "YourPath/xx.wav" wav_path = "YourPath/xx.wav"
result = model(wav_path) result = model(wav_path)
print(result) print(result)

10
funasr/runtime/python/libtorch/funasr_torch/paraformer_bin.py
View File

@ -46,6 +46,7 @@ class Paraformer():
) )
self.ort_infer = torch.jit.load(model_file) self.ort_infer = torch.jit.load(model_file)
self.batch_size = batch_size self.batch_size = batch_size
self.device_id = device_id
self.plot_timestamp_to = plot_timestamp_to self.plot_timestamp_to = plot_timestamp_to
self.pred_bias = pred_bias self.pred_bias = pred_bias
@ -58,8 +59,13 @@ class Paraformer():
end_idx = min(waveform_nums, beg_idx + self.batch_size) end_idx = min(waveform_nums, beg_idx + self.batch_size)
feats, feats_len = self.extract_feat(waveform_list[beg_idx:end_idx]) feats, feats_len = self.extract_feat(waveform_list[beg_idx:end_idx])
try: try:
outputs = self.ort_infer(feats, feats_len) with torch.no_grad():
am_scores, valid_token_lens = outputs[0], outputs[1] if int(self.device_id) == -1:
outputs = self.ort_infer(feats, feats_len)
am_scores, valid_token_lens = outputs[0], outputs[1]
else:
outputs = self.ort_infer(feats.cuda(), feats_len.cuda())
am_scores, valid_token_lens = outputs[0].cpu(), outputs[1].cpu()
if len(outputs) == 4: if len(outputs) == 4:
# for BiCifParaformer Inference # for BiCifParaformer Inference
us_alphas, us_peaks = outputs[2], outputs[3] us_alphas, us_peaks = outputs[2], outputs[3]

15
funasr/runtime/python/onnxruntime/demo.py
View File

@ -1,20 +1,15 @@
from funasr_onnx import Paraformer from funasr_onnx import Paraformer
#model_dir = "/Users/shixian/code/funasr/export/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch" model_dir = "/nfs/zhifu.gzf/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
model_dir = "/Users/shixian/code/funasr/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
model = Paraformer(model_dir, batch_size=2)
# when using paraformer-tiny-commandword_asr_nat-zh-cn-16k-vocab544-pytorch, you should set pred_bias=0 model = Paraformer(model_dir, batch_size=2, plot_timestamp_to="./", pred_bias=0) # cpu
# plot_timestamp_to works only when using speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch # model = Paraformer(model_dir, batch_size=2, plot_timestamp_to="./", pred_bias=0, device_id=0) # gpu
# model_dir = "/Users/shixian/code/funasr/export/damo/speech_paraformer-tiny-commandword_asr_nat-zh-cn-16k-vocab544-pytorch"
# model = Paraformer(model_dir, batch_size=2, pred_bias=0)
# when using paraformer-large-vad-punc model, you can set plot_timestamp_to="./xx.png" to get figure of alignment besides timestamps # when using paraformer-large-vad-punc model, you can set plot_timestamp_to="./xx.png" to get figure of alignment besides timestamps
# model_dir = "/Users/shixian/code/funasr/export/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
# model = Paraformer(model_dir, batch_size=1)
# model = Paraformer(model_dir, batch_size=1, plot_timestamp_to="test.png") # model = Paraformer(model_dir, batch_size=1, plot_timestamp_to="test.png")
wav_path = "YourPath/xx.wav" wav_path = "YourPath/xx.wav"
result = model(wav_path) result = model(wav_path)
print(result) print(result)