exoprt model

游雁

2023-02-10 8bf1a6adbb9b412615a1d121ae2fa0c1776f1c48

 funasr/export/export_model.py

@@ -1,3 +1,4 @@
import json
from typing import Union, Dict
from pathlib import Path
from typeguard import check_argument_types
@@ -25,7 +26,7 @@
        logging.info("output dir: {}".format(self.cache_dir))
        self.onnx = onnx

    def export(
    def _export(
        self,
        model: Speech2Text,
        tag_name: str = None,
@@ -60,38 +61,66 @@
        model_script = torch.jit.trace(model, dummy_input)
        model_script.save(os.path.join(path, f'{model.model_name}.torchscripts'))

    def export_from_modelscope(
        self,
        tag_name: str = 'damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
    ):
    def export(self,
               tag_name: str = 'damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
               mode: str = 'paraformer',
               ):
        
        from funasr.tasks.asr import ASRTaskParaformer as ASRTask
        from modelscope.hub.snapshot_download import snapshot_download

        model_dir = snapshot_download(tag_name, cache_dir=self.cache_dir)
        asr_train_config = os.path.join(model_dir, 'config.yaml')
        asr_model_file = os.path.join(model_dir, 'model.pb')
        cmvn_file = os.path.join(model_dir, 'am.mvn')
        model, asr_train_args = ASRTask.build_model_from_file(
            asr_train_config, asr_model_file, cmvn_file, 'cpu'
        )
        self.export(model, tag_name)

    def export_from_local(
        self,
        tag_name: str = '/root/cache/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
    ):
    
        from funasr.tasks.asr import ASRTaskParaformer as ASRTask
    
        model_dir = tag_name
        if model_dir.startswith('damo/'):
            from modelscope.hub.snapshot_download import snapshot_download
            model_dir = snapshot_download(tag_name, cache_dir=self.cache_dir)
        asr_train_config = os.path.join(model_dir, 'config.yaml')
        asr_model_file = os.path.join(model_dir, 'model.pb')
        cmvn_file = os.path.join(model_dir, 'am.mvn')
        json_file = os.path.join(model_dir, 'configuration.json')
        if mode is None:
            import json
            with open(json_file, 'r') as f:
                config_data = json.load(f)
                mode = config_data['model']['model_config']['mode']
        if mode == 'paraformer':
            from funasr.tasks.asr import ASRTaskParaformer as ASRTask
        elif mode == 'uniasr':
            from funasr.tasks.asr import ASRTaskUniASR as ASRTask
            
        model, asr_train_args = ASRTask.build_model_from_file(
            asr_train_config, asr_model_file, cmvn_file, 'cpu'
        )
        self.export(model, tag_name)
            
    # def export_from_modelscope(
    #     self,
    #     tag_name: str = 'damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
    # ):
    #
    #     from funasr.tasks.asr import ASRTaskParaformer as ASRTask
    #     from modelscope.hub.snapshot_download import snapshot_download
    #
    #     model_dir = snapshot_download(tag_name, cache_dir=self.cache_dir)
    #     asr_train_config = os.path.join(model_dir, 'config.yaml')
    #     asr_model_file = os.path.join(model_dir, 'model.pb')
    #     cmvn_file = os.path.join(model_dir, 'am.mvn')
    #     model, asr_train_args = ASRTask.build_model_from_file(
    #         asr_train_config, asr_model_file, cmvn_file, 'cpu'
    #     )
    #     self.export(model, tag_name)
    #
    # def export_from_local(
    #     self,
    #     tag_name: str = '/root/cache/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
    # ):
    #
    #     from funasr.tasks.asr import ASRTaskParaformer as ASRTask
    #
    #     model_dir = tag_name
    #     asr_train_config = os.path.join(model_dir, 'config.yaml')
    #     asr_model_file = os.path.join(model_dir, 'model.pb')
    #     cmvn_file = os.path.join(model_dir, 'am.mvn')
    #     model, asr_train_args = ASRTask.build_model_from_file(
    #         asr_train_config, asr_model_file, cmvn_file, 'cpu'
    #     )
    #     self.export(model, tag_name)

    def _export_onnx(self, model, verbose, path, enc_size=None):
        if enc_size:
@@ -116,5 +145,5 @@
if __name__ == '__main__':
    output_dir = "../export"
    export_model = ASRModelExportParaformer(cache_dir=output_dir, onnx=False)
    export_model.export_from_modelscope('damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch')
    # export_model.export_from_local('/root/cache/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch')
    export_model.export('damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch')
    # export_model.export('/root/cache/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch')