From 0856ea2ebdcb976db6e786de5cd79fae3d35cd4c Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期一, 20 二月 2023 18:18:35 +0800
Subject: [PATCH] Merge pull request #136 from alibaba-damo-academy/dev_cmz
---
funasr/export/export_model.py | 94 ++++++++++++++++++++++++++--------------------
1 files changed, 53 insertions(+), 41 deletions(-)
diff --git a/funasr/export/export_model.py b/funasr/export/export_model.py
index e5a2320..972f92f 100644
--- a/funasr/export/export_model.py
+++ b/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
@@ -8,24 +9,26 @@
from funasr.bin.asr_inference_paraformer import Speech2Text
from funasr.export.models import get_model
-
-
+import numpy as np
+import random
class ASRModelExportParaformer:
def __init__(self, cache_dir: Union[Path, str] = None, onnx: bool = True):
assert check_argument_types()
+ self.set_all_random_seed(0)
if cache_dir is None:
- cache_dir = Path.home() / "cache" / "export"
+ cache_dir = Path.home() / ".cache" / "export"
self.cache_dir = Path(cache_dir)
self.export_config = dict(
feats_dim=560,
- onnx=onnx,
+ onnx=False,
)
- logging.info("output dir: {}".format(self.cache_dir))
+ print("output dir: {}".format(self.cache_dir))
self.onnx = onnx
+
- def export(
+ def _export(
self,
model: Speech2Text,
tag_name: str = None,
@@ -41,57 +44,58 @@
model,
self.export_config,
)
- self._export_onnx(model, verbose, export_dir)
- # if self.onnx:
- # self._export_onnx(model, verbose, export_dir)
- # else:
- # self._export_torchscripts(model, verbose, export_dir)
+ model.eval()
+ # self._export_onnx(model, verbose, export_dir)
+ if self.onnx:
+ self._export_onnx(model, verbose, export_dir)
+ else:
+ self._export_torchscripts(model, verbose, export_dir)
- logging.info("output dir: {}".format(export_dir))
+ print("output dir: {}".format(export_dir))
def _export_torchscripts(self, model, verbose, path, enc_size=None):
if enc_size:
dummy_input = model.get_dummy_inputs(enc_size)
else:
- dummy_input = model.get_dummy_inputs()
+ dummy_input = model.get_dummy_inputs_txt()
# model_script = torch.jit.script(model)
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 set_all_random_seed(self, seed: int):
+ random.seed(seed)
+ np.random.seed(seed)
+ torch.random.manual_seed(seed)
+ 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(model_dir, 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)
+ self._export(model, tag_name)
+
def _export_onnx(self, model, verbose, path, enc_size=None):
if enc_size:
@@ -114,7 +118,15 @@
)
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')
\ No newline at end of file
+ import sys
+
+ model_path = sys.argv[1]
+ output_dir = sys.argv[2]
+ onnx = sys.argv[3]
+ onnx = onnx.lower()
+ onnx = onnx == 'true'
+ # model_path = 'damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch'
+ # output_dir = "../export"
+ export_model = ASRModelExportParaformer(cache_dir=output_dir, onnx=onnx)
+ export_model.export(model_path)
+ # export_model.export('/root/cache/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch')
\ No newline at end of file
--
Gitblit v1.9.1