From edec2fe85eda80ff1e24aef30b36c7bbbb55ec2a Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期一, 03 七月 2023 15:08:11 +0800
Subject: [PATCH] Update SDK_tutorial_zh.md
---
funasr/export/export_model.py | 231 +++++++++++++++++++++++++++++++++++++++++++--------------
1 files changed, 172 insertions(+), 59 deletions(-)
diff --git a/funasr/export/export_model.py b/funasr/export/export_model.py
index 1c677c9..f31f960 100644
--- a/funasr/export/export_model.py
+++ b/funasr/export/export_model.py
@@ -1,7 +1,6 @@
import json
from typing import Union, Dict
from pathlib import Path
-from typeguard import check_argument_types
import os
import logging
@@ -10,27 +9,38 @@
from funasr.export.models import get_model
import numpy as np
import random
-
+from funasr.utils.types import str2bool, str2triple_str
# torch_version = float(".".join(torch.__version__.split(".")[:2]))
# assert torch_version > 1.9
-class ASRModelExportParaformer:
+class ModelExport:
def __init__(
- self, cache_dir: Union[Path, str] = None, onnx: bool = True, quant: bool = True
+ self,
+ cache_dir: Union[Path, str] = None,
+ onnx: bool = True,
+ device: str = "cpu",
+ quant: bool = True,
+ fallback_num: int = 0,
+ audio_in: str = None,
+ calib_num: int = 200,
+ model_revision: str = None,
):
- assert check_argument_types()
self.set_all_random_seed(0)
- if cache_dir is None:
- cache_dir = Path.home() / ".cache" / "export"
- self.cache_dir = Path(cache_dir)
+ self.cache_dir = cache_dir
self.export_config = dict(
feats_dim=560,
onnx=False,
)
- print("output dir: {}".format(self.cache_dir))
+
self.onnx = onnx
+ self.device = device
self.quant = quant
+ self.fallback_num = fallback_num
+ self.frontend = None
+ self.audio_in = audio_in
+ self.calib_num = calib_num
+ self.model_revision = model_revision
def _export(
@@ -40,7 +50,7 @@
verbose: bool = False,
):
- export_dir = self.cache_dir / tag_name.replace(' ', '-')
+ export_dir = self.cache_dir
os.makedirs(export_dir, exist_ok=True)
# export encoder1
@@ -60,8 +70,19 @@
def _torch_quantize(self, model):
+ def _run_calibration_data(m):
+ # using dummy inputs for a example
+ if self.audio_in is not None:
+ feats, feats_len = self.load_feats(self.audio_in)
+ for i, (feat, len) in enumerate(zip(feats, feats_len)):
+ with torch.no_grad():
+ m(feat, len)
+ else:
+ dummy_input = model.get_dummy_inputs()
+ m(*dummy_input)
+
+
from torch_quant.module import ModuleFilter
- from torch_quant.observer import HistogramObserver
from torch_quant.quantizer import Backend, Quantizer
from funasr.export.models.modules.decoder_layer import DecoderLayerSANM
from funasr.export.models.modules.encoder_layer import EncoderLayerSANM
@@ -70,22 +91,30 @@
quantizer = Quantizer(
module_filter=module_filter,
backend=Backend.FBGEMM,
- act_ob_ctr=HistogramObserver,
)
model.eval()
calib_model = quantizer.calib(model)
- # run calibration data
- # using dummy inputs for a example
- dummy_input = model.get_dummy_inputs()
- _ = calib_model(*dummy_input)
+ _run_calibration_data(calib_model)
+ if self.fallback_num > 0:
+ # perform automatic mixed precision quantization
+ amp_model = quantizer.amp(model)
+ _run_calibration_data(amp_model)
+ quantizer.fallback(amp_model, num=self.fallback_num)
+ print('Fallback layers:')
+ print('\n'.join(quantizer.module_filter.exclude_names))
quant_model = quantizer.quantize(model)
return quant_model
+
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()
+
+ 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.trace(model, dummy_input)
@@ -101,32 +130,94 @@
random.seed(seed)
np.random.seed(seed)
torch.random.manual_seed(seed)
+
+ def parse_audio_in(self, audio_in):
+
+ wav_list, name_list = [], []
+ if audio_in.endswith(".scp"):
+ f = open(audio_in, 'r')
+ lines = f.readlines()[:self.calib_num]
+ for line in lines:
+ name, path = line.strip().split()
+ name_list.append(name)
+ wav_list.append(path)
+ else:
+ wav_list = [audio_in,]
+ name_list = ["test",]
+ return wav_list, name_list
+
+ def load_feats(self, audio_in: str = None):
+ import torchaudio
+
+ wav_list, name_list = self.parse_audio_in(audio_in)
+ feats = []
+ feats_len = []
+ for line in wav_list:
+ path = line.strip()
+ waveform, sampling_rate = torchaudio.load(path)
+ if sampling_rate != self.frontend.fs:
+ waveform = torchaudio.transforms.Resample(orig_freq=sampling_rate,
+ new_freq=self.frontend.fs)(waveform)
+ fbank, fbank_len = self.frontend(waveform, [waveform.size(1)])
+ feats.append(fbank)
+ feats_len.append(fbank_len)
+ return feats, feats_len
+
def export(self,
tag_name: str = 'damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
- mode: str = 'paraformer',
+ mode: str = None,
):
model_dir = tag_name
- if model_dir.startswith('damo/'):
+ 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')
+ model_dir = snapshot_download(model_dir, cache_dir=self.cache_dir, revision=self.model_revision)
+ self.cache_dir = model_dir
+
if mode is None:
import json
+ json_file = os.path.join(model_dir, 'configuration.json')
with open(json_file, 'r') as f:
config_data = json.load(f)
- mode = config_data['model']['model_config']['mode']
+ if config_data['task'] == "punctuation":
+ mode = config_data['model']['punc_model_config']['mode']
+ else:
+ mode = config_data['model']['model_config']['mode']
if mode.startswith('paraformer'):
from funasr.tasks.asr import ASRTaskParaformer as ASRTask
- elif mode.startswith('uniasr'):
- from funasr.tasks.asr import ASRTaskUniASR as ASRTask
+ config = os.path.join(model_dir, 'config.yaml')
+ 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(
+ config, model_file, cmvn_file, 'cpu'
+ )
+ self.frontend = model.frontend
+ self.export_config["feats_dim"] = 560
+ elif mode.startswith('offline'):
+ from funasr.tasks.vad import VADTask
+ config = os.path.join(model_dir, 'vad.yaml')
+ model_file = os.path.join(model_dir, 'vad.pb')
+ cmvn_file = os.path.join(model_dir, 'vad.mvn')
- model, asr_train_args = ASRTask.build_model_from_file(
- asr_train_config, asr_model_file, cmvn_file, 'cpu'
- )
+ model, vad_infer_args = VADTask.build_model_from_file(
+ config, model_file, cmvn_file=cmvn_file, device='cpu'
+ )
+ self.export_config["feats_dim"] = 400
+ self.frontend = model.frontend
+ elif mode.startswith('punc'):
+ from funasr.tasks.punctuation import PunctuationTask as PUNCTask
+ punc_train_config = os.path.join(model_dir, 'config.yaml')
+ punc_model_file = os.path.join(model_dir, 'punc.pb')
+ model, punc_train_args = PUNCTask.build_model_from_file(
+ punc_train_config, punc_model_file, 'cpu'
+ )
+ elif mode.startswith('punc_VadRealtime'):
+ from funasr.tasks.punctuation import PunctuationTask as PUNCTask
+ punc_train_config = os.path.join(model_dir, 'config.yaml')
+ punc_model_file = os.path.join(model_dir, 'punc.pb')
+ model, punc_train_args = PUNCTask.build_model_from_file(
+ punc_train_config, punc_model_file, 'cpu'
+ )
self._export(model, tag_name)
@@ -139,40 +230,62 @@
# model_script = torch.jit.script(model)
model_script = model #torch.jit.trace(model)
model_path = os.path.join(path, f'{model.model_name}.onnx')
-
- torch.onnx.export(
- model_script,
- dummy_input,
- model_path,
- verbose=verbose,
- opset_version=14,
- input_names=model.get_input_names(),
- output_names=model.get_output_names(),
- dynamic_axes=model.get_dynamic_axes()
- )
+ if not os.path.exists(model_path):
+ torch.onnx.export(
+ model_script,
+ dummy_input,
+ model_path,
+ verbose=verbose,
+ opset_version=14,
+ input_names=model.get_input_names(),
+ output_names=model.get_output_names(),
+ dynamic_axes=model.get_dynamic_axes()
+ )
if self.quant:
from onnxruntime.quantization import QuantType, quantize_dynamic
+ import onnx
quant_model_path = os.path.join(path, f'{model.model_name}_quant.onnx')
- quantize_dynamic(
- model_input=model_path,
- model_output=quant_model_path,
- weight_type=QuantType.QUInt8,
- )
+ if not os.path.exists(quant_model_path):
+ onnx_model = onnx.load(model_path)
+ nodes = [n.name for n in onnx_model.graph.node]
+ nodes_to_exclude = [m for m in nodes if 'output' in m]
+ quantize_dynamic(
+ model_input=model_path,
+ model_output=quant_model_path,
+ op_types_to_quantize=['MatMul'],
+ per_channel=True,
+ reduce_range=False,
+ weight_type=QuantType.QUInt8,
+ nodes_to_exclude=nodes_to_exclude,
+ )
if __name__ == '__main__':
- import sys
-
- model_path = sys.argv[1]
- output_dir = sys.argv[2]
- onnx = sys.argv[3]
- quant = sys.argv[4]
- onnx = onnx.lower()
- onnx = onnx == 'true'
- quant = quant == '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, quant=quant)
- export_model.export(model_path)
- # export_model.export('/root/cache/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch')
+ import argparse
+ parser = argparse.ArgumentParser()
+ # parser.add_argument('--model-name', type=str, required=True)
+ parser.add_argument('--model-name', type=str, action="append", required=True, default=[])
+ parser.add_argument('--export-dir', type=str, required=True)
+ 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('--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('--calib_num', type=int, default=200, help='calib max num')
+ parser.add_argument('--model_revision', type=str, default=None, help='model_revision')
+ args = parser.parse_args()
+
+ export_model = ModelExport(
+ cache_dir=args.export_dir,
+ onnx=args.type == 'onnx',
+ device=args.device,
+ quant=args.quantize,
+ fallback_num=args.fallback_num,
+ audio_in=args.audio_in,
+ calib_num=args.calib_num,
+ model_revision=args.model_revision,
+ )
+ for model_name in args.model_name:
+ print("export model: {}".format(model_name))
+ export_model.export(model_name)
--
Gitblit v1.9.1