From 111450d5e67ed01e3e7de088fe5bb3115dc7eeff Mon Sep 17 00:00:00 2001
From: mengzhe.cmz <mengzhe.cmz@alibaba-inc.com>
Date: 星期四, 23 三月 2023 19:55:15 +0800
Subject: [PATCH] update infer.py
---
funasr/export/export_model.py | 117 +++++++++++++++++++++++++++++++++++++++++++++++-----------
1 files changed, 95 insertions(+), 22 deletions(-)
diff --git a/funasr/export/export_model.py b/funasr/export/export_model.py
index 1c677c9..f6ba616 100644
--- a/funasr/export/export_model.py
+++ b/funasr/export/export_model.py
@@ -10,13 +10,19 @@
from funasr.export.models import get_model
import numpy as np
import random
-
+from funasr.utils.types import str2bool
# torch_version = float(".".join(torch.__version__.split(".")[:2]))
# assert torch_version > 1.9
class ASRModelExportParaformer:
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,
+ quant: bool = True,
+ fallback_num: int = 0,
+ audio_in: str = None,
+ calib_num: int = 200,
):
assert check_argument_types()
self.set_all_random_seed(0)
@@ -31,6 +37,10 @@
print("output dir: {}".format(self.cache_dir))
self.onnx = onnx
self.quant = quant
+ self.fallback_num = fallback_num
+ self.frontend = None
+ self.audio_in = audio_in
+ self.calib_num = calib_num
def _export(
@@ -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
@@ -70,16 +91,20 @@
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:
@@ -101,6 +126,39 @@
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',
@@ -127,6 +185,7 @@
model, asr_train_args = ASRTask.build_model_from_file(
asr_train_config, asr_model_file, cmvn_file, 'cpu'
)
+ self.frontend = model.frontend
self._export(model, tag_name)
@@ -153,26 +212,40 @@
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')
+ 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('--export-dir', type=str, required=True)
+ parser.add_argument('--type', type=str, default='onnx', help='["onnx", "torch"]')
+ 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')
+ args = parser.parse_args()
+
+ export_model = ASRModelExportParaformer(
+ cache_dir=args.export_dir,
+ onnx=args.type == 'onnx',
+ quant=args.quantize,
+ fallback_num=args.fallback_num,
+ audio_in=args.audio_in,
+ calib_num=args.calib_num,
+ )
+ export_model.export(args.model_name)
--
Gitblit v1.9.1