From 4e44c9f46e550eab4ec6b70c099dcdae44eb9d61 Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 23 三月 2023 20:34:21 +0800
Subject: [PATCH] Merge pull request #288 from alibaba-damo-academy/dev_gzf
---
funasr/export/export_model.py | 62 ++++++++++++++++++++++++++++---
1 files changed, 56 insertions(+), 6 deletions(-)
diff --git a/funasr/export/export_model.py b/funasr/export/export_model.py
index beb1efe..b1161cb 100644
--- a/funasr/export/export_model.py
+++ b/funasr/export/export_model.py
@@ -10,17 +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:
+class ModelExport:
def __init__(
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)
@@ -36,6 +38,9 @@
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(
@@ -67,8 +72,15 @@
def _torch_quantize(self, model):
def _run_calibration_data(m):
# using dummy inputs for a example
- dummy_input = model.get_dummy_inputs()
- m(*dummy_input)
+ 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.quantizer import Backend, Quantizer
@@ -114,6 +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',
@@ -140,6 +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)
@@ -188,14 +234,18 @@
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', action='store_true', 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('--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(
+ export_model = ModelExport(
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