From 69ccdd35cda4c8482e189fa350fbcb83997872f2 Mon Sep 17 00:00:00 2001
From: wanchen.swc <wanchen.swc@alibaba-inc.com>
Date: 星期一, 06 三月 2023 18:18:31 +0800
Subject: [PATCH] [Quantization] model quantization for inference
---
funasr/export/export_model.py | 83 ++++++++++++++++++++++++++++++++++-------
1 files changed, 69 insertions(+), 14 deletions(-)
diff --git a/funasr/export/export_model.py b/funasr/export/export_model.py
index e15390b..1c677c9 100644
--- a/funasr/export/export_model.py
+++ b/funasr/export/export_model.py
@@ -7,13 +7,17 @@
import logging
import torch
-from funasr.bin.asr_inference_paraformer import Speech2Text
from funasr.export.models import get_model
import numpy as np
import random
+# 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):
+ def __init__(
+ self, cache_dir: Union[Path, str] = None, onnx: bool = True, quant: bool = True
+ ):
assert check_argument_types()
self.set_all_random_seed(0)
if cache_dir is None:
@@ -24,13 +28,14 @@
feats_dim=560,
onnx=False,
)
- logging.info("output dir: {}".format(self.cache_dir))
+ print("output dir: {}".format(self.cache_dir))
self.onnx = onnx
+ self.quant = quant
def _export(
self,
- model: Speech2Text,
+ model,
tag_name: str = None,
verbose: bool = False,
):
@@ -44,24 +49,53 @@
model,
self.export_config,
)
+ 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 _torch_quantize(self, model):
+ 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
+ module_filter = ModuleFilter(include_classes=[EncoderLayerSANM, DecoderLayerSANM])
+ module_filter.exclude_op_types = [torch.nn.Conv1d]
+ 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)
+ 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_txt()
+ dummy_input = model.get_dummy_inputs()
# 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'))
+
+ if self.quant:
+ quant_model = self._torch_quantize(model)
+ model_script = torch.jit.trace(quant_model, dummy_input)
+ model_script.save(os.path.join(path, f'{model.model_name}_quant.torchscripts'))
+
def set_all_random_seed(self, seed: int):
random.seed(seed)
@@ -85,9 +119,9 @@
with open(json_file, 'r') as f:
config_data = json.load(f)
mode = config_data['model']['model_config']['mode']
- if mode == 'paraformer':
+ if mode.startswith('paraformer'):
from funasr.tasks.asr import ASRTaskParaformer as ASRTask
- elif mode == 'uniasr':
+ elif mode.startswith('uniasr'):
from funasr.tasks.asr import ASRTaskUniASR as ASRTask
model, asr_train_args = ASRTask.build_model_from_file(
@@ -104,20 +138,41 @@
# 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,
- os.path.join(path, f'{model.model_name}.onnx'),
+ model_path,
verbose=verbose,
- opset_version=12,
+ 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
+ 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 __name__ == '__main__':
- output_dir = "../export"
- export_model = ASRModelExportParaformer(cache_dir=output_dir, onnx=True)
- 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')
\ No newline at end of file
+ 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')
--
Gitblit v1.9.1