Skip to content

Commit 2d3d941

Browse files
yunyaoXYYyunyaoXYY
authored
[Quantization] Add new PaddleSeg models quantization support. (#953)
* Fix links in readme * Fix links in readme * Update PPOCRv2/v3 examples * Update auto compression configs * Add neww quantization support for paddleclas model * Update quantized Yolov6s model download link * Improve PPOCR comments * Add English doc for quantization * Fix PPOCR rec model bug * Add new paddleseg quantization support * Add new paddleseg quantization support * Add new paddleseg quantization support * Add new paddleseg quantization support
1 parent 2431890 commit 2d3d941

File tree

5 files changed

+120
-3
lines changed

5 files changed

+120
-3
lines changed

‎tools/common_tools/auto_compression/README.md

+4-2
Original file line numberDiff line numberDiff line change
@@ -103,13 +103,15 @@ FastDeploy目前为用户提供了多个模型的压缩[config](./configs/)文
103103
| [mobilenetv3_large_x1_0_quant](./configs/classification/mobilenetv3_large_x1_0_quant.yaml) | [mobilenetv3_large_x1_0](https://bj.bcebos.com/paddlehub/fastdeploy/MobileNetV3_large_x1_0_ssld_infer.tgz) | |
104104
| [pphgnet_tiny_quant](./configs/classification/pphgnet_tiny_quant.yaml) | [pphgnet_tiny](https://bj.bcebos.com/paddlehub/fastdeploy/PPHGNet_tiny_ssld_infer.tgz) | |
105105
| [pplcnetv2_base_quant](./configs/classification/pplcnetv2_base_quant.yaml) | [pplcnetv2_base](https://bj.bcebos.com/paddlehub/fastdeploy/PPLCNetV2_base_infer.tgz) | |
106-
107106
| [yolov5s_quant](./configs/detection/yolov5s_quant.yaml) | [yolov5s](https://paddle-slim-models.bj.bcebos.com/act/yolov5s.onnx) | |
108107
| [yolov6s_quant](./configs/detection/yolov6s_quant.yaml) | [yolov6s](https://paddle-slim-models.bj.bcebos.com/act/yolov6s.onnx) | |
109108
| [yolov7_quant](./configs/detection/yolov7_quant.yaml) | [yolov7](https://paddle-slim-models.bj.bcebos.com/act/yolov7.onnx) | |
110109
| [ppyoloe_withNMS_quant](./configs/detection/ppyoloe_withNMS_quant.yaml) | [ppyoloe_l](https://bj.bcebos.com/v1/paddle-slim-models/act/ppyoloe_crn_l_300e_coco.tar) | 支持PPYOLOE的s,m,l,x系列模型, 从PaddleDetection导出模型时正常导出, 不要去除NMS |
111110
| [ppyoloe_plus_withNMS_quant](./configs/detection/ppyoloe_plus_withNMS_quant.yaml) | [ppyoloe_plus_s](https://bj.bcebos.com/paddlehub/fastdeploy/ppyoloe_plus_crn_s_80e_coco.tar) | 支持PPYOLOE+的s,m,l,x系列模型, 从PaddleDetection导出模型时正常导出, 不要去除NMS |
112-
| [pp_liteseg_quant](./configs/segmentation/pp_liteseg_quant.yaml) | [pp_liteseg](https://bj.bcebos.com/paddlehub/fastdeploy/PP_LiteSeg_T_STDC1_cityscapes_without_argmax_infer.tgz) | |
111+
| [pp_liteseg_quant](./configs/segmentation/pp_liteseg_quant.yaml) | [pp_liteseg](https://bj.bcebos.com/paddlehub/fastdeploy/PP_LiteSeg_T_STDC1_cityscapes_without_argmax_infer.tgz) |
112+
| [deeplabv3_resnet_quant](./configs/segmentation/deeplabv3_resnet_quant.yaml) | [deeplabv3_resnet101](https://bj.bcebos.com/paddlehub/fastdeploy/Deeplabv3_ResNet101_OS8_cityscapes_without_argmax_infer.tgz) | |
113+
| [fcn_hrnet_quant](./configs/segmentation/fcn_hrnet_quant.yaml) | [fcn_hrnet](https://bj.bcebos.com/paddlehub/fastdeploy/FCN_HRNet_W18_cityscapes_without_argmax_infer.tgz) | |
114+
| [unet_quant](./configs/segmentation/unet_quant.yaml) | [unet](https://bj.bcebos.com/paddlehub/fastdeploy/Unet_cityscapes_without_argmax_infer.tgz) | | |
113115

114116

115117

‎tools/common_tools/auto_compression/README_EN.md

+5-1
Original file line numberDiff line numberDiff line change
@@ -103,7 +103,7 @@ To complete the quantization, developers only need to provide a customized model
103103
## 3. FastDeploy One-Click Model Auto Compression Config file examples
104104
FastDeploy currently provides users with compression [config](./configs/) files of multiple models, and the corresponding FP32 model, Users can directly download and experience it.
105105

106-
| Config文件 | 待压缩的FP32模型 | 备注 |
106+
| Config file | FP32 model | Note |
107107
| -------------------- | ------------------------------------------------------------ |----------------------------------------- |
108108
| [mobilenetv1_ssld_quant](./configs/classification/mobilenetv1_ssld_quant.yaml) | [mobilenetv1_ssld](https://bj.bcebos.com/paddlehub/fastdeploy/MobileNetV1_ssld_infer.tgz) | |
109109
| [resnet50_vd_quant](./configs/classification/resnet50_vd_quant.yaml) | [resnet50_vd](https://bj.bcebos.com/paddlehub/fastdeploy/ResNet50_vd_infer.tgz) | |
@@ -117,6 +117,10 @@ FastDeploy currently provides users with compression [config](./configs/) files
117117
| [ppyoloe_withNMS_quant](./configs/detection/ppyoloe_withNMS_quant.yaml) | [ppyoloe_l](https://bj.bcebos.com/v1/paddle-slim-models/act/ppyoloe_crn_l_300e_coco.tar) | Support PPYOLOE's s,m,l,x series models, export the model normally when exporting the model from PaddleDetection, do not remove NMS |
118118
| [ppyoloe_plus_withNMS_quant](./configs/detection/ppyoloe_plus_withNMS_quant.yaml) | [ppyoloe_plus_s](https://bj.bcebos.com/paddlehub/fastdeploy/ppyoloe_plus_crn_s_80e_coco.tar) | Support PPYOLOE+'s s,m,l,x series models, export the model normally when exporting the model from PaddleDetection, do not remove NMS |
119119
| [pp_liteseg_quant](./configs/segmentation/pp_liteseg_quant.yaml) | [pp_liteseg](https://bj.bcebos.com/paddlehub/fastdeploy/PP_LiteSeg_T_STDC1_cityscapes_without_argmax_infer.tgz) | |
120+
| [deeplabv3_resnet_quant](./configs/segmentation/deeplabv3_resnet_quant.yaml) | [deeplabv3_resnet101](https://bj.bcebos.com/paddlehub/fastdeploy/Deeplabv3_ResNet101_OS8_cityscapes_without_argmax_infer.tgz) | |
121+
| [fcn_hrnet_quant](./configs/segmentation/fcn_hrnet_quant.yaml) | [fcn_hrnet](https://bj.bcebos.com/paddlehub/fastdeploy/FCN_HRNet_W18_cityscapes_without_argmax_infer.tgz) | |
122+
| [unet_quant](./configs/segmentation/unet_quant.yaml) | [unet](https://bj.bcebos.com/paddlehub/fastdeploy/Unet_cityscapes_without_argmax_infer.tgz) | | |
123+
120124

121125
## 3. Deploy quantized models on FastDeploy
122126

‎tools/common_tools/auto_compression/configs/segmentation/deeplabv3_resnet_quant.yaml

+37
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,37 @@
1+
Global:
2+
model_dir: ./Deeplabv3_ResNet101_OS8_cityscapes_without_argmax_infer/
3+
format: paddle
4+
model_filename: model.pdmodel
5+
params_filename: model.pdiparams
6+
qat_image_path: ./train_stuttgart
7+
ptq_image_path: ./train_stuttgart
8+
input_list: ['x']
9+
qat_preprocess: ppseg_cityscapes_qat_preprocess
10+
ptq_preprocess: ppseg_cityscapes_ptq_preprocess
11+
qat_batch_size: 2
12+
13+
14+
Distillation:
15+
alpha: 1.0
16+
loss: l2
17+
node:
18+
- conv2d_225.tmp_0
19+
20+
QuantAware:
21+
onnx_format: True
22+
quantize_op_types:
23+
- conv2d
24+
- depthwise_conv2d
25+
26+
TrainConfig:
27+
epochs: 1
28+
eval_iter: 360
29+
learning_rate: 0.0001
30+
optimizer_builder:
31+
optimizer:
32+
type: SGD
33+
weight_decay: 0.0005
34+
35+
PTQ:
36+
calibration_method: 'avg' # option: avg, abs_max, hist, KL, mse
37+
skip_tensor_list: None

‎tools/common_tools/auto_compression/configs/segmentation/fcn_hrnet_quant.yaml

+36
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,36 @@
1+
Global:
2+
model_dir: ./FCN_HRNet_W18_cityscapes_without_argmax_infer
3+
format: paddle
4+
model_filename: model.pdmodel
5+
params_filename: model.pdiparams
6+
qat_image_path: ./train_stuttgart
7+
ptq_image_path: ./train_stuttgart
8+
input_list: ['x']
9+
qat_preprocess: ppseg_cityscapes_qat_preprocess
10+
ptq_preprocess: ppseg_cityscapes_ptq_preprocess
11+
qat_batch_size: 2
12+
13+
Distillation:
14+
alpha: 1.0
15+
loss: l2
16+
node:
17+
- conv2d_613.tmp_1
18+
19+
QuantAware:
20+
onnx_format: True
21+
quantize_op_types:
22+
- conv2d
23+
- depthwise_conv2d
24+
25+
TrainConfig:
26+
epochs: 20
27+
eval_iter: 360
28+
learning_rate: 0.0001
29+
optimizer_builder:
30+
optimizer:
31+
type: SGD
32+
weight_decay: 4.0e-05
33+
34+
PTQ:
35+
calibration_method: 'avg' # option: avg, abs_max, hist, KL, mse
36+
skip_tensor_list: None

‎tools/common_tools/auto_compression/configs/segmentation/unet_quant.yaml

+38
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,38 @@
1+
Global:
2+
model_dir: ./Unet_cityscapes_without_argmax_infer/
3+
format: paddle
4+
model_filename: model.pdmodel
5+
params_filename: model.pdiparams
6+
qat_image_path: ./train_stuttgart
7+
ptq_image_path: ./train_stuttgart
8+
input_list: ['x']
9+
qat_preprocess: ppseg_cityscapes_qat_preprocess
10+
ptq_preprocess: ppseg_cityscapes_ptq_preprocess
11+
qat_batch_size: 2
12+
13+
14+
Distillation:
15+
alpha: 1.0
16+
loss: l2
17+
node:
18+
- conv2d_37.tmp_1
19+
20+
QuantAware:
21+
onnx_format: True
22+
quantize_op_types:
23+
- conv2d
24+
- depthwise_conv2d
25+
26+
27+
TrainConfig:
28+
epochs: 10
29+
eval_iter: 180
30+
learning_rate: 0.0005
31+
optimizer_builder:
32+
optimizer:
33+
type: SGD
34+
weight_decay: 4.0e-05
35+
36+
PTQ:
37+
calibration_method: 'avg' # option: avg, abs_max, hist, KL, mse
38+
skip_tensor_list: None

0 commit comments

Comments
 (0)