[Quantization] Add new PaddleClas models quantization support. (#864)

* 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
2025-10-13 20:34:02 +08:00 · 2022-12-13 10:21:44 +08:00
parent 7415552cb2
commit 5fc6cf30df
8 changed files with 207 additions and 2 deletions
--- a/examples/vision/detection/yolov6/quantize/cpp/README.md
+++ b/examples/vision/detection/yolov6/quantize/cpp/README.md
@@ -23,7 +23,7 @@ cmake .. -DFASTDEPLOY_INSTALL_DIR=${PWD}/fastdeploy-linux-x64-x.x.x
 make -j

 #下载FastDeloy提供的yolov6s量化模型文件和测试图片
-wget https://bj.bcebos.com/paddlehub/fastdeploy/yolov6s_qat_model.tar
+wget https://bj.bcebos.com/paddlehub/fastdeploy/yolov6s_qat_model_new.tar
 tar -xvf yolov6s_qat_model.tar
 wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg

--- a/examples/vision/detection/yolov6/quantize/python/README.md
+++ b/examples/vision/detection/yolov6/quantize/python/README.md
@@ -17,7 +17,7 @@ git clone https://github.com/PaddlePaddle/FastDeploy.git
 cd examples/slim/yolov6/python

 #下载FastDeloy提供的yolov6s量化模型文件和测试图片
-wget https://bj.bcebos.com/paddlehub/fastdeploy/yolov6s_qat_model.tar
+wget https://bj.bcebos.com/paddlehub/fastdeploy/yolov6s_qat_model_new.tar
 tar -xvf yolov6s_qat_model.tar
 wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg

--- a/tools/common_tools/auto_compression/README.md
+++ b/tools/common_tools/auto_compression/README.md
@@ -99,6 +99,11 @@ FastDeploy目前为用户提供了多个模型的压缩[config](./configs/)文
 | -------------------- | ------------------------------------------------------------ |----------------------------------------- |
 | [mobilenetv1_ssld_quant](./configs/classification/mobilenetv1_ssld_quant.yaml)      | [mobilenetv1_ssld](https://bj.bcebos.com/paddlehub/fastdeploy/MobileNetV1_ssld_infer.tgz)           |           |
 | [resnet50_vd_quant](./configs/classification/resnet50_vd_quant.yaml)      |   [resnet50_vd](https://bj.bcebos.com/paddlehub/fastdeploy/ResNet50_vd_infer.tgz)          |     |
+| [efficientnetb0_quant](./configs/classification/efficientnetb0_quant.yaml)      |   [efficientnetb0](https://bj.bcebos.com/paddlehub/fastdeploy/EfficientNetB0_small_infer.tgz)          |     |
+| [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)          |     |
+| [pphgnet_tiny_quant](./configs/classification/pphgnet_tiny_quant.yaml)      |   [pphgnet_tiny](https://bj.bcebos.com/paddlehub/fastdeploy/PPHGNet_tiny_ssld_infer.tgz)          |     |
+| [pplcnetv2_base_quant](./configs/classification/pplcnetv2_base_quant.yaml)      |   [pplcnetv2_base](https://bj.bcebos.com/paddlehub/fastdeploy/PPLCNetV2_base_infer.tgz)          |     |
+
 | [yolov5s_quant](./configs/detection/yolov5s_quant.yaml)       |   [yolov5s](https://paddle-slim-models.bj.bcebos.com/act/yolov5s.onnx)         |     |
 | [yolov6s_quant](./configs/detection/yolov6s_quant.yaml)       |  [yolov6s](https://paddle-slim-models.bj.bcebos.com/act/yolov6s.onnx)          |     |
 | [yolov7_quant](./configs/detection/yolov7_quant.yaml)        | [yolov7](https://paddle-slim-models.bj.bcebos.com/act/yolov7.onnx)           |      |
--- a/tools/common_tools/auto_compression/README_EN.md
+++ b/tools/common_tools/auto_compression/README_EN.md
@@ -107,6 +107,10 @@ FastDeploy currently provides users with compression [config](./configs/) files
 | -------------------- | ------------------------------------------------------------ |----------------------------------------- |
 | [mobilenetv1_ssld_quant](./configs/classification/mobilenetv1_ssld_quant.yaml)      | [mobilenetv1_ssld](https://bj.bcebos.com/paddlehub/fastdeploy/MobileNetV1_ssld_infer.tgz)           |           |
 | [resnet50_vd_quant](./configs/classification/resnet50_vd_quant.yaml)      |   [resnet50_vd](https://bj.bcebos.com/paddlehub/fastdeploy/ResNet50_vd_infer.tgz)          |     |
+| [efficientnetb0_quant](./configs/classification/efficientnetb0_quant.yaml)      |   [efficientnetb0](https://bj.bcebos.com/paddlehub/fastdeploy/EfficientNetB0_small_infer.tgz)          |     |
+| [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)          |     |
+| [pphgnet_tiny_quant](./configs/classification/pphgnet_tiny_quant.yaml)      |   [pphgnet_tiny](https://bj.bcebos.com/paddlehub/fastdeploy/PPHGNet_tiny_ssld_infer.tgz)          |     |
+| [pplcnetv2_base_quant](./configs/classification/pplcnetv2_base_quant.yaml)      |   [pplcnetv2_base](https://bj.bcebos.com/paddlehub/fastdeploy/PPLCNetV2_base_infer.tgz)          |     |
 | [yolov5s_quant](./configs/detection/yolov5s_quant.yaml)       |   [yolov5s](https://paddle-slim-models.bj.bcebos.com/act/yolov5s.onnx)         |     |
 | [yolov6s_quant](./configs/detection/yolov6s_quant.yaml)       |  [yolov6s](https://paddle-slim-models.bj.bcebos.com/act/yolov6s.onnx)          |     |
 | [yolov7_quant](./configs/detection/yolov7_quant.yaml)        | [yolov7](https://paddle-slim-models.bj.bcebos.com/act/yolov7.onnx)           |      |
--- a/tools/common_tools/auto_compression/configs/classification/efficientnetb0_quant.yaml
+++ b/tools/common_tools/auto_compression/configs/classification/efficientnetb0_quant.yaml
@@ -0,0 +1,50 @@
+Global:
+  model_dir: ./EfficientNetB0_small_infer/
+  format: 'paddle'
+  model_filename: inference.pdmodel
+  params_filename: inference.pdiparams
+  qat_image_path: ./ImageNet_val_640
+  ptq_image_path: ./ImageNet_val_640
+  input_list: ['inputs']
+  qat_preprocess: cls_image_preprocess
+  ptq_preprocess: cls_image_preprocess
+  qat_batch_size: 32
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - softmax_0.tmp_0
+
+QuantAware:
+  use_pact: true
+  activation_bits: 8
+  is_full_quantize: false
+  onnx_format: True
+  activation_quantize_type: moving_average_abs_max
+  weight_quantize_type: channel_wise_abs_max
+  not_quant_pattern:
+  - skip_quant
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+  - matmul
+  - matmul_v2
+  weight_bits: 8
+
+TrainConfig:
+  epochs: 1
+  eval_iter: 500
+  learning_rate:
+    type: CosineAnnealingDecay
+    learning_rate: 0.015
+    T_max: 8000
+  optimizer_builder:
+    optimizer:
+      type: Momentum
+    weight_decay: 0.00002
+  origin_metric: 0.7738
+
+PTQ:
+  calibration_method: 'avg'   # option: avg, abs_max, hist, KL, mse
+  skip_tensor_list: None
--- a/tools/common_tools/auto_compression/configs/classification/mobilenetv3_large_x1_0_quant.yaml
+++ b/tools/common_tools/auto_compression/configs/classification/mobilenetv3_large_x1_0_quant.yaml
@@ -0,0 +1,46 @@
+Global:
+  model_dir: ./MobileNetV3_large_x1_0_ssld_infer/
+  format: 'paddle'
+  model_filename: inference.pdmodel
+  params_filename: inference.pdiparams
+  qat_image_path: ./ImageNet_val_640
+  ptq_image_path: ./ImageNet_val_640
+  input_list: ['inputs']
+  qat_preprocess: cls_image_preprocess
+  ptq_preprocess: cls_image_preprocess
+  qat_batch_size: 128
+
+
+Distillation:
+  alpha: 1.0
+  loss: soft_label
+
+QuantAware:
+  use_pact: true
+  activation_bits: 8
+  is_full_quantize: false
+  onnx_format: True
+  activation_quantize_type: moving_average_abs_max
+  weight_quantize_type: channel_wise_abs_max
+  not_quant_pattern:
+  - skip_quant
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+  - matmul
+  - matmul_v2
+  weight_bits: 8
+
+TrainConfig:
+  epochs: 2
+  eval_iter: 5000
+  learning_rate: 0.001
+  optimizer_builder:
+    optimizer:
+      type: Momentum
+    weight_decay: 0.00002
+  origin_metric: 0.7896
+
+PTQ:
+  calibration_method: 'avg'   # option: avg, abs_max, hist, KL, mse
+  skip_tensor_list: None
--- a/tools/common_tools/auto_compression/configs/classification/pphgnet_tiny_quant.yaml
+++ b/tools/common_tools/auto_compression/configs/classification/pphgnet_tiny_quant.yaml
@@ -0,0 +1,50 @@
+Global:
+  model_dir: ./PPHGNet_tiny_ssld_infer/
+  format: 'paddle'
+  model_filename: inference.pdmodel
+  params_filename: inference.pdiparams
+  qat_image_path: ./ImageNet_val_640
+  ptq_image_path: ./ImageNet_val_640
+  input_list: ['x']
+  qat_preprocess: cls_image_preprocess
+  ptq_preprocess: cls_image_preprocess
+  qat_batch_size: 32
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - softmax_1.tmp_0
+
+QuantAware:
+  use_pact: true
+  activation_bits: 8
+  is_full_quantize: false
+  onnx_format: True
+  activation_quantize_type: moving_average_abs_max
+  weight_quantize_type: channel_wise_abs_max
+  not_quant_pattern:
+  - skip_quant
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+  - matmul
+  - matmul_v2
+  weight_bits: 8
+
+TrainConfig:
+  epochs: 1
+  eval_iter: 500
+  learning_rate:
+    type: CosineAnnealingDecay
+    learning_rate: 0.015
+    T_max: 8000
+  optimizer_builder:
+    optimizer:
+      type: Momentum
+    weight_decay: 0.00002
+  origin_metric: 0.7959
+
+PTQ:
+  calibration_method: 'avg'   # option: avg, abs_max, hist, KL, mse
+  skip_tensor_list: None
--- a/tools/common_tools/auto_compression/configs/classification/pplcnetv2_base_quant.yaml
+++ b/tools/common_tools/auto_compression/configs/classification/pplcnetv2_base_quant.yaml
@@ -0,0 +1,50 @@
+Global:
+  model_dir: ./PPLCNetV2_base_infer/
+  format: 'paddle'
+  model_filename: inference.pdmodel
+  params_filename: inference.pdiparams
+  qat_image_path: ./ImageNet_val_640
+  ptq_image_path: ./ImageNet_val_640
+  input_list: ['x']
+  qat_preprocess: cls_image_preprocess
+  ptq_preprocess: cls_image_preprocess
+  qat_batch_size: 32
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - softmax_1.tmp_0
+
+QuantAware:
+  use_pact: true
+  activation_bits: 8
+  is_full_quantize: false
+  onnx_format: True
+  activation_quantize_type: moving_average_abs_max
+  weight_quantize_type: channel_wise_abs_max
+  not_quant_pattern:
+  - skip_quant
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+  - matmul
+  - matmul_v2
+  weight_bits: 8
+
+TrainConfig:
+  epochs: 1
+  eval_iter: 500
+  learning_rate:
+    type: CosineAnnealingDecay
+    learning_rate: 0.015
+    T_max: 8000
+  optimizer_builder:
+    optimizer:
+      type: Momentum
+    weight_decay: 0.00002
+  origin_metric: 0.7704
+
+PTQ:
+  calibration_method: 'avg'   # option: avg, abs_max, hist, KL, mse
+  skip_tensor_list: None