Add Quantization Function. (#256)

* Add PaddleOCR Support * Add PaddleOCR Support * Add PaddleOCRv3 Support * Add PaddleOCRv3 Support * Update README.md * Update README.md * Update README.md * Update README.md * Add PaddleOCRv3 Support * Add PaddleOCRv3 Supports * Add PaddleOCRv3 Suport * Fix Rec diff * Remove useless functions * Remove useless comments * Add PaddleOCRv2 Support * Add PaddleOCRv3 & PaddleOCRv2 Support * remove useless parameters * Add utils of sorting det boxes * Fix code naming convention * Fix code naming convention * Fix code naming convention * Fix bug in the Classify process * Imporve OCR Readme * Fix diff in Cls model * Update Model Download Link in Readme * Fix diff in PPOCRv2 * Improve OCR readme * Imporve OCR readme * Improve OCR readme * Improve OCR readme * Imporve OCR readme * Improve OCR readme * Fix conflict * Add readme for OCRResult * Improve OCR readme * Add OCRResult readme * Improve OCR readme * Improve OCR readme * Add Model Quantization Demo * Fix Model Quantization Readme * Fix Model Quantization Readme * Add the function to do PTQ quantization * Improve quant tools readme * Improve quant tool readme * Improve quant tool readme * Add PaddleInference-GPU for OCR Rec model * Add QAT method to fastdeploy-quantization tool * Remove examples/slim for now * Move configs folder * Add Quantization Support for Classification Model * Imporve ways of importing preprocess * Upload YOLO Benchmark on readme * Upload YOLO Benchmark on readme * Upload YOLO Benchmark on readme * Improve Quantization configs and readme * Add support for multi-inputs model
2025-12-24 13:28:13 +08:00 · 2022-10-08 15:45:28 +08:00
parent 8d47637541
commit 1efc0fa6b0
13 changed files with 736 additions and 0 deletions
--- a/tools/quantization/configs/classification/mobilenetv1_ssld_quant.yaml
+++ b/tools/quantization/configs/classification/mobilenetv1_ssld_quant.yaml
@@ -0,0 +1,49 @@
+Global:
+  model_dir: ./MobileNetV1_ssld_infer/
+  format: 'paddle'
+  model_filename: inference.pdmodel
+  params_filename: inference.pdiparams
+  image_path: ./ImageNet_val_640
+  arch: MobileNetV1
+  input_list: ['input']
+  preprocess: cls_image_preprocess
+
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - softmax_0.tmp_0
+
+
+Quantization:
+  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
+  weight_bits: 8
+
+
+TrainConfig:
+  train_iter: 5000
+  learning_rate:
+    type: CosineAnnealingDecay
+    learning_rate: 0.015
+    T_max: 8000
+  optimizer_builder:
+    optimizer:
+      type: Momentum
+    weight_decay: 0.00002
+  origin_metric: 0.70898
+
+
+PTQ:
+  calibration_method: 'avg'   # option: avg, abs_max, hist, KL, mse
+  skip_tensor_list: None
--- a/tools/quantization/configs/classification/resnet50_vd_quant.yaml
+++ b/tools/quantization/configs/classification/resnet50_vd_quant.yaml
@@ -0,0 +1,47 @@
+Global:
+  model_dir: ./ResNet50_vd_infer/
+  format: 'paddle'
+  model_filename: inference.pdmodel
+  params_filename: inference.pdiparams
+  image_path: ./ImageNet_val_640
+  arch: ResNet50
+  input_list: ['input']
+  preprocess: cls_image_preprocess
+
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - softmax_0.tmp_0
+
+Quantization:
+  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
+  weight_bits: 8
+
+TrainConfig:
+  train_iter: 5000
+  learning_rate:
+    type: CosineAnnealingDecay
+    learning_rate: 0.015
+    T_max: 8000
+  optimizer_builder:
+    optimizer:
+      type: Momentum
+    weight_decay: 0.00002
+  origin_metric: 0.7912
+
+
+PTQ:
+  calibration_method: 'avg'   # option: avg, abs_max, hist, KL, mse
+  skip_tensor_list: None
--- a/tools/quantization/configs/detection/yolov5s_quant.yaml
+++ b/tools/quantization/configs/detection/yolov5s_quant.yaml
@@ -0,0 +1,35 @@
+Global:
+  model_dir: ./yolov5s.onnx
+  format: 'onnx'
+  model_filename: model.pdmodel
+  params_filename: model.pdiparams
+  image_path: ./COCO_val_320
+  arch: YOLOv5
+  input_list: ['x2paddle_images']
+  preprocess: yolo_image_preprocess
+
+Distillation:
+  alpha: 1.0
+  loss: soft_label
+
+Quantization:
+  onnx_format: true
+  use_pact: true
+  activation_quantize_type: 'moving_average_abs_max'
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+
+
+PTQ:
+  calibration_method: 'avg'   # option: avg, abs_max, hist, KL, mse
+  skip_tensor_list: None
+
+TrainConfig:
+  train_iter: 3000
+  learning_rate: 0.00001
+  optimizer_builder:
+    optimizer:
+      type: SGD
+    weight_decay: 4.0e-05
+  target_metric: 0.365
--- a/tools/quantization/configs/detection/yolov6s_quant.yaml
+++ b/tools/quantization/configs/detection/yolov6s_quant.yaml
@@ -0,0 +1,35 @@
+Global:
+  model_dir: ./yolov6s.onnx
+  format: 'onnx'
+  model_filename: model.pdmodel
+  params_filename: model.pdiparams
+  image_path: ./COCO_val_320
+  arch: YOLOv6
+  input_list: ['x2paddle_image_arrays']
+
+Distillation:
+  alpha: 1.0
+  loss: soft_label
+
+Quantization:
+  onnx_format: true
+  activation_quantize_type: 'moving_average_abs_max'
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+
+
+PTQ:
+  calibration_method: 'avg'   # option: avg, abs_max, hist, KL, mse
+  skip_tensor_list:  ['conv2d_2.w_0', 'conv2d_15.w_0', 'conv2d_46.w_0', 'conv2d_11.w_0', 'conv2d_49.w_0']
+
+TrainConfig:
+  train_iter: 8000
+  learning_rate:
+    type: CosineAnnealingDecay
+    learning_rate: 0.00003
+    T_max: 8000
+  optimizer_builder:
+    optimizer:
+      type: SGD
+    weight_decay: 0.00004
--- a/tools/quantization/configs/detection/yolov7_quant.yaml
+++ b/tools/quantization/configs/detection/yolov7_quant.yaml
@@ -0,0 +1,34 @@
+Global:
+  model_dir: ./yolov7.onnx
+  format: 'onnx'
+  model_filename: model.pdmodel
+  params_filename: model.pdiparams
+  image_path: ./COCO_val_320
+  arch: YOLOv7
+  input_list: ['x2paddle_images']
+
+Distillation:
+  alpha: 1.0
+  loss: soft_label
+
+Quantization:
+  onnx_format: true
+  activation_quantize_type: 'moving_average_abs_max'
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+
+PTQ:
+  calibration_method: 'avg'   # option: avg, abs_max, hist, KL, mse
+  skip_tensor_list: None
+
+TrainConfig:
+  train_iter: 3000
+  learning_rate:
+    type: CosineAnnealingDecay
+    learning_rate: 0.00003
+    T_max: 8000
+  optimizer_builder:
+    optimizer:
+      type: SGD
+    weight_decay: 0.00004
--- a/tools/quantization/configs/readme.md
+++ b/tools/quantization/configs/readme.md
@@ -0,0 +1,48 @@
+# FastDeploy 量化配置文件说明
+FastDeploy 量化配置文件中，包含了全局配置，量化蒸馏训练配置，离线量化配置和训练配置.
+用户除了直接使用FastDeploy提供在本目录的配置文件外，可以按需求自行修改相关配置文件
+
+## 实例解读
+
+```
+#全局信息
+Global:
+  model_dir: ./yolov7-tiny.onnx     #输入模型路径
+  format: 'onnx'                    #输入模型格式，选项为 onnx 或者 paddle
+  model_filename: model.pdmodel     #paddle模型的模型文件名
+  params_filename: model.pdiparams  #paddle模型的参数文件名
+  image_path: ./COCO_val_320        #PTQ所有的Calibration数据集或者量化训练所用的训练集
+  arch: YOLOv7                      #模型系列
+
+#量化蒸馏训练中的蒸馏参数设置
+Distillation:
+  alpha: 1.0
+  loss: soft_label
+
+#量化蒸馏训练中的量化参数设置
+Quantization:
+  onnx_format: true
+  activation_quantize_type: 'moving_average_abs_max'
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+
+#离线量化参数配置
+PTQ:
+  calibration_method: 'avg' #Calibraion算法，可选为 avg, abs_max, hist, KL, mse
+  skip_tensor_list: None    #不进行离线量化的tensor
+
+
+#训练参数
+TrainConfig:
+  train_iter: 3000  
+  learning_rate:
+    type: CosineAnnealingDecay
+    learning_rate: 0.00003
+    T_max: 8000
+  optimizer_builder:
+    optimizer:
+      type: SGD
+    weight_decay: 0.00004
+
+```