mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2026-04-23 17:11:21 +08:00
[Quantization] Improve the usage of FastDeploy tools. (#660)
* 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 * Add backends and params file for YOLOv7 * Add quantized model deployment support for YOLO series * Fix YOLOv5 quantize readme * Fix YOLO quantize readme * Fix YOLO quantize readme * Improve quantize YOLO readme * Improve quantize YOLO readme * Improve quantize YOLO readme * Improve quantize YOLO readme * Improve quantize YOLO readme * Fix bug, change Fronted to ModelFormat * Change Fronted to ModelFormat * Add examples to deploy quantized paddleclas models * Fix readme * Add quantize Readme * Add quantize Readme * Add quantize Readme * Modify readme of quantization tools * Modify readme of quantization tools * Improve quantization tools readme * Improve quantization readme * Improve PaddleClas quantized model deployment readme * Add PPYOLOE-l quantized deployment examples * Improve quantization tools readme * Improve Quantize Readme * Fix conflicts * Fix conflicts * improve readme * Improve quantization tools and readme * Improve quantization tools and readme * Add quantized deployment examples for PaddleSeg model * Fix cpp readme * Fix memory leak of reader_wrapper function * Fix model file name in PaddleClas quantization examples * Update Runtime and E2E benchmark * Update Runtime and E2E benchmark * Rename quantization tools to auto compression tools * Remove PPYOLOE data when deployed on MKLDNN * Fix readme * Support PPYOLOE with OR without NMS and update readme * Update Readme * Update configs and readme * Update configs and readme * Add Paddle-TensorRT backend in quantized model deploy examples * Support PPYOLOE+ series * Add reused_input_tensors for PPYOLOE * Improve fastdeploy tools usage * improve fastdeploy tool * Improve fastdeploy auto compression tool * Improve fastdeploy auto compression tool * Improve fastdeploy auto compression tool * Improve fastdeploy auto compression tool * Improve fastdeploy auto compression tool * remove modify * Improve fastdeploy auto compression tool * Improve fastdeploy auto compression tool * Improve fastdeploy auto compression tool * Improve fastdeploy auto compression tool * Improve fastdeploy auto compression tool * Remove extra requirements for fd-auto-compress package * Imporve fastdeploy-tools package * Install fastdeploy-tools package when build fastdeploy-python * Imporve quantization readme
This commit is contained in:
yunyaoXYY
@@ -0,0 +1,54 @@
|
||||
# FastDeploy 一键自动化压缩配置文件说明
|
||||
FastDeploy 一键自动化压缩配置文件中，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，可以按照以下示例,自行修改相关配置文件, 来尝试压缩自己的模型.
|
||||
|
||||
## 实例解读
|
||||
|
||||
```
|
||||
# 全局配置
|
||||
Global:
|
||||
model_dir: ./ppyoloe_plus_crn_s_80e_coco #输入模型的路径, 用户若需量化自己的模型，替换此处即可
|
||||
format: paddle #输入模型的格式, paddle模型请选择'paddle', onnx模型选择'onnx'
|
||||
model_filename: model.pdmodel #量化后转为paddle格式模型的模型名字
|
||||
params_filename: model.pdiparams #量化后转为paddle格式模型的参数名字
|
||||
qat_image_path: ./COCO_train_320 #量化蒸馏训练使用的数据集,此例为少量无标签数据, 选自COCO2017训练集中的前320张图片, 做少量数据训练
|
||||
ptq_image_path: ./COCO_val_320 #离线训练使用的Carlibration数据集, 选自COCO2017验证集中的前320张图片.
|
||||
input_list: ['image','scale_factor'] #待量化的模型的输入名字
|
||||
qat_preprocess: ppyoloe_plus_withNMS_image_preprocess #模型量化蒸馏训练时,对数据做的预处理函数, 用户可以在 ../fdquant/dataset.py 中修改或自行编写新的预处理函数, 来支自定义模型的量化
|
||||
ptq_preprocess: ppyoloe_plus_withNMS_image_preprocess #模型离线量化时,对数据做的预处理函数, 用户可以在 ../fdquant/dataset.py 中修改或自行编写新的预处理函数, 来支自定义模型的量化
|
||||
qat_batch_size: 4 #量化蒸馏训练时的batch_size, 若为onnx格式的模型,此处只能为1
|
||||
|
||||
|
||||
#量化蒸馏训练配置
|
||||
Distillation:
|
||||
alpha: 1.0 #蒸馏loss所占权重
|
||||
loss: soft_label #蒸馏loss算法
|
||||
|
||||
Quantization:
|
||||
onnx_format: true #是否采用ONNX量化标准格式, 要在FastDeploy上部署, 必须选true
|
||||
use_pact: true #量化训练是否使用PACT方法
|
||||
activation_quantize_type: 'moving_average_abs_max' #激活量化方式
|
||||
quantize_op_types: #需要进行量化的OP
|
||||
- conv2d
|
||||
- depthwise_conv2d
|
||||
|
||||
#离线量化配置
|
||||
PTQ:
|
||||
calibration_method: 'avg' #离线量化的激活校准算法, 可选: avg, abs_max, hist, KL, mse, emd
|
||||
skip_tensor_list: None #用户可指定跳过某些conv层,不进行量化
|
||||
|
||||
#训练参数配置
|
||||
TrainConfig:
|
||||
train_iter: 3000
|
||||
learning_rate: 0.00001
|
||||
optimizer_builder:
|
||||
optimizer:
|
||||
type: SGD
|
||||
weight_decay: 4.0e-05
|
||||
target_metric: 0.365
|
||||
|
||||
```
|
||||
## 更多详细配置方法
|
||||
|
||||
FastDeploy一键压缩功能由PaddeSlim助力, 更详细的量化配置方法请参考:
|
||||
[自动化压缩超参详细教程](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/example/auto_compression/hyperparameter_tutorial.md)
|
||||
@@ -0,0 +1,56 @@
|
||||
# Quantization Config File on FastDeploy
|
||||
|
||||
The FastDeploy quantization configuration file contains global configuration, quantization distillation training configuration, post-training quantization configuration and training configuration.
|
||||
In addition to using the configuration files provided by FastDeploy directly in this directory, users can modify the relevant configuration files according to their needs
|
||||
|
||||
## Demo
|
||||
|
||||
|
||||
```
|
||||
# Global config
|
||||
Global:
|
||||
model_dir: ./ppyoloe_plus_crn_s_80e_coco #Path to input model
|
||||
format: paddle #Input model format, please select 'paddle' for paddle model
|
||||
model_filename: model.pdmodel #Quantized model name in Paddle format
|
||||
params_filename: model.pdiparams #Parameter name for quantized paddle model
|
||||
qat_image_path: ./COCO_train_320 #Data set paths for quantization distillation training
|
||||
ptq_image_path: ./COCO_val_320 #Data set paths for PTQ
|
||||
input_list: ['image','scale_factor'] #Input name of the model to be quanzitzed
|
||||
qat_preprocess: ppyoloe_plus_withNMS_image_preprocess # The preprocessing function for Quantization distillation training
|
||||
ptq_preprocess: ppyoloe_plus_withNMS_image_preprocess # The preprocessing function for PTQ
|
||||
qat_batch_size: 4 #Batch size
|
||||
|
||||
|
||||
# Quantization distillation training configuration
|
||||
Distillation:
|
||||
alpha: 1.0 #Distillation loss weight
|
||||
loss: soft_label #Distillation loss algorithm
|
||||
|
||||
Quantization:
|
||||
onnx_format: true #Whether to use ONNX quantization standard format or not, must be true to deploy on FastDeploy
|
||||
use_pact: true #Whether to use the PACT method for training
|
||||
activation_quantize_type: 'moving_average_abs_max' #Activations quantization methods
|
||||
quantize_op_types: #OPs that need to be quantized
|
||||
- conv2d
|
||||
- depthwise_conv2d
|
||||
|
||||
# Post-Training Quantization
|
||||
PTQ:
|
||||
calibration_method: 'avg' #Activations calibration algorithm of post-training quantization , Options: avg, abs_max, hist, KL, mse, emd
|
||||
skip_tensor_list: None #Developers can skip some conv layers‘ quantization
|
||||
|
||||
# Training Config
|
||||
TrainConfig:
|
||||
train_iter: 3000
|
||||
learning_rate: 0.00001
|
||||
optimizer_builder:
|
||||
optimizer:
|
||||
type: SGD
|
||||
weight_decay: 4.0e-05
|
||||
target_metric: 0.365
|
||||
|
||||
```
|
||||
|
||||
## More details
|
||||
|
||||
FastDeploy one-click quantization tool is powered by PaddeSlim, please refer to [Auto Compression Hyperparameter Tutorial](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/example/auto_compression/hyperparameter_tutorial.md) for more details.
|
||||
+50
@@ -0,0 +1,50 @@
|
||||
Global:
|
||||
model_dir: ./MobileNetV1_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: ['input']
|
||||
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
|
||||
|
||||
|
||||
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
|
||||
@@ -0,0 +1,48 @@
|
||||
Global:
|
||||
model_dir: ./ResNet50_vd_infer/
|
||||
format: 'paddle'
|
||||
model_filename: inference.pdmodel
|
||||
params_filename: inference.pdiparams
|
||||
image_path: ./ImageNet_val_640
|
||||
input_list: ['input']
|
||||
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
|
||||
|
||||
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
|
||||
@@ -0,0 +1,39 @@
|
||||
Global:
|
||||
model_dir: ./ppyoloe_plus_crn_s_80e_coco
|
||||
format: paddle
|
||||
model_filename: model.pdmodel
|
||||
params_filename: model.pdiparams
|
||||
qat_image_path: ./COCO_train_320
|
||||
ptq_image_path: ./COCO_val_320
|
||||
input_list: ['image','scale_factor']
|
||||
qat_preprocess: ppyoloe_plus_withNMS_image_preprocess
|
||||
ptq_preprocess: ppyoloe_plus_withNMS_image_preprocess
|
||||
qat_batch_size: 4
|
||||
|
||||
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: 5000
|
||||
learning_rate:
|
||||
type: CosineAnnealingDecay
|
||||
learning_rate: 0.00003
|
||||
T_max: 6000
|
||||
optimizer_builder:
|
||||
optimizer:
|
||||
type: SGD
|
||||
weight_decay: 4.0e-05
|
||||
@@ -0,0 +1,39 @@
|
||||
Global:
|
||||
model_dir: ./ppyoloe_crn_s_300e_coco
|
||||
format: paddle
|
||||
model_filename: model.pdmodel
|
||||
params_filename: model.pdiparams
|
||||
qat_image_path: ./COCO_train_320
|
||||
ptq_image_path: ./COCO_val_320
|
||||
input_list: ['image','scale_factor']
|
||||
qat_preprocess: ppyoloe_withNMS_image_preprocess
|
||||
ptq_preprocess: ppyoloe_withNMS_image_preprocess
|
||||
qat_batch_size: 4
|
||||
|
||||
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: 5000
|
||||
learning_rate:
|
||||
type: CosineAnnealingDecay
|
||||
learning_rate: 0.00003
|
||||
T_max: 6000
|
||||
optimizer_builder:
|
||||
optimizer:
|
||||
type: SGD
|
||||
weight_decay: 4.0e-05
|
||||
@@ -0,0 +1,37 @@
|
||||
Global:
|
||||
model_dir: ./yolov5s.onnx
|
||||
format: 'onnx'
|
||||
model_filename: model.pdmodel
|
||||
params_filename: model.pdiparams
|
||||
qat_image_path: ./COCO_train_320
|
||||
ptq_image_path: ./COCO_val_320
|
||||
input_list: ['x2paddle_images']
|
||||
qat_preprocess: yolo_image_preprocess
|
||||
ptq_preprocess: yolo_image_preprocess
|
||||
qat_batch_size: 1
|
||||
|
||||
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
|
||||
@@ -0,0 +1,38 @@
|
||||
Global:
|
||||
model_dir: ./yolov6s.onnx
|
||||
format: 'onnx'
|
||||
model_filename: model.pdmodel
|
||||
params_filename: model.pdiparams
|
||||
qat_image_path: ./COCO_train_320
|
||||
ptq_image_path: ./COCO_val_320
|
||||
input_list: ['x2paddle_image_arrays']
|
||||
qat_preprocess: yolo_image_preprocess
|
||||
ptq_preprocess: yolo_image_preprocess
|
||||
qat_batch_size: 1
|
||||
|
||||
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
|
||||
@@ -0,0 +1,37 @@
|
||||
Global:
|
||||
model_dir: ./yolov7.onnx
|
||||
format: 'onnx'
|
||||
model_filename: model.pdmodel
|
||||
params_filename: model.pdiparams
|
||||
qat_image_path: ./COCO_train_320
|
||||
ptq_image_path: ./COCO_val_320
|
||||
input_list: ['x2paddle_images']
|
||||
qat_preprocess: yolo_image_preprocess
|
||||
ptq_preprocess: yolo_image_preprocess
|
||||
qat_batch_size: 1
|
||||
|
||||
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
|
||||
@@ -0,0 +1,37 @@
|
||||
Global:
|
||||
model_dir: ./PP_LiteSeg_T_STDC1_cityscapes_without_argmax_infer
|
||||
format: paddle
|
||||
model_filename: model.pdmodel
|
||||
params_filename: model.pdiparams
|
||||
qat_image_path: ./train_stuttgart
|
||||
ptq_image_path: ./val_munster
|
||||
input_list: ['x']
|
||||
qat_preprocess: ppseg_cityscapes_qat_preprocess
|
||||
ptq_preprocess: ppseg_cityscapes_ptq_preprocess
|
||||
qat_batch_size: 16
|
||||
|
||||
|
||||
Distillation:
|
||||
alpha: 1.0
|
||||
loss: l2
|
||||
node:
|
||||
- conv2d_94.tmp_0
|
||||
|
||||
Quantization:
|
||||
onnx_format: True
|
||||
quantize_op_types:
|
||||
- conv2d
|
||||
- depthwise_conv2d
|
||||
|
||||
PTQ:
|
||||
calibration_method: 'avg' # option: avg, abs_max, hist, KL, mse
|
||||
skip_tensor_list: None
|
||||
|
||||
TrainConfig:
|
||||
epochs: 10
|
||||
eval_iter: 180
|
||||
learning_rate: 0.0005
|
||||
optimizer_builder:
|
||||
optimizer:
|
||||
type: SGD
|
||||
weight_decay: 4.0e-05
|
||||
Reference in New Issue
Block a user