Skip to content

Latest commit

 

History

History
50 lines (39 loc) · 1.67 KB

quant.md

File metadata and controls

50 lines (39 loc) · 1.67 KB
Raw

We provide quantization aware training(QAT) interface based on MQBench.

Get Started

Training hyper-parameters for example

  • batch size = the batch size of the fp pretrained model
  • learning rate per image = the final convergent learning rate of the float pretrained model(YOLOX may need smaller learning rate)
  • weight decay = 0
  • optimizer: SGD
  • epochs = 6
  • lr_scheduler: MultiStepLR
  • milestones: [2, 4]
  • type of bn: freeze_bn
  • type of head: You can train a float pretrained model with shared head while fine-tune a quantized model with an unshared head.
  • training transformer: [*train_resize, *to_tensor]
  • ema: False
  • the template of retinanet’s qat config is configs/det/retinanet/retinanet-r18-improve_quant_trt.yaml
  • the config of retinanet’s pretrained fp model is configs/det/retinanet/retinanet-r18-improve.yaml
  • the template of yolox’s qat config is configs/det/retinanet/yolox_s_ret_a1_comloc_quant_trt.yaml
  • the config of yolox’s pretrained fp model is configs/det/retinanet/yolox_s_ret_a1_comloc.yaml

train

Quantization-aware train a model based on a float pretrained model.

git clone https://github.com/ModelTC/EOD.git
git clone https://github.com/ModelTC/MQBench.git
cd scripts
# Follow the prompts to set config in train_quant.sh.
sh train_quant.sh

deploy

export the quantized model to ONNX

cd scripts
# Follow the prompts to set config in quant_deploy.sh.
sh quant_deploy.sh

result

Model map@fp32 ptq\int8 qat\int8
retinanet-r18-improve 40.7 40.5 40.7
yolox_s 40.4 39.4 39.8