Skip to content

Latest commit

 

History

History
40 lines (38 loc) · 1.48 KB

OTHERS.md

File metadata and controls

40 lines (38 loc) · 1.48 KB
Raw

Export Pb

Example:

cd $PATH_ROOT/tools/retinanet
python exportPb.py

SWA Object Detection

Paper: SWA Object Detection
Usage:

  1. Step1: Train additional n epochs (MAX_ITERATION = SAVE_WEIGHTS_INTE*20 -> MAX_ITERATION = SAVE_WEIGHTS_INTE*(20+n) in cfgs.py) 
    cd $PATH_ROOT/tools/retinanet
python train_swa.py
  2. Step2: Average n trained weights 
    cd $PATH_ROOT/tools
python swa.py
  3. Step: Test 
    cd $PATH_ROOT/tools/retinanet
python test_dota.py --test_dir='/PATH/TO/IMAGES/'  
                    --gpus=0,1,2,3,4,5,6,7  
                    -ms (multi-scale testing, optional)
                    -s (visualization, optional)

Pyrotch Pretrain Weights Conversion via MMdnn

Take resnet50 as an example (torch 1.5.1, torchvision 0.6.1):

  1. Step1: 
    pip install mmdnn
mmdownload -f pytorch
mmdownload -f pytorch -n resnet50 -o ./
mmtoir -f pytorch -d resnet50 --inputShape 3,224,224 -n imagenet_resnet50.pth
mmtocode -f tensorflow --IRModelPath resnet50.pb --IRWeightPath resnet50.npy --dstModelPath tf_resnet50.py
  2. Step2: Migrate the generated network structure script to resnet_pytorch.py, and make some modifications, including the freezing of bn and the first few blocks, the construction of feature_dict variables, etc.