- modify
train_config.pyandeval_config.py:- change
backbone_typeyou want use [se-resnext50,vgg,resnet50,resnext50,resnext101] you can add your backbone inbackbones, or add torchvision supported model by modifyutils/model_utils.py - train_datasets_bpath =
'data/to/your/path'and same to test_datasets_bpath - change dataLoader_util to you want use(cv2 by default) [
cv2/PIL/jpeg4py(jpeg4py need compile )] - change
batch_sizeandgpu_idsto you want use - if you wanna use FP16 Mixed Precision, apex requirement, then change
fp16_using= True - if train with center loss, change
additive_loss_type='CenterLoss' - if train with COCO loss, change
additive_loss_type='COCOLoss' - if train with center loss and COCO loss, change
additive_loss_type='COCOLoss&CenterLoss' - if train only use softmax, change
additive_loss_type=Noneor'' - if train with focal loss, change
use_focal_loss=True
- change
- run
nohup python -m visdom.server &on linux shell then gotolocalhost:8097to see your model visual output - then run
python train.pyto trainor runfast_train.pyto train withDALIthat3x~20x(still debuging only support single GPU now) faster thanpytorch dataloader DALIspeed up training supportTriplet Model, if you have 2 GPU card, here's example:python -m torch.distributed.launch --nproc_per_node=2 fast_triplet_train.pychange--nproc_per_node=nif you have n GPU card.
- modify
triplet_train_config.pyreferenceSiamese-Network - run
python triplet_train.py
- support multi-GPU DALI speedup