Siamese-Network and Triplet-Network

how to run Siamese-Network

• modify train_config.py and eval_config.py:
  • change backbone_type you want use [se-resnext50, vgg, resnet50, resnext50, resnext101] you can add your backbone in backbones, or add torchvision supported model by modify utils/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_size and gpu_ids to 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 = None or ''
  • if train with focal loss, change use_focal_loss = True
• run nohup python -m visdom.server & on linux shell then goto localhost:8097 to see your model visual output
• then run python train.py to train or run fast_train.py to train with DALI that 3x~20x(still debuging only support single GPU now) faster than pytorch dataloader
• DALI speed up training support Triplet Model, if you have 2 GPU card, here's example:python -m torch.distributed.launch --nproc_per_node=2 fast_triplet_train.py change --nproc_per_node=n if you have n GPU card.

Triplet-Network

• modify triplet_train_config.py reference Siamese-Network
• run python triplet_train.py

requirement

• pytorch >= 1.0
• opencv
• PIL
• visdom
• jpeg4py
• tqdm
• apex
• DALI

TODO

• support multi-GPU DALI speedup