Pytorch

How to finetune a network

1. Modify dataLoader / Class for custom dataset
   1. segmentation tasks
      1. modify the palette for rendering the result labels
   2. modify num_classes
   3. dataloader part: modify __init__ and __getitem__
2. Modify training code:
   1. Change dataset.
      • import new dataset functions and classes
      • change initialization for the dataset
      • change mean-std
      • modify all parts associated with dataset: eg. xxxDataset.num_class
   2. Net Surgery
      • Load initial net
      • modify
      • net_init.final._modules['4'] = nn.Conv2d(512, UAVDataset.num_classes, kernel_size=1).cuda()
3. Hook https://pytorch.org/tutorials/beginner/former_torchies/nn_tutorial.html#forward-and-backward-function-hooks
4. Extract Features: https://becominghuman.ai/extract-a-feature-vector-for-any-image-with-pytorch-9717561d1d4c