Question about task-specific pretrained model

[Fengyee]

Hi there,

I wonder if you can discuss about how is the task-specific model is obtained. Is it pretrained weights from torchvision or you have some pretrain pipeline not included in this repo?

Because currently the code for train_image is just using resnet34 arch and network weights are not frozen.

Best,