kuangliu

This is some experiment code I wrote for learning, not aiming for accuracy. If you do use pretrained VGG model, to my best knowledge the reason might be: 1. potential...

@dydokamil You can try my fork: https://github.com/kuangliu/vim-easymotion

No. I know the loss is not working. I would like to know is there any bug in the `encoder` first.

Cool. I update the loss function. I'm still testing it. The loss vibrates. Not sure what's going wrong.

Yeah. Just as Faster RCNN.

1. Run script `get_state_dict.py` first to get a base model. 2. Put the generated model in `model/net.pth`. 3. Then start the training with `python train.py --lr=1e-3`.

Yeah, it's a bug, should add a sigmoid layer to normalize the cls_output. Notice sigmoid don't output the class probability, just normalize it to avoid numeric instability. And the final...

I tried replacing softmax with only sigmoid. It seems working better. I'll look into it carefully and report back later.

Forget what I say... Sigmoid works fine. See `loss.py` for more details. Although you may find a softmax version of focal loss. But I stick with the original paper using...