when and when not to freeze

[ThisIsIsaac]

You freeze vgg16_bn weights in trainic15data.py but not in trainSyndata.py.

1. Why do you freeze in trainic15data.py and not in trainSyndata.py?
2. What is the benefit of freezing? I assumed we should ALWAYS let gradients flow through both vgg16 and CRAFT.
3. Have you trained vgg16 with CRAFT starting from Pytorch's pretrained weights? If so, how was the performence?
4. saving model like this:

net = CRAFT()
net.load_state_dict(copyStateDict(torch.load('CRAFT-Reimplemetation/pretrain/SynthText.pth')))
torch.save(net.state_dict(), os.path.join("pretrain","test.pth"))

only saves weights of CRAFT and not vgg16_bn. If I were to train both CRAFT and vgg16, both unfrozen, how would I save both CRAFT and vgg16's network together in a single model?

[backtime92]

@ThisIsIsaac The author freeze the model to train the linkRefiner (see the paper P11 and P12). When you train my reimplementation-model that you do not need freeze anything.

[backtime92]

@ThisIsIsaac I've changed the mistake. Em... I think you can train for MLT. Maybe IC15 train has some mistakes. I will refine it when I am no busy.

[ThisIsIsaac]

I've changed the mistake.

what mistake are you referring to?

Maybe IC15 train has some mistakes. I will refine it when I am no busy.

Maybe I can help out. I've also fixed some critical errors and done some cleanup. I would like to send a PR soon.

[backtime92]

@ThisIsIsaac In the trainic15data.py, code in line 114 should be freeze=False. Sorry, I do not know what's the meaning of PR.

[ThisIsIsaac]

I do not know what's the meaning of PR.

PR stands for pull request

In the trainic15data.py, code in line 114 should be freeze=False

will apply that to my copy of the code. thanks