The learning rate is too small

[youngfly11]

I found that when you trained the FCN8s model, the learning rate is too small (1e-14). I remember that the learning rate is set to 1e-4 in the original FCN paper. I am a little confused.
Can you give me some answer? Thank you for advance

[wkentaro]

I refer to its journal version:
https://github.com/shelhamer/fcn.berkeleyvision.org/blob/affc85ff245099dbde06609d66ed08d73411f4ad/voc-fcn8s/solver.prototxt#L10

[youngfly11]

Thank you for your answer. The loss does not decay when i train the FCN8s on the CamVid dataset. i am so confused. I have another question. Do you use the pretrained model or not?

[wkentaro]

Yeah, I use pretrained model of fcn16s: https://github.com/wkentaro/pytorch-fcn/blob/master/examples/voc/train_fcn8s.py#L76
which is required to train fcn8s.

[youngfly11]

Thanks you so much. Can i directly use the pretrained model of fcn8s? i remember that you give a url to download the fcn8s_from_caffe.pth in FCN8s.py?

[wkentaro]

You can do like below:

import torchfcn
model = torchfcn.FCN8s()
pretrained_model = model.download()
model.load_state_dict(torch.load(pretrained_model))