Why the FCN needs a pre-trained weight? about pixelrl HOT 7 OPEN

@machanic Hi, can you share us your code of PyTorch version of PixelRL? Thank you very much.

from pixelrl.

Thank you for your question.

Why needs the pre-trained FCN weight?

The pre-trained weights are used in order to make the training of the network stable.

Can I train the FCN directly without loading?

Maybe yes, but the training will be unstable especially at the beginning.

from pixelrl.

@rfuruta I have an idea of my paper to use your code, but the reward is only a scalar value for the whole image, rather than the pixel-wise rewards. Can I use PixelRL? Maybe the difficult is that many agent & action has to determine one reward.
Thank you.

from pixelrl.

I do not think PixelRL will work in such a situation. Training PixelRL with one reward is a remaining problem.

from pixelrl.

@rfuruta Why doesn't work? Does it because that so many actions have to be made to obtain one reward?
I have implemented that idea, but it works poorly, maybe my implementation is incorrect?

from pixelrl.

I guessed it would not work because the reward value is determined by many actions.
But, my guess may be wrong if your idea is working.

from pixelrl.

我注意到您使用chainer.serializers.load_npz('../denoise_with_convGRU/model/pretrained_15.npz', net)加载预训练的 FCN 重吗？为什么需要预训练的 FCN 重？ 我现在将代码翻译成 PyTorch 版本，我可以直接训练 FCN 而不是加载预训练的重吗？ 谢谢！
can you share us your code of PyTorch version of PixelRL? Thank you very much.

from pixelrl.