rfuruta / pixelrl Goto Github PK

Hi!
This is more general doubt related to the concept and not directly the implementation. Can you please explain how the inferencing works? From the paper it seems that you already have to know the final denoised image (to calculate rewards and to improve iteratively). So how does this work when you are on a test set in which case you don't have access to the actual groundtruths?

If yes, can you provide the link to the same.

how can it be run in colab, have issues with cupy

Very interesting project. I've been recently trying to test it with 3D images, but I didn't have much progress. I believe it has to do with the shape of the output of the Policy network being pout.sample().shape = (batch_size, crop_size, crop_size). I think it should include the 3 channels too, however I don't know how to change the shape of the SoftmaxPolicy layer. Is that the right approach? And, if it is, could you give me a hint on how to proceed?

Can you publish the code for the Image Restoration and Color Enhancement sections?

May I ask what is your hand-design reward functions which are used in image denoising problems?

Thanks

Trying the test function but all images are in gray, is color not supported?

What does moved_image = self.image + move[:,np.newaxis,:,:] mean, which is located in in https://github.com/rfuruta/pixelRL/blob/master/denoise/State.py#L17

Can you explain the function of def step(self, act) a bit more to make it easier to understand? It is located in
https://github.com/rfuruta/pixelRL/blob/master/denoise/State.py#L13

Really interesting project of implementing image processing tasks using DeepRL network.
Can this work be used for efficient pixel-level image compression instead of normal downsizing an image. If yes, can anyone plz help me saying how to achieve that?

I notice that you use chainer.serializers.load_npz('../denoise_with_convGRU/model/pretrained_15.npz', net) to load a pre-trained FCN weight? Why needs the pre-trained FCN weight?
I currently translate the code into PyTorch version, Can I train the FCN directly without loading pretrained weights?
Thank you!