Transfer your keras model to pytorch about yad2k HOT 4 CLOSED

An equivalent PyTorch operation is called PixelShuffle. You may also want to take a look at this existing YOLOv2 PyTorch implementation.

I don't think there's a NumPy equivalent to this function, but this white paper by Shi et al. has some good diagrams for what this operation is doing and how it's motivated by the upscaling convolution operation (often erroneously called deconvolution).

from yad2k.

@allanzelener Thanks for your response!
Could you please tell me what is the range of your network output layers value? I mean do you have negative number as the result of forward path when you forward an input image? Another thing which is ambiguous for me is that, Is your input to the network normalized (I mean between -1 and 1) or not?

Another thing which I would like to know ,if it is okay for you, is that What is the order of concatenation in your implementation. I mean in concatenate_1 layer!

from yad2k.

Please see this link which i provide an example of usage this function in stackoverflow. In addition to reshaping, it also changes the value of the input tensors! Is it logical to change the value?

from yad2k.

Ah, I see. I didn't realize they only implemented it for scaling factors > 1. I answered your question on Stackoverflow and opened an issue on the PyTorch Github here.

None of these operations should change any values, only their position within the array.

Yes, negative outputs are theoretically possible for the width and height parameters. I haven't checked the possible ranges. It's also possible to have a bounding box with width and height such that it extends beyond the image boundary.

I'm going to close this issue now since this is more related to PyTorch than this project.

from yad2k.