new wrapper

[nithin127]

to resize images as 80x80

[maximecb]

This is a good idea, but there are two things I would do differently:

• Put the import cv2 into the function that needs it. This is because OpenCV has been problematic to install on some systems, and I would rather make it so only people who need to use that functionality need to have it installed.

• Don't try to also combine the PyTorch observation reshaping into this wrapper. Keep them separate. If people need to both resize and transpose the observations, they should use both wrappers.

[nithin127]

So, import cv2 would be called every time the env.reset() or env.step() is called; should I go ahead with this?

[maximecb]

Yes, put the import inside both reset and step.

[maximecb]

Don't add the RewardWrapper to the repo. It's clearly specific to your own use-case/experiments.

[nithin127]

done;

[bhairavmehta95]

Why not just use the Pytorch transforms for this?

[bhairavmehta95]

resize = T.Compose([T.ToPILImage(),
                    T.Resize(40, interpolation=Image.CUBIC),
                    T.ToTensor()])

https://pytorch.org/tutorials/intermediate/reinforcement_q_learning.html

(You'd put 80 instead of 40)

Also, looking through this, only axes 0 and 2 are swapped, which would make your image, CxWxH instead of CxHxW, no?

https://pytorch.org/docs/stable/torchvision/transforms.html#torchvision.transforms.ToPILImage

[nithin127]

@bhairavmehta95 Florian suggested that we try 80; And I swap(0,2) just to resize it, then I swap it back -- (0,2) again; so essentially the Resize wrapper is not changing any dimensions

But the PyTorchObsWrapper does use an observation.transpose(2, 1, 0); do you think it should be observation.transpose(2, 0, 1)

[bhairavmehta95]

Yeah dimensions don't matter, so I think he's right.

I don't think it matters, as long as it's consistent. Only reason I'd say 2, 0, 1 is then you're only moving the channel. Makes it better to visualize, otherwise, if you feed your image tensor into something like tensorboardX, it will show up rotated. (As it's expecting 2, 0, 1, but you're giving it 2, 1, 0)

[nithin127]

done!

[maximecb]

The new wrapper is fine, but you've made a one-line change to the PyTorchObsWrapper. Can you remove that?