cc-ai / maskgenerator Goto Github PK

I see all inputs in set_input_display are unsqueezed. Why is that? Don't they come in the shape batch x channels x height x width?

Why is there a need for self.image = input.data.x.unsqueeze(0)?

I'm seeing undocumented branches...!

Regarding the comet image log:
It seems that we are loading number of images to display * batch size and only displaying the first image of each batch loaded. And we only want to display number of images to display.
Does it make sense to change the code to load the images in this way:
test_display_images =[Dict(val_loader.dataset[i]) for i in range(opts.comet.display_size)]
rather than:
test_display_images = [Dict(iter(val_loader).next()) for i in range(opts.comet.display_size)]
It seems to me that would be more efficient.
If you agree I'll make changes and PR!

Code on a repo always needs a license

Pull request to add depth map input

Currently the images of depth maps are not logged in comet. Here is how depth maps are set as input:

• A depth map a depth map is read as an image (can be encoded either in Megadepth way (1 channel) or our simulator way (4 channels))
• Converted to a normalized depth map tensor T with values between [-1, 1]
  Now what is a good way to recover the depth map image ?
  Ideally, I would like to normalize T between [0,1], and plot log(1/T). (you get better visualizations of depth maps if you plot the log inverse depth)
  How do I handle zeros ? Can I just add an arbitratily small epsilon ?