torch.FloatTensor type error about yolo_v3_tutorial_from_scratch

[ayooshkathuria]

Okay, there was a bug in NMS which caused to give > 0 IOU threshhold for two boxes that didn't overlap. I pushed it just yesterday, but couldn't check it on GPU as I don't have access to one on weekdends. It works fine on CPU.

Your solution would then break the code on CPU, since .cuda() will raise an error. The best way to go about it would be to use torch.maximum or something. I'll look into it in a couple of hours and push the changes. Thanks for pointing out the bug though. If you could verify it with torch.max, perhaps you could make a pull request too. It'll be a few hours before I can push the fix.

[ayooshkathuria]

I've fixed the issue. It works fine on GPU now.

[e-sha]

if torch.cuda.is_available():
inter_area = torch.max(inter_rect_x2 - inter_rect_x1 + 1,torch.zeros(inter_rect_x2.shape).cuda())*torch.max(inter_rect_y2 - inter_rect_y1 + 1, torch.zeros(inter_rect_x2.shape).cuda())
else:
inter_area = torch.max(inter_rect_x2 - inter_rect_x1 + 1,torch.zeros(inter_rect_x2.shape))*torch.max(inter_rect_y2 - inter_rect_y1 + 1, torch.zeros(inter_rect_x2.shape))

It's better to use torch.clamp, so you don't actually need memory allocation for tensor of zeros:
nter_area = torch.clamp(inter_rect_x2 - inter_rect_x1 + 1,min=0)*torch.clamp(inter_rect_y2 - inter_rect_y1 + 1, 0)

[ayooshkathuria]

Can you make a pull request? I'm. A bit busy with college stuff that I can't access my work PC right now. I made the above pushes directly from the website. I agree with the choice of ,torch.clamp ( I have used the function to clip bounding boxes that exceed dimensions of the image). Plus. If you are gonna do the PR, please make sure the code works both on CUDA and CPU. I guess clamp is device agnostic but I can't exactly nail it down. And yes, the min arg to clamp should be 0.0 instead of 0 since it's a FloatTensor. Thanks for the suggestion.

[cloudliu01]

Thank you both for the quick fix