Pytorch 1.2 compatible

[nesvera]

Nice tutorial. It helped me a lot while I was learning about the implementation of SSD, thanks :D

I encountered the following error when I tried to use your code with pytorch 1.2

Traceback (most recent call last):
File "eval.py", line 88, in
evaluate(test_loader, model)
File "eval.py", line 63, in evaluate top_k=200)
File "/home/feaf-seat-1/Documents/nesvera/object_detection/a-PyTorch-Tutorial-to-Object-Detection/model.py", line 494, in detect_objects
suppress = torch.max(suppress, overlap[box] > max_overlap)
RuntimeError: Expected object of scalar type Byte but got scalar type Bool for argument #2 'other'

I changed some lines of the model.py code to be able to run the code with pytorch 1.2. Now, it is compatible with both versions: 0.4 and 1.2.

Beyond, I created an algorithm to make predictions from a video (based in your detect code), and I noticed that the FPS increase almost 2 times using pytorch 1.2.

Running on a Quadro M5000, I achieved the following results:
pytorch 0.4 (original code)

• detect.py = Mean Average Precision (mAP): 0.739
• detect_video.py = ~19 FPS

pytorch 1.2 (modified code)

• detect.py = Mean Average Precision (mAP): 0.739
• detect_video.py = ~35 FPS

I didn't know exactly if the increase of the FPS comes from the code (that was changed from uint8 to bool), or it comes from the pytorch upgrade. Probably the latter.