Training breaks with Pytorch version 1.13.0 and torchvision version 0.8.0

[dt1729]

I was trying to run the coco dataset training example and I came across the issue of

  File "test.py", line 126, in <module>
    prefetch_test(opt)
  File "test.py", line 58, in prefetch_test
    detector.model.v2_transform()
  File "/home/dt/DetectionNN/CenterNet-HarDNet/src/lib/models/networks/hardnet.py", line 494, in v2_transform
    self.base[i].transform(blk, self.trt)
  File "/home/dt/DetectionNN/CenterNet-HarDNet/src/lib/models/networks/hardnet.py", line 227, in transform
    self.conv_layers[i].weight[0:part[0], :, :,:] = w_src[:, 0:in_ch, :,:]
RuntimeError: a view of a leaf Variable that requires grad is being used in an in-place operation.

both w_src and weight are leaf variables and pytorch doesn't allow the same. Go through this and this for more detail.

A temporary solution is virtual env with pytorch version 1.7.0 and torchvision version 0.8.1 works with that.

Sourcing pycocotools can be difficult so this line can help in the virtual environment:

pip3 install pycocotools

otherwise, add pycocotools to the requirements.txt