AssertionError during the validation phase about pytorch-retinanet HOT 8 OPEN

The bug comes from 270 line in model.py
return [torch.zeros(0), torch.zeros(0), torch.zeros(0, 4)]
Because it does not return the value in GPU device. Change this line like this:
return [torch.zeros([1]).cuda(0), torch.zeros([1]).cuda(0), torch.zeros([1, 4]).cuda(0)]
Do not forget to change torch.zeros(0) to torch.zeros([1]), otherwise there will be another bug about array bound

from pytorch-retinanet.

I got this error too

from pytorch-retinanet.

emmmmm , I got the same error

from pytorch-retinanet.

My solution is moving this mAP = csv_eval.evaluate(dataset_val, retinanet) out of for iter_num, data in enumerate(dataloader_train): because calculating mAP should be done after one epoch.

from pytorch-retinanet.

@Kwongrf the train.py already has that line of code out of the for loop.. I am still getting the same error.. Any idea on what else is wrong?

from pytorch-retinanet.

@mahaling Have you modified other parts of train.py... In my case, I tried to change train.py to train on my own data but I made a mistake that I moved that line in the for loop. So it just trained one batch and then it started to evaluate. It seems like yours can train one entire epoch and then get the assertionerror. Maybe you can print the inputs in .../torch/nn/parallel/_functions.py .

from pytorch-retinanet.

For now I turned off csv_eval, so that I get the data trained for 100 epochs.

from pytorch-retinanet.

I get the assertion error again after I have trained for 4 epochs...It seems like there are some faults in csv_eval.py

from pytorch-retinanet.