Fixed bug in model when evaluating if no detection was found

[flifuehu]

Hi yhenon,

First of all, thanks for your work here, it's amazing. Now, I think I may have found an error when evaluating the accuracy of the network. Please correct me if I'm wrong, but I think line 271 of model.py:

pytorch-retinanet/model.py

Line 271 in 71f8e0a

return [torch.zeros(0), torch.zeros(0), torch.zeros(0, 4)]

should read:

return [torch.zeros(1), torch.zeros(1), torch.zeros(1, 4)]

Otherwise, DataParallel will give an error when asserting that all variables are CUDA after gathering them if no scores_over_thresh > 0 is found.

[AljoSt]

Should it not rather contain the batch size?
so return [torch.zeros(batch_size, 0), torch.zeros(batch_size, 0), torch.zeros(batch_size, 0, 4)]