RuntimeError when 'inference_batch_size' > 1

[tomrunia]

I believe there is a small bug with setting the inference batch size > 1. For example, when running:

python main.py --model FlowNet2 --resume /path/to/dir/flownet2-pytorch/pretrained/FlowNet2_checkpoint.pth.tar --skip_training --skip_validation --inference_dataset ImagesFromFolder --inference_dataset_root ./path/with/images/  --inference_dataset_iext jpg --inference --save_flow --inference_batch_size 2

The following error is thrown:

File "/path/to/dir/pytorch/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/path/to/dir/flownet2-pytorch/losses.py", line 21, in forward
    lossvalue = torch.abs(output - target).mean()
RuntimeError: The size of tensor a (2) must match the size of tensor b (3) at non-singleton dimension 1

[tomrunia]

I fixed the error and created pull request #84