aabbas90 / cops Goto Github PK

Hi, is there any script for inference that can be used to see the results on new images?

I don't have available any gpu with enough memory to execute your training scripts with the config you provided.
Because of that I have executed the pretrain on cityscapes with a batchsize of 2 (instead of 12), ending with a network with the following validation metrics (note that not having the small small validation set, I simply used the whole cityscapes validation set):

(the training total_loss was 1.501 at the final iteration)
I then used the last checkpoint of this pretrain for the fully differentiable training, with a bachsize of 8 (instead of 24), but during the training the total_loss went up to around 9.3, while the pq and sq went down to 0 (both the training and the evaluation ones at the end). I also tried lowering the learning rate, but the results were always the same.

Is this a known issue? Are there any solutions?

PS: in the paper you said you used {1, 4, 16, 32, 64, 128} as edge distances with cityscapes, but in your training configuration the maximum distance is 64. Is it intentional?