Fariha

Same issue.

I am using 4 Tesla T4 gpus and I could only get to a batch size of 2 per gpu for distributed training. All the other hyper parameters are same...

Thank you so much for such a detailed answer. 1-I am only calculating the f1-score for segmentation only, from code it may seem like I am using boundary and distance...

I tried the gradient accumulation way and got to accumulate gradients after 32 samples. The training is still not stable and I tried with a really low learning rate as...

Here, '#' before a line represents that it is commented. ```python from mxnet import nd from models.changedetection.mantis.mantis_dn import * from mxnet import autograd,optimizer,gluon import mxnet as mx import horovod.mxnet as...

Thank you once again for such a detailed reply. Okay so when I calculate f1 score against class axis, it is actually good. So for visualising results and data, I...

  These are the results I got on a test image. I used the sliding window method to perform inferences. The third and fourth images represent ground truth and...

Thank you, you helped a lot. Yes that is true, it needs more training. I wanted to ask if I lower the image size and then apply the sliding window...

@humayunah The 'none' value comes out when score for that label does not exist because that class does not exit in the data point. So you will have to decide...