Yang Zou

 In original DeepLab, iterations rather than epochs are used. As long as the network see enough images, it does not matter if a full epoch in the very end...

What are your specific numbers? And I think you change your network structures. And a trick here is that not training model for to long since the model will overfit...

Sometimes you can use the target labels to find a source model that has a good transfer learning performance, although this may not be true in practice.

Do you know what is iou, confusion matrix, how to calculate confusion matrix for multiple classes and how to calculate iou from confusion matrices? If you do know, you can...

Yes. You get the point.

Did you use correct label mapping function to map label to train_id? And also I resized data of SYNTHIA to 1024x2048

'600,900' is the size of the cropped size. 1024,2048 is the size of the image.

Yes, I pretrained on source and then self-train on target. For visda17, I just train 3 epochs on source and 20 epochs on target. For Office-31, pretrain for 200 epochs...

Since pseudo-label generation is time-costly for semantic segmentation, we generate pseudo-label for segmentation every two epochs. However, for pseudo-label generation for image classification, it's much faster. Thus we generate pseudo-label...

The code runs well in my desktop. The bug seems to caused by the string format of 'loss.data.cpu().numpy()'. But that's weird since 'loss.data.cpu().numpy()' should be a float number. You may...