valeoai / mtaf Goto Github PK

Thanks for enlightening work agian.

I train the Mdis method for one source and one target, but I am confused for the loss, and I plot by tensorboard.
And as I think, the adv loss should walk low and the discrimitor loss should walk higher.
but in the loss below, the two losses oscillate around a number. whats wrong with it?

Besides, I infer the training results should be better when training in manner of 1source 1target instead of 1source multi target. But in my training, I dont get good results.

So hope your thought sincerely.

And my training config:
adv loss weight: 0.5
adv learning rate: 1e-5
seg learning rate: 1.25e-5

adversarial loss of one source and one target

dicriminator loss of one source and one target

Thanks for the great research!

I have a question though, the mIoU you report in your paper is for 7 classes, but the segmentation colour map in the qualitative analysis seems to be for the 19 classes commonly used in domain adaptive semantic segmentation.

In other words, how can a model trained on 7 classes be used to generate a 19-class segmentation colour map? Or am I wrong in my understanding?

I look forward to your response.

Thank you!

In train_UDA.py
758 line

        d_main_list[i] = d_main
        optimizer_d_main_list.append(optimizer_d_main)
        d_aux_list[i] = d_aux
        optimizer_d_aux_list.append(optimizer_d_aux)

If this were done（d_main_list[i] = d_main and d_aux_list[i] = d_aux）, it would make all the discriminators in the list use the same one, shouldn't there be one discriminator for each classifier?

Thank you for sharing the code for your excellent work. I have some basic questions about your implementation.
pred_trg_main = interp_target(all_pred_trg_main[i+1]) ## what does [i+1] mean? pred_trg_main_list.append(pred_trg_main) pred_trg_target = interp_target(all_pred_trg_main[0]) ## what does [0] mean? pred_trg_target_list.append(pred_trg_target)

In train_UDA.py, line 829-836, why should we use index[i+1] and [0]? What's the meaning of that? Also, where is the definition of the target-agnostic classifier in your code?

Thanks again and look forward to hearing back from you!

Hello, thanks for sharing the code and such a good contribution. I would like to run your method on a setup that is a bit different, specifically adapting from Cityscapes ---> BDD, Mapillary. I have seen that the code accepts Cityscapes for both source and target, so that shouldnt be a problem, and I have added a dataloader for BDD to be the target 1.

In order to get the best performance, do I need to train the baseline and then train the method using MTKT or MDIS loading the baseline as pretrained? Or do I get the best performance directly by running the training script for MTKT or MDIS without the baseline?

Thanks for sharing your codes.

I was impressed with your good research.

Could you explain why the output map is not resized for target size(cfg.TEST.OUTPUT_SIZE_TARGET) in the case of Mapillary dataset in line 57 of eval_UDA.py?

When I tested the trained model on Mapillary dataset, inference took a long time due to the large resolution.

I'm looking forward to hearing from you.

Thank you!

thank you for you sharing!

Thanks for enlightening and practical work about multi-target DA !
I have read your paper, and I found one source dataset and 3 target datasets of unequal quantity, does the quantity of data for every domain matters? And what is the appropriate amount of training data for MTKT?
Another question, I want to know why KL loss is used for knowledge transfer? If I want to train an embedding word instead of a segmentation map, is the KL loss appropriate, and is there a better alternative?

Hello! @SportaXD
Thank you for your great work!

I was reproducing the code and noticed: the labels in the IDD dataset are in JSON file format instead of segmentation label form.

How is this problem solved?

Thank you for sharing the code for your excellent work. I have some basic questions about your implementation.

d_main = get_fc_discriminator(num_classes=num_classes)
d_main.train()
d_main.to(device)
d_aux = get_fc_discriminator(num_classes=num_classes)
d_aux.train()
d_aux.to(device)

Can you tell me why the multi-domain baseline code does not use multiple discriminators but only one discriminator.
It looks like a single domain approach.
Thanks！