A training problem about Global alloc not supported yet

[xiehousen]

I created a new running environment for mask2former according to the steps. When I train the COCO dataset, I can train normally, but when I train my dataset, I encounter the following problems.

I've been looking for a solution on Google for a long time, so I'd like to ask if you have any similar problems. Thank you very much for your reply.

[bowenc0221]

Does it work if you use batch_dice_loss instead of batch_dice_loss_jit?

[xiehousen]

if I use batch_dice_loss, that's worked.

[haotian-liu]

Hi @bowenc0221, I met the same issue when training the model on YouTube-VIS dataset (with official code). Do we need to turn JIT off or is there some way to fix this RuntimeError issue? Thanks!

[bowenc0221]

Hi @bowenc0221, I met the same issue when training the model on YouTube-VIS dataset (with official code). Do we need to turn JIT off or is there some way to fix this RuntimeError issue? Thanks!

I have never met this error. I think it is not necessary to use the JIT version of the loss function, so turning it off is the simplest solution. If you really want to fix this error, I would suggest posting it to the PyTorch team for help.

[haotian-liu]

Hi @bowenc0221, I met the same issue when training the model on YouTube-VIS dataset (with official code). Do we need to turn JIT off or is there some way to fix this RuntimeError issue? Thanks!

I have never met this error. I think it is not necessary to use the JIT version of the loss function, so turning it off is the simplest solution. If you really want to fix this error, I would suggest posting it to the PyTorch team for help.

Thank you!

[xiehousen]

Hi @bowenc0221, I met the same issue when training the model on YouTube-VIS dataset (with official code). Do we need to turn JIT off or is there some way to fix this RuntimeError issue? Thanks!

I have never met this error. I think it is not necessary to use the JIT version of the loss function, so turning it off is the simplest solution. If you really want to fix this error, I would suggest posting it to the PyTorch team for help.

Thank you!

I did not use the JIT version, and the The training speed becomes very slow. Do you have this question?

[ShijieVVu]

I created a new running environment for mask2former according to the steps. When I train the COCO dataset, I can train normally, but when I train my dataset, I encounter the following problems.

I've been looking for a solution on Google for a long time, so I'd like to ask if you have any similar problems. Thank you very much for your reply.

If your custom training set includes zero annotation images, this error would show up.

[deeptig84]

Hi @bowenc0221 bowenc0221

I would like to know how to turn off the batch_dice_loss , is it a config change in the Mask2Former , or I need to go and change the code itself? For Now, I have made changes in the matcher.py file and changed the code to call batch_dice_loss in place of batch_dice_loss_jit. same replacement I have done for batch_sigmoid_ce_loss_jit as well. I wanted to validate the changes with you. My training ran fine after these changes.

[YellowPig-zp]

Just to add to the comments with my personal experience regarding the issue. Seems like even if your data doesn't contain empty annotations, the codes would still throw out the same bug(I manually removed all the images/annotations for ade20k that do not have labels and encountered the same bug).

Seems like it also has something to do with the pytorch version. For v1.10 the bug appears, but when I down-grade to 1.9.1, it runs like a charm, and the training time is reduced for a few hours.

Hope this could help!

[Jianghanxiao]

Based on above comments, I also find that this case happens when some images have no gt. Below is my modification, which can somehow reduce the influence. It can still use jit most of the time and don't need to modify the dataset.

[lsnls]

Hi, @Jianghanxiao , your modification is effective. But I am confused about why tgt_mask.shape[0] == 0 might appear?

Hope this could be help! Thanks a lot!