'MaskFormer' object has no attribute 'module' about eclipse

[qjadud1994]

If you use a single GPU, you should remove .module. due to the DDP wrapper.

Also, we are currently checking the ADE20K 100-50 setting.

[zhengyuan-xie]

Hi, I also got this error. Do you use a single GPU? I think when a single GPU is used, the DDP training process won't be launched, so I simply change the corresponding code to:

if hasattr(self.model, "module"):
    self.model.module.copy_prompt_embed_weights()
    self.model.module.copy_mask_embed_weights()
    self.model.module.copy_no_obj_weights()
    self.model.module.copy_prompt_trans_decoder_weights() 
else: 
    self.model.copy_prompt_embed_weights()
    self.model.copy_mask_embed_weights()
    self.model.copy_no_obj_weights()
    self.model.copy_prompt_trans_decoder_weights() 

However, I got another error: RuntimeError: Global alloc not supported yet, one solution is replacing batch_dice_loss_jit with batch_dice_loss but I'm not sure whether it is correct.

[wendyhuang2021]

Yes! I used single GPU and similar code like you written. It remain 3.5 hr to finish the ADE20k 100-50 setting.
However, the author have produce error in this setting, and I cannot check if those edited harm the performance even finished above setting.
Maybe I would tried another setting later.