UDOP pretraining with MAE decoder about i-code HOT 6 CLOSED

I'm not the author of the work. From their introduction, the weights are result of pretraining

from i-code.

Please ignore my question, I found the answer. At the beginning of forward() pass this loop

if input_dict is not None:
            return_task_outputs = []
            for task in input_dict:
                return_task_outputs.append(self.forward(**input_dict[task]))
            return return_task_outputs

computes the loss for each training task separately, then they are summed up in trainer.training_step()

from i-code.

Hi @haixpham
Can you please share you managed to reproduce the pretraining?
BTW, are you sure that the supplied checkpoints are pretraining? I saw that they used them for inference in RVL-CDIP for doc. classification.
Thanks!

from i-code.

@ofir1080 Unfortunately this is for a company project so I'm not allowed to share code at the moment. The code in this repository is for downstream finetuning, the code for pretraining is at a different repo

from i-code.

Yes sure, I was just asking if the given checkpoints are already finetuned, or only pretrained?

from i-code.

@ofir1080 Unfortunately this is for a company project so I'm not allowed to share code at the moment. The code in this repository is for downstream finetuning, the code for pretraining is at a different repo

@haixpham Hi, could you please give the link of the other repo for the pretraining code? I can't find it.

from i-code.