Junyang Lin comments

Results 173 comments of


                                            Junyang Lin

Pile Dataset

It is not available for us to release the processed data. You can try downloading from the official website.

Do the training data for the pretrained OFA include samples from COCO val set?

No, such data are not included

Information leakage occurs in snli-ve training

The related setup of this task is mentioned in the paper.

IS A RTXA6000 ALONE able to finetune this model in a custom dataset?

Never used this before. I think it is good enough, as it has 48GB memory. Maybe for huge models you should use relatively small batch sizes.

how the model perform on object detection?

Did not try with this task by finetuning seriously. We'll figure our a solution in the near future.

Question about fintune Document Task

I have provided the code but not the script. I'll update it soon.

Question about label_smoothed_cross_entropy.py

Which version of pytorch are you using? This is because the problem of in-place operation in the new version of pytorch, for example, >=1.10

Is there any onnx configuration or converted onnx model for OFA?

Not yet. Sorry about that. Would you mind telling us in which scenarios you need such models?

question about how to use topp sampling?

For what reason you consider about using topp sampling? For this repo, we do not have relevant experience. Perhaps it is still better to use beam search following our practice...

Question about the Visual Grounding inference result

Yes, this is one significant problem of the current model. One way to tackle this problem is to compute the average probabilities of the output logits, and set a threshold...