model overfitting issue

[alchemz]

Dear authors,
We have tested the COCO model with different frames, but outputting the same caption results. It seems to be a overfitting issue. Could you suggest if any parameters need to be specified at inference time, or it could be a data issue that the input frames are similar to some extent?

[ArminBaz]

Hi, did you find a solution to this problem?

[ret7020]

@alchemz Can you show your hyperparameters?

[zzk2021]

Me too. I have the same problem

[zzk2021]

It seems like the model isn‘t train completely. The loss at a high level. the outputs are high-frequency words

[alchemz]

Hi all, unfortunately I haven't find a solution to this.

@ret7020 For your question, we did not perform hyparameter tuning on the model with new datasets, but simply run inference with the existing model.