attention head weight shapes in paddle and pytorch model are different

[iazdan]

I was able to convert the PP-structure paddle train model (best_accuracy.pdparams, best_accuracy.pdopt, best_accuracy.stated) to pytorch model (model.pth). However I realized that the model output is not the same as the paddle model. I did some investigation and I seems the paddle and the backbone weights are the same however the shape of the attention head in the pytorch model is different than the paddle model. This is a bug in the code. Would you be able to fix it?

Thanks

[frotms]

table model is not supported.