GAT is related with batch size about frt-pad

[WentianZhang-ML]

Hi, thanks for trying our work.
About your question, you are right. The architecture of GAT is related to batch size. If you use GAT as the cross-modal adapter, the batch size must be the same in the training and testing phase. But you can copy 1 sample to 32 in the channel-wise to continue the test and prediction when you trained the model with a batch size of 32.

[Wei-Baldwin-Zeng]

Hi, thanks for trying our work. About your question, you are right. The architecture of GAT is related to batch size. If you use GAT as the cross-modal adapter, the batch size must be the same in the training and testing phase. But you can copy 1 sample to 32 in the channel-wise to continue the test and prediction when you trained the model with a batch size of 32.

Thank you so much for your quick response. But why should the GAT is related to batch size? As I understand, multiple batches in deep learning are just for speed up the training process, now the GAT seems to reshape the multiple batches to one huge batch, does it not support multiple batches for training? Or is there any way that I can also change the architecture of GAT so that it is not related to batch size.

Again, very wonderful work. Hope to hear your response.

[WentianZhang-ML]

Sorry for the late response. There was another way to support multiple batches, but we didn't find it until a year after we completed the project. You can refer to the solution in the https://github.com/Diego999/pyGAT/issues/36. But we have not tried it out. Please let me know if it works with our method after you use it!