abhrac / relational-proxies Goto Github PK

Inrelational-proxies.py
Line 93 mean_logits = global_logits + summary_logits + relation_logits / 3. Do you need to calculate the average of these three logits? A parenthesis is missing here. This value directly affects the prediction results, and the experimental results of the entire article are based on wrong code. Or is this the original calculation method?

Thanks for you codes, but i have a small question about the results. I have trained the relational-proxies model from strach on CUB_200_2011 on 1 3090 gpu, but the acc was only about 59%, not the reported 92% in you paper. The test result was slightly better than 92% (92.19%) in your checkpoints, and the START_EPOCH was 96 which means the 'epoch200.pth' should have been trained for 96 epochs, but not 200. So this 'epoch200.pth' was trained not only with CUB_200_2011? or you just have used other tricks didn't mentationed in the paper to get this 92% acc?