Hi @Yuxin-CV
DTN is an inspired work and it's a very good idea. But I have some issues that hopefully will be replied. The experiment setting is not clearly explained in the paper. In your code, I find that your model is trained on the train_val joint splits, not only using the train splits as shown in common protocol. Not as described in the paper that "The hyper-parameters are optimized on the validation set". I also find that your model uses the 224x224 as the input, not 84x84. I don't know if you didn't mention this in your paper deliberately? I think your comparison is unfair. When I corrected the above two points and retried the experiment, I find that the accuracy was very poor. Can you give me some good advice?
Looking forward to your reply!
Hi,
I run the model you provide in folder DTN_SEED#5 and get accuracy only around 71%. Ps, I run 1*600 tasks when testing. But it should around 77% according your report.
The forward pass of the GeneratorNet network and Net network (the feature extractor) multiply A_rebuild with self.s which is set to 10 and then passes it to the classifier.
Can you explain what this is? Why can't we pass it directly to the FC layer for classification?