pricexu / stager Goto Github PK

1. In the meta-testing phase, you use both base and novel data. Is this reasonable in few-shot learning? You have used base data in the meta-training phase.Generally speaking, the tasks in the meta-training and meta-testing phases are different, right?
2. What's more, in the test function, why do you still perform 300 epoch of training? In fact, you train the model with the data of base and novel at the same time, and then test it. Is this reasonable? In my opinion, the upper bound of generalized few-shot may be obtained by training the two data together. This is also one of the reasons why I think your experimental results are high.
3. In terms of training times, when the model is trained for 1000 times in meta-train, it will be trained for 3000 times in meta-test.

Hello, first thank you for sharing the code! It was so glad to see your works.

Could you give me the hyperparameter settings you used that can reproduce the results in the paper?

Although you have given some hyperparameter search ranges, it is still difficult to reproduce the results. And the results is much lower.

I hope I can get your help and look forward to your reply very sincerely！