Vision-Kek
commented
4 months ago Is it the case that train in one episode and test in another 200 episodes with the same category?
Yes.
Is there a requirement for the other query images to have any correlation with the query image used for training?
After fitting the adaption layers to one episode(=query+support), technically you are able to forward pass any other episode without running the adaption(training) again - also episodes featuring other classes. (If you want to prevent the re-adaption for new classes, you would need to modify the code, though, since currently even in quick-infer mode re-adaption will be invoked once an episode from a never-seen class appears.)
Regarding the requirements for the queries, naturally the model will be most suitable for episodes that are similar to the one used for fitting. In practice, the results depend on the dataset characteristics like intra-class variations and I suggest you to compare it empirically for your target dataset to decide what procedure is appropriate for your application.
In section 4.3, you point: We therefore evaluate the scenario where task-adaption is only done once, and the thus learned parameters are reutilized for every subsequent query. Table 5 reports the results averaged over 200 runs, where a run samples first an episode for training and then infers on further 200 queries.
I would like to inquire about the scenario for reusage. Is it the case that train in one episode and test in another 200 episodes with the same category? Is there a requirement for the other query images to have any correlation with the query image used for training?