Question about the training strategy in the stardard FSL setting about fsl-global

[SY-Xuan]

I have the same question with you!

[tiangeluo]

We would fix the 5 samples from novel classes in our training and such a way help the model perform in the novel classes in general. For the incremental classes setting, like add extra 20 classes to 100 classes of Mini-ImageNet, we have discussed in Section 4.3.2. In that case, we would only learn the new global representations for the new 20 classes while keeping the model parameters and the 100 global representations fixed.

[xssdss]

Before classifying the query samples in the test phase, has the model seen all 100 support samples (20 novel classes, 5 support samples for each class)?