Since few-shot learning mostly works on episodes and it tries to find the closest embedding of a given test image with support image. Is it possible to plot confusion matrix, as every time model takes a different subset of classes(say 5 out of 100 classes in 5 way classification). Currently my model takes subset of classes and it converts all actual labels to 0-4(for 5 way classification) and then based on that it gives final classification of test image.

To get confusion matrix what can I do?