Hey @kontrabas,
if you read the definition SaveFeatures() class, you will see this -
self.features = np.row_stack((self.features, out))
This essentially is stacking each batches prediction in sf.features array.
Hope this answers your question.
Hello @sgugger,
since fastai v2 got out - is this still the valid way of getting the activations of a specific layer?
Thanks in advance
@r2d2 can you share the link of less 12 ? Not sure which version of course is talked here.