How to use this code when infer? about fgvc-pim HOT 1 CLOSED

"We adopt a very simple design, which passes each feature point through a fully connected layer to predict the category.
When the highest probability of the predicted result after softmax is greater than a certain value, the feature point
is considered as a helpful feature and will be reserved for subsequent fusion. Conversely, the feature point is considered less helpful for fine-grained classification." -- https://arxiv.org/pdf/2202.03822v1.pdf

To know which area is discriminative or not, the PIM module does not need the ground truth.

By the way, we update our code. Hope the new version can provide more clearer explanation. Thanks.

from fgvc-pim.