search

G-U-N / PyCIL

PyCIL: A Python Toolbox for Class-Incremental Learning
Other
815 stars 138 forks source link

about model ，about cifa10 #83

Closed Hephaestusxg closed 6 months ago

Hephaestusxg commented 6 months ago

Sorry to bother yo, is saving the weights done in the after task function? I'd like to ask for some guidance on this. However, I noticed that when using DER, this function is not utilized. Can I save the model weights at the initial stage as well as after each incremental step? For example, the weights for 8 classes, 9 classes, and 10 classes (on CIFAR-10).

zhoudw-zdw commented 6 months ago

See https://github.com/G-U-N/PyCIL/issues/77#issuecomment-2098657237

Hephaestusxg commented 5 months ago

See [#77 (comment)](https://github.com/G-U-N/PyCIL/issues/77#issuecomment-2098657

See #77 (comment) Thank you for your response. When I printed the model structure, I noticed that the fully connected (fc) layer is set for 10-class classification for CIFAR-10. My previous understanding was that during the initial training stage, for instance, if there are initially 7 classes, the fc layer should be set for 7-class classification. Based on what the code shows, the accuracy calculated this way would be inaccurate, right?