search

edgarschnfld / CADA-VAE-PyTorch

Official implementation of the paper "Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders" (CVPR 2019)
MIT License
283 stars 57 forks source link

Ablation study exmperiment #19

Open GanAndreas opened 4 years ago

GanAndreas commented 4 years ago

Hi, first of all, thank you for your work. It's so great!!!

I want to recreate your ablation study by train the model with only CA-VAE and DA-VAE. I tried to do it by disable the distance loss or cross-reconstruction loss in the total loss, but I can't get the result on your paper. I only get harmonic mean accuracy of 2.18% for CA-VAE and 39.52% for DA-VAE. Can you point out my mistake?

Thank you

edgarschnfld commented 4 years ago

Hi, it's very much possible that the hyperparameters have to be adjusted for each ablation, especially the early stopping step for the final classifier. That should be done on the validation split (=the split that is not used for the final testing), see here https://github.com/edgarschnfld/CADA-VAE-PyTorch/issues/10.

angela000 commented 4 years ago

Hi,have you got the true ablation result ? can you show me the details ?