Hi,
I trained a denoising autoencoder on mnist dataset corrupted by random gaussian noise and built a classifier with encoder as backbone(The layers are freezed), and a linear classifier layer trained on mnist dataset corrupted by random gaussian noise. Its performance on test dataset is 10%. May I know how to improve this network. I am sharing the code.
https://colab.research.google.com/drive/1fnLJxtzuT2cNeWXycjn3_pVej5DE3z-y?usp=sharing