KL_coeff = 0 and encoded dimension
Created on July 14|Last edited on July 21
Comment
- I compared the performance in terms of reconstruction of the models with kl_coefficient = 0, varying the encoded dimension and testing the standard CNN architecture against a different architecture inspired to ResNet18 (both encoder and decoder), including skip connections (NB: to remain faithful to the original implementation this has more filters and more downsampling for now).
- The results show a clear correlation between the encoded dimension and the reconstruction capabilities of the model, perhaps implying that more dimensions that we originally assumed are necessary for this dataset.
- I notice a difference between in the reconstructions in that the CNN images tend to look more 'blurry' while the resnet outputs present more 'pixelated' images.
- Although the reconstruction loss is higher for the resnet-inspired architecture, the generations seem to have at least capture some features (tonality)
Add a comment