IFT6266 - Convolutional Auto-Encoder
This CAE is for the class project for IFT6266 : Conditional Image Generation
[Insert Image Here]
You need to modify the paths in the code to point to you inpainting folder that should contain:
- train2014/ - Images for training. (64x64 RGB - not cropped)
- val2014/ - Images for validation.
- *.pkl - Pickle file with labels.
The Grayscale Images will be excluded by the script. The Images will be cropped by the script.
It seems that more training could be good if the features learned where actually good, because when it's wrong, less training gives a better image by just roughly inpainting the context colors... Wrong guesses trying to paint something that doesn't fit makes it worse. Next step would be to try a deeper network, to learn more features.
Most of this work is build on top of Massimiliano Comin's work Thanks for putting simple code that actually works!