README.md

ConvDAE-use-in-notMNIST

Win10 Python3.5 Tensorflow-1.1.0-gpu

This is a ConvDAE using in notMNIST dataset to denoise.

The notMNIST dataset you can find in https://github.com/hankcs/udacity-deep-learning.

The ConvDAE you can see in https://github.com/NELSONZHAO/zhihu/tree/master/denoise_auto_encoder.
The differences :

1. The loss function, we use the l2_loss, not the sigmoid_cross_entropy.
2. We use 5X5 kernel size with 64 filters in all convolutional layers.
3. Our structure is: conv->pool->conv->pool->resize->conv->resize->conv

Here is the result. The first row is add noisy images, the second row is the denoise images after ConvDAE processing, the third row is the original images.