search

microsoft / CNTK

Microsoft Cognitive Toolkit (CNTK), an open source deep-learning toolkit
https://docs.microsoft.com/cognitive-toolkit/
Other
17.53k stars 4.28k forks source link

Denoising autoencoder #2524

Open L-Ramos opened 7 years ago

L-Ramos commented 7 years ago

I want to create a denoising autoencoder using CNTK deconv and Unpooling functions.

The example for autoencoders only use the input image for comparison with the output from the autoencoder: (err = z - features)

Several papers report that the autoencoders easily learn the identity function.

Since I'm interested in extracting features from those images I added noise to them to make the autoencoder more robust and prevent it from learning the identity function.

My current problem is, the output from the autoencoder shoud be compared to the normal image, without noise, but I don't know how to load both datasets (with and without noise).

I'm using brainscript and the textfilereader

cha-zhang commented 7 years ago

First I suggest you use the Python API instead of BrainScript.

Loading two datasets shouldn't be an issue. If you are using txt reader, follow the CTF format here: https://docs.microsoft.com/en-us/cognitive-toolkit/brainscript-cntktextformat-reader You case should be just the classification case.