Mnist-Anomaly-detection
Goal
The Goal of this notebook is to identify anomalies in given dataset.
Data
the dataset contains 100 images. 90 from mnist, 10 from other dataset
Method
I used pre-trained VAE that has been trained on mnist dataset to generate new images.
In variational autoencoders, inputs are mapped to a probability distribution over latent vectors, and a latent vector is then sampled from that distribution. The decoder becomes more robust at decoding latent vectors as a result.
So, if I will have data from different distributions, the VAE won't succeed to reconstruct the image. And the loss will be big.
Like that, I can find anomalies if their loss will be bigger than normal data.
Pre-Trained
I Used pre-trained NN that i found here https://github.com/csinva/gan-vae-pretrained-pytorch