Cifar Image Reconstruction using Autoencoder Models

Take 1000 images from CIFAR10. Then, given a pair of images from CIFAR10 𝑥1 and 𝑥2, build a network that can return both images given their average (𝑥1+𝑥2)/2 as the only input. The design of the architecture and of the input-output mapping is your choice.

Pipline

1. At first we need to generate dataset. for building the whole dataset we need to generate 1000*999/2 =(550000) picture that is large dataset. so we use the subset of this whole dataset. we generate the 50 image from every image that is combined by 50 different other image for all of the 1000 images.
2. At second task we need to config the dataloader class that should return image , (image_label1 , image_label2). Actually we have 2 label for each dataset that expect the model to create them in the output.
3. At the third step we should create the autoencoder model by pytorch that input is 32*32 and the output is 2*32*32 which remind us some state-of-the-art models like U-net model and ... . We use some basic models to do it . Let's see our models that have been use to reach the best result: