Autoencoder-classifier task

This is a basic CV problem.

Project Description

Problem Statement

rus:
Написать и обучить модель-автокодировщик на датасете на выбор: CIFAR10, CIFAR100.
Обучить модель-классификатор на латентных представлениях обученного автокодировщика.

eng:
Write and train an autoencoder model on a dataset of your choice: CIFAR10, CIFAR100.
Train a classifier model on latent representations of the trained autoencoder.

Data

To see full data statistics you can see run/data_exploration.ipynb. https://www.cs.toronto.edu/~kriz/cifar.html

Solution

In this project a few small models was trained on CPU to check the parametrs like dimentions of latent space classifier input.

For classifier training is on the "bottleneck" autoencoder latent space representations:

The scheme of solution pipeline:

Encoder architectures

Convolutional Autoencoder

Pros:

• Can handle spatial information effectively due to the convolutional layers.
• Can potentially provide better performance when trained on image data like CIFAR10 or CIFAR100.
• Easy to understand and implement.

Cons:

• May fail to generate new samples that are as diverse as the original dataset because it lacks a mechanism for ensuring that the distribution of latent variables has good coverage of the space.
• The structure and size of the network can greatly influence the model's ability to learn.

Variational Autoencoder:

Pros:

• Incorporates probabilistic encoders and decoders, and thus handles uncertainties better.
• Can generate new samples by drawing from the latent space.

Cons:

• More complex to implement due to the reparameterization trick.
• Sometimes can generate blurrier images compared to other generative models.
• The choice of the latent space dimension is critical and can significantly influence the model's performance.

Classifier

There is all-pairs testing for 2 autoencoder architectures and 1 classifier for different latent space dimention on both CIFAR10 and CIFAR100 normalised datasets.

Results

To see the final results of training process for all model pairs on different latent space dimentions you can see run/master_notebook.ipynb.

Example:

AE_loss for CIFAR10 dataset:

AE_loss for CIFAR100 dataset:

Classifier performance:

Structure

.                              
├── data                                # CIFAR10/CIFAR100 datasets
├── run 
│   ├── data_exploration.ipynb          # Explore dataset classes and structure
│   └── master_notebook.ipynb           # Full pipeline
├── reports                             # Solution plots
├── models                              # Trained models and checkpoints
└── src
    ├── data_loader.py                  # Load and preprocess the CIFAR10/CIFAR100 dataset
    ├── autoencoder.py                  # Autoencoder architectures
    ├── train.py                        # Autoencoder training loop
    ├── classifier.py                   # Classifier architectures and training procedure
    ├── baseline.py                     # Baseline solution
    └── utils.py                        # Useful functions