Deep Learning – Project 1

Image classification with convolutional neural networks

Codes for the Deep Learning project 1 about CNNs on CIFAR-10.

Codes description

Folder models contains the implementation of the models used in the project. The models are implemented as classes that inherit from torch.nn.Module.

Script train_model.py contains a class that you can pass model, dataset and training hyperparameters and it will train the model and return the results. It is used by tests_script.py.

Script tests_script.py can be run from terminal to train a model of a given architecture with a given set of hyperparameters and upload the results to neptune.ai. It has a parameter to repeat the training multiple times and also allows for trying different augmentation methods.

Folder mixup_scripts contains scripts for training models with mixup augmentation as they required a few changes in the training loop.

Other files

ensembles* files are related to creating ensembles and evaluating their accuracy.

kaggle* files are related to the Kaggle CIFAR-10 competition. The predicted labels achieved a score of 0.8119

Folder dl_trained_models contains some of the models used for ensembling. Bigger models (ResNet18 architecture) were not uploaded.

Folder plots contains all the charts created during the project.

export.csv contains most important results exported from neptune.ai.

Reproducibility

This section contains description of how to repeat the experiments using the provided scripts.

Environment

Experiments were run mainly in following settings:

Ubuntu 20.04, Python 3.10.9, PyTorch 2.0.0, CUDA 11.7

Code was also tested in following settings:

Ubuntu 22.04, Python 3.9.16, PyTorch 2.1.0, CUDA 11.7
Debian 10, Python 3.9.16, PyTorch 1.12.1, CPU only

requirements.txt and environment.yml files contain the list of required packages installed in the first of the environments listed above.

Using pip:

pip install -r requirements.txt

Using conda:

conda env create -f environment.yml

How to run the experiments

Script experiment_baseline.sh runs training of baseline models with default parameters and repeates the training 5 times. It also uploads the results to neptune.ai. Default parameters include learning rate = 0.003, batch size = 16 and number of epochs = 25. Run with:

./experiment_baseline.sh

Script experiment_hyperparameters.sh runs training of the pretrained ResNet18 with different hyperparameters and augmenation methods. It also uploads the results to neptune.ai. Run with:

./experiment_hyperparameters.sh

You may directly run the tests_script.py script with different parameters. To see the examples, open experiment_baseline.sh.

There is no separate script for ensembles. They were created and evauated in the ensembles.ipynb Jupyter Notebook.

The codes for final plots, which were created in order to study the influence of hyperparameters and data augmentation, are available in summary.ipynb notebook.