Image-Classifier

General Information

Project:

• Main purpose of that project was to learn how Convolutional Neural Networks work and how to build custom one.
• We also evaluated existing and well-known solutions as residual networks family (ResNets) and ViTbase (based on transformers) in topic of transfer-learning
• Results and conclusions are stored in short report - "Projekt 5 - GOLEM CNN.pdf"

Authors:

• Jakub Kwaśniak
• Jakub Mieczkowski

Run and locally reproduce the results

prepare environment

# Ubuntu / MacOS
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

run script / reproduce results

I option:

python3 main.py --optim 'SGD' --loss 'CrossEntropy' --backbone 'resnet18'

• optim - optimizer chosen from - "SGD", "RMS", "ADADELTA", "ADAMAX"
• loss - loss function - Cross Entropy Loss if flag is equal to "CrossEntropy", Mean Square Error otherwise
• backbone - model's backbone - can be chosen from pretrained ("ViT", "resnet18", "resnet34") or our custom one if other flag than one of shown is passed

II option :

• Run notepad.ipynb prepared for models tests and evaluation (all results collected as plots)