GGSN: assignments and final project

This repository holds four assignments and the final project (CWP) from the GGSN course (Deep and Graph Neural Networks) at AGH, covering classical machine learning, convolutional networks, sequence models, graph neural networks and neural-network compression.

Author: Jakub Kierznowski

Contents

File	Topic	Data	Key techniques
GGSN_Jakub_Kierznowski_Zad1.ipynb	House price prediction (Kaggle House Prices)	House Prices (Kaggle)	Feature engineering, DBSCAN (outlier removal), target encoding, a parallel neural network (Keras), XGBoost + LightGBM stacking, hyperparameter tuning with Optuna, blending
GGSN_Jakub_Kierznowski_Zad2.ipynb	Image classification (Intel Image Classification)	Intel Image Classification (Kaggle)	CNN from scratch, regularisation (BatchNorm/Dropout/L2), augmentation, transfer learning and fine-tuning (MobileNetV2, EfficientNetB0), Grad-CAM
GGSN_Jakub_Kierznowski_Zad3.ipynb	Tweet sentiment analysis (Sentiment140)	Sentiment140 (~1.6M tweets)	Text encoding (OHE, learned embeddings, GloVe, BERT tokenizer), RNN, LSTM, BiGRU, DistilBERT fine-tuning, K-fold CV, ensembling
GGSN_Jakub_Kierznowski_Zad4.ipynb	Graph neural networks on Yelp data	Yelp Open Dataset (restaurants)	Graph construction (geographic adjacency + shared users), GCN, GAT, GraphSAGE, GIN, GRNN; node classification, rating regression, cuisine-type classification, anomaly detection
GGSN_Jakub_Kierznowski_CWP.ipynb	Final project: neural-network compression	COCO 2017 (classification derived from detection)	Unstructured and structured pruning (magnitude, Network Slimming, He), quantisation (ACIQ, QAT, AdaRound/BRECQ), knowledge distillation, a combined pipeline (Deep Compression); compared across 7 architectures (ResNet, EfficientNet, MobileNet, DenseNet)

Material accompanying the CWP project:

• GGSN_Jakub_Kierznowski_CWP_prezentacja.pptx.pdf / Kompresja_Sieci_Neuronowych_prezentacja.pptx — the summary presentation
• articles_full.zip — the full set of source papers (20 entries) cited in the CWP notebook

The images first_part.png, second_part.png, parallel_model_architecture.png and result.png are diagrams and plots embedded in the Zad1 notebook (the parallel network architecture and the model results).

Environment

The base dependencies are in requirements.txt:

numpy, pandas, matplotlib, seaborn, tensorflow, scikit-learn, keras, shap, ipython, xgboost, scipy

GGSN_Jakub_Kierznowski_CWP.ipynb (network compression) additionally needs PyTorch + torchvision/timm — those experiments were run on a GPU, for reference an RTX 4070 Ti SUPER on CUDA 12.1 — and Zad4 needs graph neural network libraries (e.g. PyTorch Geometric). Install those manually before running either notebook.

Setup:

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
jupyter notebook

Data

The datasets are not included in the repository (apart from articles_full.zip); each notebook downloads its own data from the links given in its introductory section (Kaggle, COCO 2017, Yelp Open Dataset).