This repository contains a series of hands-on labs exploring various deep learning paradigms, including supervised learning, reinforcement learning, and transformers. It was developed and tested locally on a high-performance workstation to ensure smooth execution and compatibility across environments.

The labs are structured as independent Jupyter notebooks and cover the following topics:

• Train MLPs and CNNs on MNIST and CIFAR-10
• Experiment with residual connections and knowledge distillation

• Implement REINFORCE and Deep Q-Network (DQN) agents
• Use OpenAI’s gymnasium environments for training and evaluation

• Perform sentiment analysis and feature extraction with Hugging Face Transformers (DistilBERT)
• Apply Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA

All experiments were run and validated locally using the following setup:

• GPU: NVIDIA RTX 4070 Ti Super (16 GB VRAM)
• CPU: AMD Ryzen 7 7800X3D
• RAM: 32 GB DDR5 (Corsair Vengeance)
• Storage: NVMe SSD Kingston

This configuration allows for fast model training and smooth interaction with large transformer models.

git clone https://github.com/marcopibbes/DLA_labs_1-2-3
cd LabDLA

⚠️ Note: Due to compatibility issues between libraries, you may need separate environments.

conda create -n dla_lab python=3.12.8 -y
conda activate dla_lab
conda install --file requirements.txt -c conda-forge -y

Every file is accurately commented (+ additional Markdown) in order to give the reader a clear view of the phases of the experiments!

.
├── Lab1-CNNs.ipynb           # Supervised learning (MLPs, CNNs)
├── Lab2-DRL.ipynb            # Reinforcement learning agents
├── Lab3-Transformers.ipynb   # Transformers & PEFT
├── BaseTrainingPipeline.py   # Core trainer class
├── SLTrainingPipeline.py     # Supervised training pipeline
├── RLTrainingPipeline.py     # REINFORCE trainer
├── QLTrainingPipeline.py     # DQN trainer
├── MLP.py                    # MLP and ResMLP definitions
├── CNN.py                    # CNN and ResCNN definitions
├── data/                     # Dataset storage (auto-downloaded)
├── checkpoints/              # Saved models
├── logs/                     # TensorBoard logs
└── requirements.txt          # Dependency list

1. Activate the appropriate Python environment.
2. Launch Jupyter:

jupyter lab  # or jupyter notebook

3. Open the desired notebook (Lab1-CNNs.ipynb, Lab2-DRL.ipynb, or Lab3-Transformers.ipynb) and start running the cells.

💡 Tip:
If no checkpoints are available, make sure to comment out pipeline.load() and use pipeline.train() instead.

Each notebook also defines a clear() function to clean temporary data (checkpoints, logs, datasets).

To visualize logs and monitor training:

tensorboard --logdir=./logs

Then open http://localhost:6006 in your browser.

All required packages are listed in requirements.txt, including:

• torch, torchvision, transformers, tensorboard
• gymnasium, datasets, scikit-learn, matplotlib, pandas
• Utilities for training, evaluation, and visualization

Some dependencies are version-pinned for compatibility and were installed via conda-forge.

• Expect first runs to download datasets and pretrained weights.
• Some visualizations (e.g., Lab 2) may require pygame.
• Tested across Python 3.9 and 3.12.8 environments.

• I found this strange yet useful method: autogenerating some comments using Copilot before the warmup code.
• This helped me to understand more about the context and some choices.
• Obviously, I made sure to have the most solid background possible.
• Told this both for transparency and to be helpful to people like me