Deep Learning for Computer Vision

Companion repository for 4F12: Computer Vision (Engineering Tripos Part IIB, University of Cambridge).

Lectures

#	Topic	Video
1	Multi-layer Perceptrons	YouTube
2	Convolutional Neural Nets	YouTube

Notebooks

Interactive lecture notes are provided as Jupyter Notebooks:

• lecture_notes_0.ipynb — Perceptrons, Classification, Backpropagation, Gradient Descent, Optimisation
• lecture_notes_1.ipynb — Convolution, Pooling, ResNet, Segmentation, VAE, Object Detection, Face Recognition, Adversarial Attacks
• lecture_notes_2.ipynb — RNNs, LSTM, Transformers, Vision Transformers, Generative Pretraining, CLIP, DINOv2

Topics Covered

Topic	Module(s)
Perceptrons & MLPs	perceptron.py, multiperceptron.py, mlp.py
Activations	activations.py
Gradient Descent & Optimisation	gradient_descent.py
Convolution & Pooling	filters.py, cnn.py, pooling.py
ResNet	resnet.py
Salience Maps	salience.py
Datasets (MNIST, CIFAR-10, LFW, EMNIST, COCO)	datasets.py
Variational Autoencoders	vae.py
FaceNet & Triplet Loss	facenet.py
Object Detection (YOLO)	yolo.py
Adversarial Examples (FGSM)	adversarial.py
Recurrent Neural Networks	rnn.py
Transformers (ViT, GPT, Encoder-Decoder)	transformers_model.py
Patch Clustering (K-means Tokenisation)	patch_clustering.py
CLIP (Zero-shot Classification)	clip.py
DINO / DINOv2 (Self-supervised Vision)	dino.py

Getting Started

It is recommended to use a virtual environment:

python -m venv .env
source .env/bin/activate   # Linux/macOS
# .env\Scripts\activate    # Windows

Install dependencies:

python -m pip install pip --upgrade
pip install -r requirements.txt

Then follow the instructions to install PyTorch.

Pre-computed Results

The notebooks rely on pre-computed training results (saved weights, embeddings, etc.). To download them:

python download_results.py

This downloads ~69 MB of results files into results/. Use --force to re-download existing files.

If you prefer to generate the results yourself, run the corresponding Python modules directly (e.g. python facenet.py), though this requires a GPU and considerably more time.

Warning — regeneration dependencies: Some data and results files depend on each other. Regenerating them will invalidate downstream results, requiring those to be retrained as well.

If you regenerate…	…you must also regenerate
data/emnist_patch128.npz	transformer_emnist_gpt_128.results, transformer_emnist_128_finetune.results, rnn_emnist_patch_128.results
data/lfw_minitrain.npz	facenet.results
transformer_language_128.results	transformer_language_128_finetune.results
transformer_emnist_gpt_128.results	transformer_emnist_128_finetune.results

The scripts can be run standalone, but are primarily designed to be called from the Jupyter notebooks. Learn more about Jupyter.

Requirements

See requirements.txt. Key dependencies: PyTorch ≥ 2.0, torchvision, matplotlib, numpy, ultralytics (YOLO), open-clip-torch (CLIP).

License

See LICENSE.