🧠 Deep Learning — M.Sc. Coursework

From-scratch implementations of foundational and modern deep learning systems — including Transformers, vision-language models (CLIP, DINO), diffusion models (DDPM), and contrastive self-supervised learning, built in NumPy and PyTorch.

---

✨ Highlights

Area	What's implemented
🔢 Foundations	Backprop, BatchNorm, Dropout, Conv — all from scratch in NumPy
🔤 Sequence & Attention	Transformer from scratch, RNN/LSTM image captioning on COCO
🖼️ Vision-Language (Multimodal)	CLIP zero-shot classification, DINO self-supervised ViT
🎨 Generative Models	DDPM with classifier-free guidance, U-Net denoiser, EMA training
🔍 Self-Supervised Learning	SimCLR contrastive loss + augmentation pipeline

---

📂 Repository Structure

assignments/
├── assignment1/   # Neural Network Fundamentals (NumPy → PyTorch)
├── assignment2/   # Sequence Models & Vision-Language
└── assignment3/   # Generative Models — DDPM Diffusion

---

Assignment 1 — Neural Network Fundamentals

Builds the core building blocks of neural networks from scratch in NumPy, then reimplements them in PyTorch.

Notebook	Topics
01_softmax.ipynb	Softmax classifier, cross-entropy loss, analytic gradients
02_two_layer_net.ipynb	Two-layer fully connected network, backpropagation
03_FullyConnectedNets.ipynb	Modular layer design, arbitrary-depth FC nets, SGD / Adam / RMSProp
04_Dropout.ipynb	Inverted dropout (forward & backward passes)
05_BatchNormalization.ipynb	Batch norm, layer norm, group norm
06_ConvolutionalNetworks.ipynb	Convolution (im2col), max pooling, spatial batch norm
07_PyTorch.ipynb	PyTorch re-implementation of all the above

Key implementations (dl/):

• Forward & backward passes for affine, ReLU, softmax, conv, pooling, dropout, and all normalization variants
• Modular Solver class with pluggable optimizers and learning-rate scheduling
• Fully connected and convolutional network classifiers

---

Assignment 2 — Sequence Models & Vision-Language (Multimodal)

Implements the full spectrum from recurrent models to modern vision-language architectures.

Notebook	Topics
rnn_lstm_captioning.ipynb	Vanilla RNN, LSTM, attention mechanism — image captioning on COCO
Transformers.ipynb	Transformer from scratch — self-attention, multi-head attention, encoder/decoder, positional encoding
Self_Supervised_Learning.ipynb	SimCLR contrastive learning framework
CLIP_DINO.ipynb	CLIP zero-shot classification, DINO self-supervised ViT feature extraction

Key implementations:

• rnn_lstm_captioning.py — RNN/LSTM cells, temporal softmax loss, CaptioningRNN with attention
• transformers.py — Scaled dot-product attention (loop & vectorized), SelfAttention, MultiHeadAttention, LayerNorm, FeedForwardBlock, full encoder/decoder Transformer
• dl/simclr/ — SimCLR contrastive loss and data augmentation pipeline
• dl/clip_dino.py — CLIP and DINO feature extraction, zero-shot evaluation

💡 CLIP is a multimodal model that jointly embeds images and text — this notebook covers zero-shot image classification using OpenAI's pretrained CLIP. DINO uses self-supervised ViT representations without labels.

---

Assignment 3 — Generative Models (DDPM)

Full implementation of Denoising Diffusion Probabilistic Models for conditional image generation.

Notebook	Topics
DDPM.ipynb	Gaussian diffusion process, U-Net denoiser, classifier-free guidance, emoji generation

Key implementations (dl/):

• gaussian_diffusion.py — Forward diffusion (q-sample), reverse sampling (p-sample), noise scheduling (linear / cosine / sigmoid), training loss
• unet.py — U-Net with sinusoidal time embeddings, ResNet blocks, context conditioning, classifier-free guidance
• ddpm_trainer.py — Full training loop with EMA (Exponential Moving Average)

---

🛠 Tech Stack

Tool	Usage
NumPy	From-scratch neural network layers (Assignment 1)
PyTorch	All deep learning models (Assignments 1–3)
Torchvision	Pretrained feature extractors for captioning & CLIP/DINO
Jupyter Notebooks	Experiments, visualizations, and analysis

---

🚀 Getting Started

# Clone the repository
git clone https://github.com/TalSomech/Deep-Learning-Assignments.git
cd Deep-Learning-Assignments

# Install core dependencies
pip install numpy torch torchvision matplotlib jupyter

# Assignment 3 has additional requirements
pip install -r assignment3/requirements.txt

# Launch notebooks
jupyter notebook

Note: Some notebooks download datasets (CIFAR-10, COCO) on first run.