Deep Learning Course Homeworks

This repository contains the completed homework assignments for the Deep Learning course at SUT, instructed by Dr. Soleymani.

Notes:

• Each homework consists of both theoretical and practical parts. My solutions to the theoretical questions are available in the theoretical directory, formatted in LaTeX.

• Most notebooks are designed to run on both CPU and GPU:

  • GPU-Only Sections: Computationally intensive tasks are wrapped in conditions to run only when a GPU is available.
  • CPU Execution: If running on a CPU, the notebooks will attempt to load pretrained weights saved from a previous GPU run.
  • Pretrained Weights: At the beginning of the notebooks, there is a link to download the pretrained weights that I have saved.

HW1: Deep Learning Basics

• Optimization Methods: This notebook explores several optimization methods commonly employed in deep learning tasks and assesses their effectiveness using two test functions.

• NN From Scratch: In this notebook, we will implement fully connected networks using a modular approach. We will implement a forward and a backward function for each layer. The forward function will receive inputs, weights, and other parameters and will return both output and a cache object storing data needed for the backward pass. The backward pass will receive upstream derivatives and the cache object and will return gradients with respect to the inputs and weights.

• Pytorch Basics: This notebook will walk you through many of the important features of PyTorch.

HW2: Convolutional Neural Networks

• Vanilla CNN: In this notebook, we will learn about CNN's basics by working on a problem of great importance in computer vision: classifying images of cats and dogs :).

• ResNet and U-Net: In this notebook, we will implement the ResNet18 and U-Net using PyTorch, and use them for image classification and colorization on the CIFAR10 dataset.

• YOLO: In this notebook, we will implement the YOLO object detection algorithm using PyTorch. YOLO is a real-time object detection system that can recognize multiple objects in an image while also predicting their bounding boxes.

HW3: RNNs & Transformers

• RNN: In this notebook, we will implement both a vanilla RNN and an LSTM using PyTorch. Additionally, we will develop a sentiment analysis model by simply adding a linear layer to the hidden vector of an LSTM.

• Simple GPT: The objective of this notebook is to create and train a decoder-only model, which is a custom and scaled-down version of GPT, using a dataset consisting of dialogues from the TV series Friends.

• MLM and Sequence Classification with BERT: In this notebook, we use the "rotten_tomatoes" dataset from HuggingFace, which contains movie reviews, to perform two tasks: Masked Language Modeling (MLM) and building a classifier to differentiate between positive and negative reviews. We utilize a smaller, customized version of BERT to ensure efficient and feasible training.

• iTransformer: In this notebook, we implement both the inverted Transfromer (iTransformer) and the standard Transformer using PyTorch and compare their performance in time series forecasting.

HW4: Generative Models

• GAN and VAE: In this notebook, we are going to implement Variational AutoEncoder (VAE) and Generative Adversarial Network (GAN) on the MNIST dataset. VAEs learn a latent variable model from input data. They sample from this distribution and decode it to generate new data. GANs uses a generator to make images based on a prior distribution.

• Denoising Diffusion Probabilistic Models (DDPMs): In this notebook, we will implement both a Denoising Diffusion Probabilistic Model (DDPM) and a Conditional DDPM. For more details, please refer to the original DDPM paper.

HW5: More Recent Models and Topics

• DINO: In this assignment, we will use the self-supervised DINO model to extract visual features from satellite imagery. These features will be used to train a classifier to predict the presence of solar panels in the images. Additionally, we will analyze the transformer's attention maps from the DINO model to estimate the size of the solar panels in positive examples.

• Stable Diffusion: This notebook consists of three parts: implementing the stable diffusion pipeline with classifier-free guidance, enhancing blue color in the generated images, and fine-tuning the model using the DreamBooth technique for personalized image generation based on paired tokens and images.

• Adversarial Attacks and Training: This notebook explores adversarial attacks and training in deep learning. It covers training a ResNet18 model on the CIFAR-10 dataset, implementing FGSM and PGD attacks, generating adversarial examples, and conducting adversarial training. Finally, it compares the performance and robustness of the adversarially trained model with the original model.