Deep Learning 1 Assignments at UvA, edition 2022

This repository contains the practical assignments of the Deep Learning 1 course at the University of Amsterdam.

Course website: https://uvadlc.github.io/
Course edition: Fall 2022
Course taught by: Dr. Yuki M. Asano

Assignment 1 - MLPs and Backpropagation

In the first assignment, the equations of backpropagation for basic modules in a vanilla neural network are derived from scratch, and then implemented as modules in order to create a simple multi-layer perceptron (MLP). The dataset used for this assignment is CIFAR-10. Model accuracies, hyperparameters, and other statistics can be automatically generated using compare_models.py.

Assignment 2 - Transfer Learning for CNNs, Visual Prompting, Graph Neural Networks

The assignment is organized in two parts. The first part covers transfer learning for CNNs, where you fine-tune an existing network to adapt it to a new dataset. In the second part, you prompt CLIP to perform image classification, both in the zero-shot setting and by learning a visual prompt. The corresponding results can be found under the results directory.

Deliverables for Transfer Learning with CNNs

• Compare popular CNN architectures on ImageNet using data from the PyTorch website.
• Use the official PyTorch implementations to compute and compare the inference speed, memory usage, and parameter count for the same models.
• Adapt a ResNet-18 model from ImageNet to CIFAR-100 by resetting its last layer.
• Perform augmentations to improve performance.

Deliverables for Visual Prompting

• Evaluate CLIP-B/32 in the zero-shot setting on CIFAR-10 and CIFAR-100 (clipzs.py).
• Prompt CLIP on two new downstream tasks by changing the text template (--prompt) and class labels (--class_names). You can visualize your predictions with --visualize_predictions.
• Learn visual prompts for the CIFAR-10 and CIFAR-100 datasets (learner.py, vpt_model.py, vp.py).
• Experiment with the prompt design to get near-random performance on CIFAR-100.
• Evaluate the robustness of the learnt prompts to distributional shifts (robustness.py).
• Evaluate each dataset's learnt prompt on the concatenation of both CIFAR-10 and CIFAR-100 (cross_dataset.py).

Assignment 3 - Deep Generative Models

The final assignment contains 2 parts. Part 1 is to implement in pytorch a variational autoencoder (VAE), and part 2 an adversarial autoencoder (AAE).

For the VAE part, we will train the model on generating 4-bit MNIST images. The original MNIST dataset contains images with pixel values between 0 and 1. To discretize those, we multiply pixel values with 16 and map the result to the closest integer value (rounding down 16 to 15). This is a 4-bit representation of the original image. Standard RGB images are usually using 8-bit encodings (i.e. values between 0 and 255), but to simplify the task, we only use 4 bits here.

For the AAE part, We will train the model on generating MNIST images.