This repository contains a university project where we implement a data selection method known as "Importance sampling based on loss". Effectively, we reweight the data sampling distribution for training batches on gradient descent in favour of datapoints that yield the highest training loss. We experimented on the CIFAR10 dataset and found that the while the convergence speed improves slightly, it does not enhance test accuracy and ultimately there are more valuable methods to obtaining better convergence and accuracy. For contrast, we also show the effect data augmentation has on the accuracy (see below plot and report for more details).

To run the code, you will need to install all packages in the requirements.txt, which can be done by:

 pip install -r requirements.txt

Note that torch==1.12.1+cu113 version inside assumes CUDA 11.3 is installed. If using a different CUDA version, then replace these dependencies with the corresponding ones.

torch==1.12.1+cu113
torchaudio==0.12.1+cu113
torchvision==0.13.1+cu113

The main file that trains models is train.py. The additional files have the following purposes:

• arg_parser.py - parses commandline arguments (see next section)
• datasets.py - contains datasets to train on
• models.py - contains model architectures to train with
• samplers.py - contains sampling methods to experiment with
• tutorials/* - a dumping folder for tutorial code from other sources

The code is designed to be run a commandline where the full extent of commands can be seen from python train.py new -h. Below are two example commands:

This runs the baseline standard mini-batch sampling

python train.py new --model-name ResNetMNIST --sampler-name RandomSamplerBase --training-runs 3 --wandb

This runs a sampling using the most recent loss

python train.py new --model-name ResNetMNIST --sampler-name LastLossWeightedSampler --sampler-config "{'num_samples':10000, 'replacement':true}" --training-runs 3 --wandb

The project uses wandb to centrally track training metrics with results visible here. Using it we can aggregate results and output them in whatever format we like.