CIFAR-100 image classification with ResNet

ResNet implementation trained on CIFAR-100 dataset. Achieved 73% test accuracy with test-time augmentation applied

Usage

Try on Google Colab

The easiest way to just run a few predictions and try the model @ Open Google Colab

Running locally

Python 3.10 is recommended. Clone the repo and install dependencies:

# PIP
pip install .

# Poetry
poetry install

Download model weights from Github release and put the file in ./bin folder

Prepare some image and run the inference:

# PIP
python -m image_classifier.main classify --image-path ./image.jpg

# Poetry
poetry run image-classifier classify --image-path ./image.jpg

Implementation details

Basic info
Neural net architecture	ResNet 18 with bottleneck blocks
Optimizer	SGD with CosineAnnelingLR scheduler
Optimizer params	Initial learning rate: 0.02. Weight decay: 0.0005. Momentum: 0.9
Loss function	Cross entropy loss

Data processing
Dataset	CIFAR-100
Batch size	128
Transforms	Upscaling to 64x64, applying TrivialAugmentWide and normalization

Test-time augmentation is used for inference to increase test accuracy by 3 percent

State of the model

The ~73% test accuracy is far from ideal. The model is currently overfitting, probably due to shortage of data to train on

The possible solution is to train on other additional larger dataset (e.g. ImageNet). Switching to more performant neural net architecture is also a great idea (e.g. PyramidNet)

References

• hf project with ResNet 18 hyperparams tuned for CIFAR-100 - https://huggingface.co/edadaltocg/resnet18_cifar100
• "Deep Residual Learning for Image Recognition" paper - https://arxiv.org/abs/1512.03385
• "How to Use Test-Time Augmentation to Make Better Predictions" - https://machinelearningmastery.com/how-to-use-test-time-augmentation-to-improve-model-performance-for-image-classification/