SUR project

This repository contains the code for the SUR project.

Classification problem

In this project we had to create a 2 classifiers for custom dataset with very few examples without using outside data. Data contains 2 classes: target and non-target. We had to create a classifier that would be able to classify the data into these 2 classes. Image data has 3 channels and is in RGB format. Their size is 80x80 pixels. All images are closeups of faces.

Dataset:

• train: 151 examples
  • target: 20
  • non-target: 131
• test: 70 examples
  • target: 10
  • non-target: 60

Eval dataset: This data was provided to us after submitting the final model.

• eval: 1342 examples

Models:

1. ResNet9 - 9 layer ResNet model - 6 571 266 params
2. PreActResNet18 - 18 layer PreActResNet model - 11 167 042 params
3. PreActResNet34 - 34 layer PreActResNet model - 21 275 202 params
4. ConvMixer - ConvMixer model (substantially worse than the other 2 models)

They key was to use heavy data augmentation to increase the number of examples in the dataset. Here are the augmentations used:

TRANSFORM_TRAIN: transforms.Compose = transforms.Compose([
                        transforms.RandomCrop(80, padding=8, padding_mode='constant'),
                        transforms.GaussianBlur(3, sigma=(0.01, 1)),
                        transforms.RandomRotation(degrees=(0, 8)),
                        transforms.RandomGrayscale(),
                        transforms.RandomAutocontrast(),
                        transforms.RandomAdjustSharpness(sharpness_factor=2),
                        transforms.RandomPosterize(4),
                        transforms.RandomEqualize(),
                        transforms.Lambda(lambda x: cutout(x, 17, 17)),
                        transforms.RandomHorizontalFlip(),
                        transforms.ToTensor(),
])

All the hyperparameter configurations can be found in src/config.py. Code for models are stores in src/models/ and the training code is in src/trainmodel.py.

Results

Best models from checkpoints:

[ResNet9]           - Test Loss: 0.01912, Test Acc: 1.0000, Train Loss: 0.00067, Train Acc: 1.0000
[Preact ResNet18]   - Test Loss: 0.00766, Test Acc: 1.0000, Train Loss: 0.00012, Train Acc: 1.0000
[Preact ResNet18 A] - Test Loss: 0.00539, Test Acc: 1.0000, Train Loss: 0.00006, Train Acc: 1.0000
[Preact ResNet34 A] - Test Loss: 0.00009, Test Acc: 1.0000, Train Loss: 0.00002, Train Acc: 1.0000

Eval results:

[ResNet9]           - Eval Loss: 0.38013, Eval Acc: 0.9426, Error: 0.0574
[Preact ResNet18]   - Eval Loss: 0.45691, Eval Acc: 0.9523, Error: 0.0477
[Preact ResNet18 A] - Eval Loss: 0.36617, Eval Acc: 0.9545, Error: 0.0455
[Preact ResNet34 A] - Eval Loss: 0.55418, Eval Acc: 0.9553, Error: 0.0447

How to run

The main code is in jupyter notebook image_cnn.ipynb. The code is written in Python 3.11.7.

Authors

• Jan Holáň
• Matej Horník