trshstar/VisionCorruptor

VisionCorruptor

VisionCorruptor is an end-to-end image corruption classification pipeline built using PyTorch. It classifies 64×64 images from the Tiny ImageNet-C dataset into 15 corruption types (e.g., gaussian_noise, motion_blur, jpeg_compression) using a CNN architecture.

Includes:

Custom PyTorch Dataset for Tiny ImageNet-C (non-standard nested structure)
ResNet-18 training with early stopping, validation tracking, and test evaluation
Visualization: loss/accuracy plots, confusion matrix, sample images

📊 Results Summary

Model: ResNet-18

Validation Accuracy: 97.37%

Test Accuracy: 97.28%

Parameters: ~11M

Notes: Baseline CNN

Best model checkpoint: best_model.pth

Trained on 15 corruption types from Tiny ImageNet-C

mage size: 64×64 | Batch size: 64 | Optimizer: Adam | Early stopping: 3 epochs