ResNet10 Run Report

Summarized Findings

We introduced MixUp-GradCam loss from Yu-Ting Chang et al. [1], but this was not particularly useful in our top-1 classification task. We saw little improvement over our baseline.

Possible Improvements.

In our next implementation:

• We’ll replace our augmentation pipeline with off-the-shelf TrivialAugment
• We’ll introduce label-smoothing for additional regularization
• We’ll introduce CosineAnnealingWarmRestarts

Notebook

Please see the companion notebook. It contains:

• Training Confusion Matrix
• Validation Confusion Matrix
• Training Classification Report
• Validation Classification Report
• Training GradCam
• Validation GradCam