task.md

Post-Training Checklist

• Generate Predictions & Calculate Metrics

  • Evaluate the ResNet-18 (Task 1.1) baseline model on the test set.
  • Evaluate the ResNet-18 (Task 1.1) FocalLoss ablation model on the test set.
  • Evaluate the ResNet-18 + SE Blocks (Task 1.2) baseline model on the test set.
  • Evaluate the ResNet-18 + SE Blocks (Task 1.2) FocalLoss ablation model on the test set.
  • Evaluate the DeiT-3 Small (Task 2.1) baseline model on the test set.
  • Evaluate the DeiT-3 Small (Task 2.1) FocalLoss ablation model on the test set.
  • Evaluate the DeiT-3 Small + DyT (Task 2.2) baseline model on the test set.
  • Evaluate the DeiT-3 Small + DyT (Task 2.2) FocalLoss ablation model on the test set.
  • Metric requirements: Accuracy, Macro F1-score, Macro ROC-AUC (using multi_class='ovr' and average='macro').

• Bonus Visualizations (Task 3)

  • Generate Grad-CAM visualizations (Task 3.1) for ResNet-18 (Task 1.1).
    • Requirement: Run on a sample image for each of the 10 classes in the test set.
  • Generate Grad-CAM visualizations (Task 3.1) for ResNet-18 + SE Blocks (Task 1.2).
    • Requirement: Run on a sample image for each of the 10 classes in the test set.
  • Generate Attention Map visualizations (Task 3.2) for DeiT-3 Small (Task 2.1).
    • Requirement: Run on a sample image for each of the 10 classes in the test set.

• Formatting the Codebase

  • Create separate training and testing scripts for each subtask.
  • Export any .ipynb notebooks used to .py scripts.
  • Generate a requirements.txt containing all dependencies.
  • Create a README.md with:
    • Clear instructions to run each training/testing script.
    • Usage of command line arguments for paths/hyperparameters (No hardcoded paths).
    • Example commands for training and testing every task.

• Finalizing the Submission ZIP (<EntryNumber>.zip)

  • Include all .py training/testing scripts.
  • Include report.pdf.
  • Include requirements.txt.
  • Include README.md.
  • Verify dataset is not included.
  • Verify model weights are not included if > 25MB (use file-sharing service link in report instead, timestamped before deadline).

---

Report Specifications (report.pdf)

To ensure full marks as per the rubric, your report MUST include the following sections:

1. Metric Reporting

For every task (1.1, 1.2, 2.1, 2.2) and their ablations (Focal Loss), present the following metrics on the test set:

• Accuracy
• Macro F1-score
• Macro ROC-AUC (multi_class='ovr' and average='macro')

2. Model Comparisons & Analysis

• Task 1.2 vs 1.1: Compare the test results of ResNet-18 with SE blocks against the baseline ResNet-18.
• Task 2.1 vs Task 1: Compare the DeiT-3 Vision Transformer results against the CNN-based models.
• Task 2.2 vs Task 2.1: Compare the performance of the Transformer without Normalization (DyT) against the standard DeiT-3 model.

3. Visualization Analysis (Bonus Tasks)

• Grad-CAM Analysis (Task 3.1): Present the generated Grad-CAM images (1 for all 10 classes) from models trained in 1.1 and 1.2. Write an analysis of the regions the models refer to in making their predictions.
• Attention Map Analysis (Task 3.2): Present the generated attention maps (1 for all 10 classes) from the model trained in 2.1. Write an analysis mapping which regions the model attends to while making final predictions.

4. Academic Formatting & Integrity

• Cite all your sources properly using a standard citation format. Ensure anti-plagiarism compliance.

5. Weights Access Linking

• If your trained model weights exceed the 25MB ZIP limit, include a publicly accessible download link to the weights directly inside the report. (Ensure the timestamp on the uploaded files is strictly before the submission deadline: 11:59PM, 10th April, 2026).