CSE 144 Final Project

Kaggle competition

Team

• Name:
• Name:

What's in the repo

Path	Description
CSE-144_Final_Project.ipynb	Training and inference notebook
submission.csv	Group predictions for Kaggle (1036 rows)
checkpoints/best_model_On_Kaggle.pt	Best weights used for submission (epoch 21)
data/sample_submission.csv	Kaggle format reference
requirements.txt	Python dependencies

Download locally: data/train/ and data/test/ from Kaggle (not committed).

Setup

pip install -r requirements.txt

Place competition data here:

data/
  train/          # folders 0..99
  test/           # unlabeled .jpg files
  sample_submission.csv

Quick submit

Upload submission.csv on the Kaggle Submit Predictions page. No training needed.

Notebook workflow

Open CSE-144_Final_Project.ipynb and run:

1. Setup — imports, paths, hyperparameters
2. Data — loaders and train/val split (needs data/train/)
3. Model — EfficientNet-V2-S with a custom classifier head
4. Training — phase 1 warms up the head (6 epochs), phase 2 fine-tunes the full model (up to 50 epochs with early stopping)
5. Plots — loss and accuracy curves (optional)
6. Submission — loads checkpoints/best_model_On_Kaggle.pt, runs 5-view TTA, writes submission.csv

To regenerate submission.csv without retraining, run setup → data → model → submission.

Training is slow on CPU. Lower BATCH_SIZE if you run out of memory.

Model summary

• Backbone: ImageNet EfficientNet-V2-S
• Validation: stratified split, 1 image per class (100 images)
• Augmentation: random resized crop, flip, color jitter, random erasing
• Phase 2: mixup/cutmix, label smoothing, AdamW, cosine LR schedule
• Inference: 5-view test-time augmentation
• Best val accuracy: 71% (epoch 21)
• Kaggle test accuracy: 76%
• Kaggle position: 40th (as of submission, see Kaggle_Pos.png)