ML2022Spring-HW3b-Food-11-Image-Classification

KO | EN

gitlite — search

#javascript #python #hacktoberfest #typescript #android #machine-learning #deep-learning #linux #docker #java #windows #awesome #mcp #open-source

ML2022Spring-HW3b-Food-11-Image-Classification

(★ 26)

Our imbalance-aware ViT model achieves 0.91035 accuracy on the public leaderboard and 0.87750 on the private leaderboard of the ML2022Spring-HW3b Kaggle competition.

ML2022Spring-HW3b-Food-11-Image-Classification Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# ML2022Spring‑HW3b ‑ Food‑11 Image Classification (Imbalance‑Aware ViT)

> **Competition** – [ML2022Spring‑HW3b on Kaggle](https://www.kaggle.com/competitions/ml2022spring-hw3b)  
> **Focus** – Tackle **severe class imbalance** in Food‑11 with dual‑strategy re‑balancing (loss weighting + oversampling) on a ViT‑B/16 backbone.  
> **Framework** – PyTorch 1.10 + TorchVision ViT  
> **Leaderboard**

| Public | Private |
| :----: | :-----: |
| **0.91035** | **0.87750** |

---

## Key Features

* **Class‑Imbalance Remedies** – inverse‑freq weighted loss *and* `WeightedRandomSampler` oversampling.  
* **Custom Classifier Head** – 1024‑512‑11 MLP with GELU & dropout.  
* **Augmentation Policy** – category‑aware strong aug for minority classes.  
* **Early‑Stopping + LR Plateau Scheduler** – stable convergence in ≤ 3 h on a single 24 GB GPU.  
* **One‑Click Submission** – `test.py` generates `submission.csv` & visualises predictions.

---

## Table of Contents

- [ML2022Spring‑HW3b ‑ Food‑11 Image Classification (Imbalance‑Aware ViT)](#ml2022springhw3b--food11-image-classification-imbalanceaware-vit)
  - [Key Features](#key-features)
  - [Table of Contents](#table-of-contents)
  - [Project Structure](#project-structure)
  - [Environment](#environment)
  - [Quick Start](#quick-start)
    - [1 Download Data](#1download-data)
    - [2 Train (handles imbalance automatically)](#2train-handles-imbalance-automatically)
    - [3 Infer \& Create `submission.csv`](#3infer--create-submissioncsv)
  - [Methodology](#methodology)
  - [Reproduce My Scores](#reproduce-my-scores)
  - [Results \& Logs](#results--logs)
  - [License](#license)
  - [Acknowledgements](#acknowledgements)

---

## Project Structure

```text
.
├── vit_base.py              # training / fine‑tuning script
├── test.py                  # inference & CSV generation (Kaggle submission)
├── data/                    # Food‑11 dataset root
│   ├── training/
│   ├── validation/
│   └── test/
├── output/                  # checkpoints, logs, figures (auto‑created)
├── LICENSE                  # MIT license
└── README.md
```

---

## Environment

```bash
conda create -n food11_vit python=3.9 -y
conda activate food11_vit
```

Main deps: PyTorch ≥1.10, TorchVision ≥0.11, albumentations, scikit‑learn, matplotlib.

---

## Quick Start

### 1 Download Data

```bash
kaggle competitions download -c ml2022spring-hw3b -p data/
unzip data/ml2022spring-hw3b.zip -d data/
```

### 2 Train (handles imbalance automatically)

```bash
python vit_base.py   --data_path data   --output_dir output/vit_b16   --epochs 100   --batch_size 32   --balance_method both   --pretrained
```

### 3 Infer & Create `submission.csv`

```bash
python test.py   --data_path data   --output_dir output/vit_b16   --batch_size 32
```

`test.py` loads `output/vit_b16/best_model.pth`, writes `submission.csv` & `test_predictions.json`, and saves a preview figure `sample_predictions.png`.

---

## Methodology

| Component | Setting |
|-----------|---------|
| **Backbone** | ViT‑B/16 (`torchvision.models.vit_b_16`) |
| **Imbalance Handling** | 1️⃣ Inverse‑freq class weights in CE loss  2️⃣ `WeightedRandomSampler` oversampling |
| **Augmentation** | Resize‑224 → Normalize; minority classes additionally get flip/rotate/ColorJitter/GaussianBlur |
| **Classifier Head** | 1024 → 512 → 11 MLP + GELU + 0.5/0.3 dropout |
| **Optimizer** | AdamW (1 e‑4, wd 1 e‑2) |
| **Scheduler** | ReduceLROnPlateau (factor 0.5, patience 5) |
| **Early Stop** | patience 10 |
| **Metrics** | Accuracy, confusion matrix |
| **Hardware** | 1× RTX 4090 24 GB, 100 epochs ≈ 3 h |

---

## Reproduce My Scores

```bash
git clone https://github.com/<your-name>/food11-imbalance-vit.git
cd food11-imbalance-vit
pip install -r requirements.txt

# train
python vit_base.py --data_path data --output_dir output/run --epochs 100 --batch_size 32 --balance_method both

# inference
python test.py --data_path data --output_dir output/run --batch_size 32
```

---

## Results & Logs

| Split | Top‑1 |
|-------|-------|
| Public LB | 0.91035 |
| Private LB | 0.87750 |

Detailed metrics & curves are stored in `output/*`.

---

## License

This project is released under the **MIT License** – see the [LICENSE](LICENSE) file for details.

---

## Acknowledgements

* NTU 李宏毅老师 & TA team for the Food‑11 assignment scaffold  
* TorchVision ViT implementation  
* Community discussions on imbalance learning in vision

ML2022Spring-HW3b-Food-11-Image-Classification

ML2022Spring-HW3b-Food-11-Image-Classification Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

ML2022Spring-HW3b-Food-11-Image-Classification

ML2022Spring-HW3b-Food-11-Image-Classification Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README