README.md

metadata

title: Food Spoilage Detection
emoji: 🍎
colorFrom: green
colorTo: red
sdk: docker
app_file: app.py
pinned: false

siPeka — Food Spoilage Detection API

A REST API that classifies fruits and vegetables as fresh or spoiled using a fine-tuned Swin Transformer Tiny.

Live demo: huggingface.co/spaces/crtal/siPeka-model-api

---

Tech Stack

Layer	Technology
API Framework	FastAPI
Model Architecture	Swin Transformer Tiny (torchvision)
Training	PyTorch 2.10 · AdamW · Cosine Annealing LR
Inference	CUDA / CPU (auto-detected)
Containerization	Docker
Deployment	Hugging Face Spaces
CI/CD	GitHub Actions

---

API Reference

GET /

Returns basic API info.

Response

{
  "message": "Food Spoilage Detection API is running.",
  "usage": "POST an image to /predict",
  "supported_classes": ["fresh", "spoiled"]
}

---

GET /health

Health check endpoint.

Response

{ "status": "ok" }

---

POST /predict

Classifies an image as fresh or spoiled.

Request

Field	Type	Description
file	multipart/form-data	Image file (.jpg, .jpeg, .png)

Constraints

• Max file size: 20 MB
• File must be a valid image readable by PIL

Responses

200 OK — Successful prediction

{
  "status": "success",
  "prediction": "fresh",
  "confidence": 0.95,
  "model": "swin_transformer_tiny_finetuned"
}

200 OK — Low confidence (image may not be a food item)

{
  "status": "ambiguous",
  "confidence": 0.61,
  "message": "Model confidence too low. The image might not be a recognized food item or is too low quality.",
  "suggestion": "Please provide a clearer image of a fruit or vegetable."
}

400 Bad Request — Invalid input

{ "detail": "File too large. Maximum size is 20 MB." }
{ "detail": "Could not read image. Please upload a valid image file." }

500 Internal Server Error — Inference failure

{ "detail": "Error processing image: <reason>" }

Example

curl -X POST "https://crtal-sipeka-model-api.hf.space/predict" \
     -F "file=@apple.jpg"

---

Run Locally

git clone https://github.com/yukienjoyer7/sipeka-model-api.git
cd sipeka-model-api
pip install -r requirements.txt
uvicorn app:app --host 0.0.0.0 --port 7860

---

Training

1. Prepare data

python scripts/prepare_data.py

Restructures the raw dataset into binary fresh / spoiled classes and splits into train / val / test (70 / 15 / 15).

2. Train

python scripts/train.py \
  --arch swin_t_finetune \
  --output models/spoilage_model_swin_finetuned.pth \
  --epochs 20 \
  --lr 0.0001 \
  --scheduler

--arch	Description
mobilenet	MobileNetV2, head only
swin_t	Swin Transformer Tiny, head only
swin_t_finetune	Swin Transformer Tiny, last block + head unfrozen
swin_s	Swin Transformer Small, last block + head unfrozen

3. Evaluate

python scripts/evaluate.py \
  --arch swin_t_finetune \
  --weights models/spoilage_model_swin_finetuned.pth

Prints a full classification report and confusion matrix on the test set.