srilankan-food-trainer 0.2.2

Easy-to-use library for extending Sri Lankan food recognition models

🍛 Sri Lankan Food Trainer

Easy-to-use Python library for extending the Sri Lankan Food Recognition model with your own vegetable classes — no deep learning expertise required.

Built on a custom Prototypical Network trained on 8 Sri Lankan vegetable–cooking state combinations. Add a new class with as few as 20–30 images, without retraining the full model from scratch.

---

Why This Library?

Traditional Sri Lankan cooking — red curry, white curry, tempering — transforms vegetables so dramatically that standard food recognition models fail to identify them. This library lets you extend a model specifically built for that challenge, and adapt it to new vegetables with minimal data.

Approach	Images needed	Time
Train from scratch	500–1000+ per class	Days
This library	20–50 per class	30–60 minutes

---

Installation

pip install srilankan-food-trainer

Requirements: Python 3.8+, PyTorch 2.0+, internet connection on first run (downloads pre-trained model from Hugging Face automatically)

GPU is recommended but not required — CPU training works, just slower.

---

Quick Start

from srilankan_food_trainer import FoodModelExtender

# 1. Create extender — downloads pre-trained model automatically
extender = FoodModelExtender(verbose=True)

# 2. Add your class
#    ZIP filename becomes the class name:
#    e.g. potato_tempered.zip → class "potato_tempered"
extender.add_class("potato_tempered", "potato_tempered.zip", auto_extract=True)

# 3. Train (preserves original 8 classes + adds your new one)
extender.train(epochs=50)

# 4. Save
extender.save("extended_model.pth")

The saved model recognises all 9 classes — original 8 + your new one.

---

Step-by-Step Guide

Step 1 — Prepare your images

Collect 30–50 images of your food class. Organise them in a folder, then zip it:

potato_tempered/          ← folder name = class name
├── image_001.jpg
├── image_002.jpg
├── image_003.jpg
└── ... (30–50 images)

zip -r potato_tempered.zip potato_tempered/

Important: The ZIP filename becomes the class name. Use the format vegetable_cookingstate (e.g. beetroot_white_curry.zip).

---

Step 2 — Initialise the extender

from srilankan_food_trainer import FoodModelExtender

extender = FoodModelExtender(verbose=True)
# Output:
# 🚀 Sri Lankan Food Model Extender initialized!
#    Device: cuda
#    Base model: ranasinghehashini/srilankan-food-recognition

---

Step 3 — Add your class

extender.add_class(
    class_name="potato_tempered",
    images_path="potato_tempered.zip",
    auto_extract=True
)
# Output:
# 📸 Adding new class: potato_tempered
#    Found 37 images
#    Split complete:
#       Train: 25 images
#       Val:   5 images
#       Test:  7 images

The method automatically validates images, checks formats, and splits into train / validation / test sets (70% / 15% / 15%).

---

Step 4 — Train

results = extender.train(epochs=50, save_checkpoints=True)

print(f"Best validation accuracy: {results['best_val_acc']:.2f}%")
print(f"Training time: {results['total_time']/60:.1f} minutes")

Expected training time: 30–45 min with GPU, 60–90 min on CPU.

---

Step 5 — Save and use

extender.save("extended_model.pth")

Load it later with:

import torch

checkpoint = torch.load("extended_model.pth")
# checkpoint contains:
# - model_state_dict   → neural network weights
# - original_classes   → the 8 original class names
# - new_classes        → your added class names
# - training_history   → loss and accuracy per epoch
# - base_model         → source model reference
# - timestamp          → when trained

---

Image Requirements

Requirement	Details
Minimum images	20 per class
Recommended	30–50 per class
Format	JPG or PNG
Resolution	Minimum 224 × 224
Packaging	All images in one folder, zipped
ZIP naming	ZIP filename = class name

Tips for good results:

• Use clear, well-lit photos
• Include different angles (top view, side view, close-up)
• Keep the subject consistent — all images should be the same food type and cooking state
• Remove blurry or dark images before zipping

---

Naming Convention

Follow the same pattern as the original 8 classes — vegetable_cookingstate:

potato_tempered        ✅
beetroot_white_curry   ✅
lentil_red_curry       ✅
"Potato Curry"         ❌  spaces not allowed
potatocurry            ❌  missing underscore separator

---

Pre-trained Model — Original 8 Classes

The model this library extends was trained on:

#	Class	Vegetable	Cooking state
1	carrot_raw	Carrot	Raw
2	carrot_white_curry	Carrot	White curry (coconut milk)
3	greenbeans_raw	Green beans	Raw
4	greenbeans_tempered	Green beans	Tempered (high-heat stir-fry)
5	greenbeans_white_curry	Green beans	White curry (coconut milk)
6	pumpkin_raw	Pumpkin	Raw
7	pumpkin_red_curry	Pumpkin	Red curry (turmeric-based)
8	pumpkin_white_curry	Pumpkin	White curry (coconut milk)

Base model accuracy: 90.25% validation · 84.91% full test · 87.75% few-shot test

---

API Reference

FoodModelExtender(model_path=None, verbose=False)

Parameter	Type	Default	Description
model_path	str or None	None	Path to local .pth checkpoint. If None, downloads pre-trained model automatically.
verbose	bool	False	Print detailed progress messages.

---

.add_class(class_name, images_path, auto_extract=True)

Registers a new food class for training.

Parameter	Type	Default	Description
class_name	str	—	Name for the new class. Should match ZIP filename.
images_path	str	—	Path to the ZIP file containing images.
auto_extract	bool	True	Automatically extract the ZIP file.

Returns True on success, False on failure.

---

.train(epochs=50, save_checkpoints=True)

Fine-tunes the model to include the new class.

Parameter	Type	Default	Description
epochs	int	50	Training rounds. Try 75–100 for low accuracy.
save_checkpoints	bool	True	Auto-save the best model during training.

Returns a dict with best_val_acc, total_time, and training_history.

---

.save(output_path)

Saves the extended model as a PyTorch .pth checkpoint.

Parameter	Type	Description
output_path	str	Filename for the saved model (e.g. extended_model.pth).

---

Understanding Your Results

Validation accuracy	Meaning
90%+	Excellent — very reliable
85–90%	Very good — suitable for most applications
80–85%	Good — consider adding more images
Below 80%	Needs improvement — see troubleshooting

---

Troubleshooting

"Not enough images" error You need at least 20 images. Add more images and re-zip.

"Model download failed" error Check your internet connection and re-run the cell. If it persists, restart your runtime.

"Out of memory" error Restart your runtime and try again. Keep image count under 100 per class.

Low accuracy (below 80%)

• Add more high-quality images (aim for 50)
• Increase epochs to 75 or 100
• Check that all images are the same food type and cooking state
• Remove blurry or poorly lit images

Wrong ZIP structure

✅ Correct:                     ❌ Wrong:
potato_tempered.zip             potato_tempered.zip
└── potato_tempered/            ├── img1.jpg  ← images directly in ZIP
    ├── img1.jpg                └── img2.jpg
    └── img2.jpg

---

Google Colab Tutorial

A complete step-by-step tutorial is available — no local setup needed:

📓 Open in Google Colab

The notebook covers:

1. Installing the library
2. Initialising the model extender
3. Uploading your images (ZIP upload via Colab)
4. Adding your class and validating images
5. Training the extended model
6. Saving and downloading the result

Expected total time: ~1 hour (including 30–60 min training).

---

Related Resources

Resource	Link
Pre-trained model	Hugging Face — ranasinghehashini/srilankan-food-recognition
Training dataset	Kaggle — Sri Lankan Food Recognition Dataset
Research	A Transformation-Aware Deep Learning Approach for Recognizing Cooked Vegetable Ingredients in Sri Lankan Cuisine, Ranasinghe G. H. C., 2026

---

Citation

@misc{ranasinghe2026foodshot,
  author    = {Ranasinghe, G. H. C.},
  title     = {A Transformation-Aware Deep Learning Approach for Recognizing
               Cooked Vegetable Ingredients in Sri Lankan Cuisine},
  year      = {2026},
  note      = {BSc Hons in Computing, Coventry University / NIBM},
}

---

License

MIT License — free to use, modify, and distribute with attribution.