yusasif/Nigerian-food-recognision

Nigerian Food Lens — Multi-Label Food Detector

Fine-tuned EfficientNetV2-S model for recognising Nigerian foods in photos. Given a plate image it returns every food item visible along with a confidence score.

Model Details

• Architecture: EfficientNetV2-S (torchvision), fine-tuned classifier head
• Training: Multi-label classification with weighted BCE loss
• Input: RGB image, resized to 384×384
• Output: Sigmoid probabilities per class — all items above threshold are returned
• Threshold: 0.3 (recommended)

How to Use

Python (direct)

import torch
import torch.nn as nn
from torchvision import models, transforms
from torchvision.models import EfficientNet_V2_S_Weights
from huggingface_hub import hf_hub_download
from PIL import Image
import json

# Download model files
model_path = hf_hub_download("yusasif/Nigerian-food-Lens", "best_model.pth")
vocab_path = hf_hub_download("yusasif/Nigerian-food-Lens", "label_vocab.json")

# Load vocab
with open(vocab_path) as f:
    vocab = json.load(f)
labels = vocab["labels"]

# Load model
checkpoint = torch.load(model_path, map_location="cpu", weights_only=False)
m = models.efficientnet_v2_s(weights=None)
m.classifier = nn.Sequential(
    nn.Dropout(p=0.3, inplace=True),
    nn.Linear(m.classifier[1].in_features, len(labels)),
)
m.load_state_dict(checkpoint["model_state_dict"])
m.eval()

# Preprocess image
transform = transforms.Compose([
    transforms.Resize((384, 384)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
image = Image.open("your_food_image.jpg").convert("RGB")
tensor = transform(image).unsqueeze(0)

# Predict
with torch.no_grad():
    probs = torch.sigmoid(m(tensor)).squeeze(0).numpy()

# Top 3 results
top3 = sorted(enumerate(probs), key=lambda x: x[1], reverse=True)[:3]
for idx, score in top3:
    print(f"{labels[idx]}: {score*100:.1f}%")

REST API (Docker)

# Run the API server
docker compose up --build

# Send an image
curl -X POST http://localhost:8003/detect \
  -F "file=@plate.jpg"

Response:

{
  "detected": [
    {"food": "Egusi soup", "confidence": 0.87},
    {"food": "Eba",        "confidence": 0.81},
    {"food": "Stockfish",  "confidence": 0.74}
  ]
}

Detectable Foods (41 classes)

• Akara
• Akara And Bread
• Amala
• Banga soup
• Beans
• Beef
• Bitterleaf soup
• Bread
• Chicken
• Cow foot
• Cow skin
• Eba
• Edikaikong
• Efo Riro
• Egg
• Egusi soup
• Ewedu
• Fish
• Fufu
• Garri and Groundnut
• Gbegiri
• Iyan
• Jollof rice
• Moi Moi
• Nkwobi
• Ofe Oha
• Ofe Owerri
• Ogbono soup
• Ogi
• Ogi and Akara
• Okra soup
• Onion
• Plantain
• Rice
• Shrimp
• Snail
• Spaghetti
• Stew
• Stockfish
• Tomato
• Tripe

Training Data

Trained on the NutriMama Nigerian food dataset — a multi-label annotated collection of Yoruba, Igbo, and Hausa dishes covering swallows, soups, proteins, rice dishes, snacks and street food.