You can deploy this model and start earning money today!

cicerothoma/
nigerian_food_classification

Name: nigerian_food_classification
Availability: InStock

image-classification

cicerothoma/nigeria_food

google/efficientnet-b4

finetune

mit

Views

No views yet

Nigerian Food Classification — EfficientNet-B4

Classifies images of Nigerian food into 18 classes using transfer learning with EfficientNet-B4. This model is fine-tuned on a curated dataset and optimized for balanced precision and recall across diverse dishes and plating styles.

Model Card

Architecture: EfficientNet-B4 (feature extractor) + custom classifier head (18 classes)
Framework: PyTorch
Loss: Cross-Entropy
Optimizer: Adam
Mixed precision: Autocast (FP16) during training/eval where supported
Input size: typically 320–380 px (center-crop + random resized crop recommended)
Classes (18): Jollof Rice, Egusi Soup, Moi Moi, Akara, Suya, Efo Riro, Okra Soup, Ofada Rice, Pounded Yam, Banga Soup, Pepper Soup, Nkwobi, Amala, Ewedu Soup, Ogbono Soup, Yam Porridge, Puff Puff, Chin Chin

Performance

Validation results (EfficientNet-B4):

Accuracy: ~72.2%
Macro F1: ~0.715
Weighted F1: ~0.724

Class highlights:

Strong/balanced: Moi Moi, Akara, Jollof Rice, Puff Puff, Suya, Yam Porridge
Improved recall: Efo Riro (0.727), Ewedu Soup (0.643), Pounded Yam (0.950)
Still challenging: Okra Soup (recall 0.412), Ogbono Soup (0.476), Banga Soup (0.385)

Interpretation:

The larger capacity of B4 improves texture/structure cues (leafy soups, starches), lifting recall and macro F1 versus smaller backbones.
Viscous and visually similar soups (okra/ogbono/banga) remain prone to confusion without calibration or class-focused augmentation.

Limitations

General:

Sensitive to low-quality images (poor lighting, occlusions, unusual angles).
Performance can vary with domain shift (different sources, plating, camera settings).

EfficientNet-B4–specific:

Model size and memory: Larger than B0/B2; higher VRAM/RAM footprint. On macOS with MPS, long runs may hit allocator high-watermark OOM limits.
Latency/throughput: Deeper network increases inference time; consider FP16 and small batch sizes for server deployment.
Input resolution dependence: Performs best at higher resolutions; aggressive downscaling can reduce accuracy on texture-heavy dishes.
Class-specific recall: Tends to under-predict visually similar viscous soups without class-wise thresholds, focal loss, or targeted augmentation.
Deployment trade-offs: Quantization (FP16/INT8) and pruning reduce size/latency but may slightly impact accuracy; knowledge distillation to EfficientNet-B0/B2 can retain performance with a smaller footprint.

Dataset

18 Nigerian food classes sourced from publicly available images on the internet.
Train/validation split with standard image augmentations (recommendations below).
Note: Dataset bias may exist due to source variability; continued curation is encouraged.

Recommended Training Setup

Preprocessing:
- Resize to 320–380 px with RandomResizedCrop and CenterCrop for eval
- Normalize with ImageNet means/stds
Augmentations:
- Horizontal flips, color jitter, random brightness/contrast
- Class-focused augmentation for soups: variance in lighting, bowl/container types, and close-up crops to emphasize texture/viscosity
Regularization:
- Label smoothing (0.05–0.1)
- Mixup/CutMix for generalization
Loss and scheduling:
- Cross-entropy; optionally class-weighted or focal loss (gamma 1.5–2.0) to lift low-recall classes
- Warmup with frozen early layers, then unfreeze with lower LR on backbone (e.g., 1/5 of head LR)
Calibration:
- Temperature scaling or class-wise thresholds on a held-out set to rebalance recall vs precision

Usage

Inference (PyTorch)

python

import torch
from torchvision import transforms
from PIL import Image

# Load model
model = torch.load("best.pt", map_location="cpu")
model.eval()

# Class names
CLASSES = [
    "jollof_rice", "egusi_soup", "moi_moi", "akara", "suya", "efo_riro",
    "okra_soup", "ofada_rice", "pounded_yam", "banga_soup", "pepper_soup",
    "nkwobi", "amala", "ewedu_soup", "ogbono_soup", "yam_porridge",
    "puff_puff", "chin_chin"
]

# Preprocessing (ImageNet-style)
preprocess = transforms.Compose([
    transforms.Resize(380),
    transforms.CenterCrop(380),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225]),
])

def predict(image_path, topk=5, device="cpu"):
    img = Image.open(image_path).convert("RGB")
    x = preprocess(img).unsqueeze(0)
    x = x.to(device)
    with torch.no_grad():
        logits = model(x)
        probs = logits.softmax(dim=1).squeeze(0)
    topk_probs, topk_idx = probs.topk(topk)
    return [(CLASSES[i], float(topk_probs[j])) for j, i in enumerate(topk_idx)]

print(predict("examples/jollof.jpg"))

Autocast (MPS/FP16)

python

from torch import autocast

device = torch.device("mps" if torch.backends.mps.is_available() else "cpu")
model.to(device)

def predict_fp16(image_path, topk=5):
    img = Image.open(image_path).convert("RGB")
    x = preprocess(img).unsqueeze(0).to(device)
    with torch.no_grad(), autocast(device_type="mps", dtype=torch.float16):
        logits = model(x)
        probs = logits.softmax(dim=1).squeeze(0)
    topk_probs, topk_idx = probs.topk(topk)
    return [(CLASSES[i], float(topk_probs[j])) for j, i in enumerate(topk_idx)]

Deployment Notes

Quantization:
- FP16 inference recommended for speed/memory
- INT8 quantization via PTQ on linear layers can reduce footprint; validate class-specific recall
Distillation:
- Train a smaller student (EfficientNet-B0/B2 or MobileNetV3-Large) using B4 logits to preserve recall with lower latency
Export:
- TorchScript or ONNX for server deployment; CoreML with MLProgram for Apple devices
Server tips:
- Batch=1, stream requests; preprocess on CPU
- Monitor memory; on macOS MPS you can adjust PYTORCH_MPS_HIGH_WATERMARK_RATIO (use cautiously)

Example Space

Try the demo: upload an image and see the top-5 predictions with probabilities. Recommended images are in examples/.

Citation

If you use this model, please cite this repository/model card and acknowledge the dataset sources.

Acknowledgements

Developed to promote Nigerian cuisine and culture through technology. Thanks to the open-source community for tools and resources that made this model possible.

cicerothoma/nigerian_food_classification