ibrahimdaud/raw-food-recognition-models

Raw Food Recognition Models: Single-Class and Multi-Label

This repository contains both single-class and multi-label classification models trained for raw food ingredient recognition.

Single-Class Classification Models

Model	Parameters	Validation Accuracy	Architecture
ResNet-50	~25.6M	97.84%	Standard residual network
SE-ResNet-50	~26.0M	95.72%	ResNet-50 with SE attention

Both models were trained on the ibrahimdaud/raw-food-recognition dataset, which contains 90+ raw food categories.

Multi-Label Classification Models

Model	Training Mode	Parameters	Best mAP	Architecture
Multi-Label ResNet-50	Freeze Encoder	~24,656,463	0.3747	ResNet-50 encoder (frozen) + classifier

Multi-label models were trained on the ibrahimdaud/multi-label-food-recognition dataset for recognizing multiple ingredients in a single image.

Usage

Download Single-Class Models

from huggingface_hub import hf_hub_download
import torch
from models.resnet50 import create_resnet50
from models.se_resnet50 import create_se_resnet50

# Download ResNet-50 checkpoint
resnet_path = hf_hub_download(
    repo_id="ibrahimdaud/raw-food-recognition-models",
    filename="resnet50_pytorch_model.bin"
)
resnet_checkpoint = torch.load(resnet_path, map_location='cpu')

# Download SE-ResNet-50 checkpoint
se_resnet_path = hf_hub_download(
    repo_id="ibrahimdaud/raw-food-recognition-models",
    filename="se_resnet50_pytorch_model.bin"
)
se_resnet_checkpoint = torch.load(se_resnet_path, map_location='cpu')

Load Single-Class Models

# Create ResNet-50 model
resnet_model = create_resnet50(
    num_classes=90,
    pretrained=False
)
resnet_model.load_state_dict(resnet_checkpoint['model_state_dict'])
resnet_model.eval()

# Create SE-ResNet-50 model
se_resnet_model = create_se_resnet50(
    num_classes=90,
    pretrained=False,
    reduction=16
)
se_resnet_model.load_state_dict(se_resnet_checkpoint['model_state_dict'])
se_resnet_model.eval()

Download Multi-Label Models

from huggingface_hub import hf_hub_download
import torch
from models.multilabel_resnet50 import create_multilabel_resnet50

# Download Freeze Encoder model
freeze_path = hf_hub_download(
    repo_id="ibrahimdaud/raw-food-recognition-models",
    filename="multilabel_freeze_pytorch_model.bin"
)
freeze_checkpoint = torch.load(freeze_path, map_location='cpu')

# Download Full Training model
full_path = hf_hub_download(
    repo_id="ibrahimdaud/raw-food-recognition-models",
    filename="multilabel_full_pytorch_model.bin"
)
full_checkpoint = torch.load(full_path, map_location='cpu')

# Download Fine-Tuning model
finetune_path = hf_hub_download(
    repo_id="ibrahimdaud/raw-food-recognition-models",
    filename="multilabel_finetune_pytorch_model.bin"
)
finetune_checkpoint = torch.load(finetune_path, map_location='cpu')

Load Multi-Label Models

# Load Freeze Encoder model
freeze_model = create_multilabel_resnet50(
    num_classes=freeze_checkpoint['num_classes'],
    pretrained=False
)
freeze_model.load_state_dict(freeze_checkpoint['model_state_dict'])
freeze_model.eval()

# Load Full Training model
full_model = create_multilabel_resnet50(
    num_classes=full_checkpoint['num_classes'],
    pretrained=False
)
full_model.load_state_dict(full_checkpoint['model_state_dict'])
full_model.eval()

# Load Fine-Tuning model
finetune_model = create_multilabel_resnet50(
    num_classes=finetune_checkpoint['num_classes'],
    pretrained=False
)
finetune_model.load_state_dict(finetune_checkpoint['model_state_dict'])
finetune_model.eval()

Multi-Label Inference

import torch
from PIL import Image
import torchvision.transforms as transforms

# Preprocess image
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                        std=[0.229, 0.224, 0.225])
])

image = Image.open('path/to/image.jpg').convert('RGB')
image_tensor = transform(image).unsqueeze(0)

# Get multi-label predictions
with torch.no_grad():
    logits = freeze_model(image_tensor)  # or full_model, finetune_model
    probs = torch.sigmoid(logits)  # Multi-label probabilities
    
    # Get top-k predictions
    top_k = 5
    top_probs, top_indices = torch.topk(probs[0], top_k)
    
    # Assuming you have class names
    for prob, idx in zip(top_probs, top_indices):
        print(f"Class {{idx.item()}}: {{prob.item():.4f}}")

Compare Single-Class Predictions

import torch
from PIL import Image
import torchvision.transforms as transforms

# Preprocess image
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                        std=[0.229, 0.224, 0.225])
])

image = Image.open('path/to/image.jpg').convert('RGB')
image_tensor = transform(image).unsqueeze(0)

# ResNet-50 prediction
with torch.no_grad():
    resnet_outputs = resnet_model(image_tensor)
    resnet_probs = torch.nn.functional.softmax(resnet_outputs[0], dim=0)
    resnet_pred_id = torch.argmax(resnet_probs).item()
    resnet_pred_class = resnet_checkpoint['class_names'][resnet_pred_id]
    resnet_confidence = resnet_probs[resnet_pred_id].item()

# SE-ResNet-50 prediction
with torch.no_grad():
    se_resnet_outputs = se_resnet_model(image_tensor)
    se_resnet_probs = torch.nn.functional.softmax(se_resnet_outputs[0], dim=0)
    se_resnet_pred_id = torch.argmax(se_resnet_probs).item()
    se_resnet_pred_class = se_resnet_checkpoint['class_names'][se_resnet_pred_id]
    se_resnet_confidence = se_resnet_probs[se_resnet_pred_id].item()

# Compare results
print("ResNet-50 Prediction:")
print(f"  Class: {resnet_pred_class}")
print(f"  Confidence: {resnet_confidence*100:.2f}%")

print("\nSE-ResNet-50 Prediction:")
print(f"  Class: {se_resnet_pred_class}")
print(f"  Confidence: {se_resnet_confidence*100:.2f}%")

Model Details

Single-Class Models

ResNet-50

• Architecture: Standard residual network with bottleneck blocks
• Parameters: ~25.6M
• Pretrained: ImageNet weights
• Best Validation Accuracy: 97.84%

SE-ResNet-50

• Architecture: ResNet-50 with Squeeze-and-Excitation attention blocks
• Parameters: ~26.0M
• Pretrained: ImageNet weights (excluding SE blocks)
• SE Reduction Ratio: 16
• Best Validation Accuracy: 95.72%

Multi-Label Models

Freeze Encoder Mode

• Training Strategy: Encoder frozen, only classifier trained
• Use Case: Fast training, preserves encoder features
• Best for: When you have limited data or want quick results

Full Training Mode

• Training Strategy: Both encoder and classifier trained from scratch
• Use Case: Maximum flexibility, learns task-specific features
• Best for: When you have sufficient data and compute

Fine-Tuning Mode

• Training Strategy: Encoder trained with lower learning rate, classifier with higher rate
• Use Case: Balanced approach, preserves some encoder knowledge while adapting
• Best for: General-purpose multi-label classification

Training Details

Single-Class Models

• Dataset: ibrahimdaud/raw-food-recognition
• Number of Classes: 90
• Image Size: 224x224
• Optimizer: Adam
• Learning Rate: 0.001
• Batch Size: 32

Multi-Label Models

• Dataset: ibrahimdaud/multi-label-food-recognition
• Image Size: 224x224
• Optimizer: Adam
• Loss Function: BCEWithLogitsLoss
• Evaluation Metrics: Mean Average Precision (mAP), F1-Score, Hamming Loss

Files in Repository

Single-Class Models

• resnet50_pytorch_model.bin - ResNet-50 model weights
• se_resnet50_pytorch_model.bin - SE-ResNet-50 model weights
• resnet50_metadata.json - ResNet-50 metadata
• se_resnet50_metadata.json - SE-ResNet-50 metadata

Multi-Label Models

• multilabel_freeze_pytorch_model.bin - Multi-label ResNet-50 (Freeze Encoder)

• multilabel_freeze_metadata.json - Freeze Encoder metadata

• README.md - This file

Citation

If you use these models, please cite:

@model{raw_food_recognition_models_2024,
  title={Raw Food Recognition Models: Single-Class and Multi-Label Classification},
  author={Ibrahim Daud},
  year={2024},
  publisher={HuggingFace},
  url={https://huggingface.co/ibrahimdaud/raw-food-recognition-models}
}

License

MIT License