|
--- |
|
license: mit |
|
datasets: |
|
- ethz/food101 |
|
language: |
|
- en |
|
metrics: |
|
- accuracy |
|
base_model: |
|
- google/efficientnet-b4 |
|
pipeline_tag: image-classification |
|
library_name: keras |
|
tags: |
|
- computer-vision |
|
- classification |
|
- deep-learning |
|
- efficientnet |
|
--- |
|
|
|
# EfficientNetB4 Fine-Tuned on Food101 |
|
|
|
This repository contains a fine-tuned EfficientNetB4 model trained on the [Food101 dataset](https://huggingface.co/datasets/mhamza-007/multi-class-food-dataset). The Food101 dataset comprises 101 different classes of food, making it an excellent benchmark for image classification tasks in the food domain. |
|
|
|
## Model Details |
|
|
|
- **Base Architecture**: EfficientNetB4 (pre-trained on ImageNet) |
|
- **Fine-Tuning Layers**: Last 10 layers unfrozen |
|
- **Number of Classes**: 101 (Food101) |
|
- **Input Shape**: (224, 224, 3) |
|
|
|
## Training Configuration |
|
|
|
- **Epochs**: 10 |
|
- **Batch Size**: 32 |
|
- **Optimizer**: Adam |
|
- **Learning Rate**: 0.0001 |
|
- **Loss Function**: `sparse_categorical_crossentropy` |
|
- **Metrics**: `accuracy` |
|
- **Validation Split**: 0.15 |
|
- **Fine-Tuning**: Unfreezing last 10 layers of the base model |
|
|
|
## Performance |
|
|
|
| Phase | Loss | Accuracy | |
|
|--------------|---------|----------| |
|
| **Train** | 0.4790 | 87.40% | |
|
| **Test** | 0.9583 | 74.28% | |
