dimgag/deepfood

Deep Food Image Recognition Project

🍕 DeepFood 🍝

🥖 Multiclass Classification using Tensorflow on Food-101 Dataset 🥐

🍟 Download & Extract Food-101 Dataset 🍔

!wget http://data.vision.ee.ethz.ch/cvl/food-101.tar.gz
!tar xzvf food-101.tar.gz

To split the dataset into Train and Test run the python/split_train_test.py in your terminal after selecting the desired directory.

🌯 Install requirements 🌮

pip install -r requirements.txt

🍪 Files Overview 🥛

food-101 > images - Format of the Food-101 dataset and how to be splitted into Train and Test
         > meta
         > test
         > train 

models   > EfficientNetV2L          > assets
                                    > variables 
                                    > EfficiencyNetV2L.hdf5
                                    > EfficiencyNetV2L.log
                                    > kears_metadata.pb
                                    > saved_model.pb

         > EfficientNetV2S          > assets
                                    > variables 
                                    > EfficiencyNetV2S.hdf5
                                    > EfficiencyNetV2S.log
                                    > kears_metadata.pb
                                    > saved_model.pb

         > EfficientNetV2S_25Epochs > EfficiencyNetV2S.hdf5
                                    > EfficiencyNetV2S.log

python   > evaluate.py - Evaluate the model on the test set
         > main.py - Main script to run the model
         > models.py - Models definition + Fine Tuning
         > split_train_test.py - Create the data folders in DSRI persistent folder
         > train.py - Train the model
         > visualization.py - Visualize the model output
readme_images               - Images used in the README.md
test_images                 - images used for testing the model

vision_transformer     > vit_cifar100.py - Vision Transformer model definition
                       > vit.py - Vision Transformer model definition

DeepFood_Food101.ipynb          - Code Notebook with Models and Data
EfficientNetV2_Evaluation.ipynb - Code Notebook with Models and Data Evaluation
logs_analysis.ipynb             - Code Notebook with Logs Analysis
model_predictions.ipynb         - Code Notebook with Model Predictions

README.md                       - README
requirements.txt                - Requirements for the repository

⚙️ Training Configuration ⚙️

Number of Classes	101
Number of training samples	75750
Number of validation samples	25250
Input image dimensions	(299, 299)
Batch size	32
Number of Epochs	100
Learning Rate	0.0001
Momentum	0.9
Optimizer	Stochastic gradient descent
Loss Function	Categorical crossentropy
Evaluation Metric	Accuracy
GPU	NVIDIA Tesla V100 SXM2 32 GB

🚀 Models Results

Model	Training Accuracy	Validation Accuracy	Training Loss	Validation Loss	# Parameters	# Layers
EfficientNetV2-S	0.9129	0.8337	0.4350	0.7551	20.5M	516
EfficientNetV2-L	0.9411	0.8463	0.3394	0.7650	117.9M	1031

📊 Visualization EfficientNetV2-S vs EfficientNetV2-L 📈

🍽 Predictions

🚀 Test the model

# Activate your python env with the requirement.txt
# Then Run 
python food_app.py

Future code to do

• ViT for food recognition task
• Test models on images with added noise (filters, rotations, etc).

🍺 Acknowledgements 🍻

• Food-101 Dataset
• EfficientNetV2