🍔 Food Vision 📷

An end-to-end CNN Image Classification Model which identifies the food in your image.

I exercised using a pre-trained Keras image classification model after that retrained it using the Food101 Dataset.

Objective:

Make a better model than the DeepFood Paper's model which also trained on the same dataset.

Dataset : Food101

Model: EfficientNetB0

Training time : 47 min.

Accuracy: 80.40 %

How Food-vision build ?

• Downloading Food101 dataset from Tensorflow Dataset Module.
• Knowing dataset: Visualise-Visualise-Visualise
• Setup Mixed Precision In order to train model faster setup global dtype policy to mixed_float16(Implementing Mixed Precision Training)
• Build feature extraction Model
• Fit feature extraction Model
• Load and evaluate checkpoint weights
• Save model and use later on
• Preparing above model for fine-tuning
• Model Callbacks(Minimising resources unnecessary use)
  • Tensorboard Callback: TensorBoard provides the visualization and tooling needed for machine learning experimentation.
  • EarlyStoppingCallback: Used to stop training when a validation loss has stopped reducing.
  • ReduceLROnPlateau: Reduce learning rate when a model is not finding better prediction than previous epochs.
• Building and Training of a Fine Tuning Model: In this, we use pretrained models weights from above model and tweaked it get better
  results. Architecture : EffficientNetB0
• Evaluating results using latest result.