rishikeshmane / seefood2 Goto Github PK

This is a Streamlit based application that provides basic information about Indian Food (origin, ingredients & preparation time). This is based on a fictional app based on the HBO series "Silicon Valley". This can be found here -> https://adhok-seefood-streamlit-app-7lw6nk.streamlitapp.com/

This uses the Indian Food Image Dataset that consists of 85 categories that can be found on Kaggle(https://www.kaggle.com/datasets/iamsouravbanerjee/indian-food-images-dataset). Apart from the images that are already present, we have also augmented the dataset using different transformations (rotation, channel shift, vertical & horizontal flip , shearing). You can recreate this by running the code python augment.py on your terminal. 90% of the data was used for training, while 10% was used for validation.

1. The pixels of each image are normalized to be in the range 0-1, by dividing the images by 255.

2. These normalized images are split into batches of 128 and fed into a pre-trained EfficientNetV2 Network, where we train the model for 100 epochs.

   • The first 25 epochs were trained with only the last layer unfrozen
   • The second 25 epochs were trained with the last two layers unfrozen
   • The third 25 epochs were trained with the last three layers unfrozen
   • The last 25 epochs were trained with the last four layers unfrozen

   The model weights were re-used from the previous steps (E.g Model weights for the second 25 epoch run were obtained from the first 25 epoch run)

3. We use Adam optimizer with a learning rate of 0.003

4. The model weights are then automatically stored in the file training_1/cp.ckpt.

5. The training can be done by running the command python train.py

1. streamlit_app.py -> This contains the front end function of the app, such as layout , text color and the drag and drop functionality. The app can be invoked using the command streamlit run streamlit_app.py.

2. data_prep_for_prediction.py -> This contains the code that maps all 80 types of Indian Food to its prep time, region of origin and ingredients.

3. inference.py -> This uses the model weights file created in the folder training_1/ and outputs the desired results(Food Type, Region of Origin, Prep time and Ingredients) based on the input image.

The train F1 score was ~90%