• We have seen a popular meme that tries to represent the similarity shared between animals and food such as "Shiba Inu dog or toasted marshmallow?" So, We would like to develop the deep learning model that removes the uncertainty of an image that could be like a loaf of bread or corgi butt. But for sure, We just do it for fun.

• We used the PyTorch framework with GPU to develop our model using Google Colaboratory.

• Bread dataset was acquired from the combination of Google Open Images V5 + Google Images search

• Corgi dataset was acquired from the combination of ImageNet + Stanford Dogs + Google Images search

  After that, we manually remove incorrect images and apply phash (Perceptual Hashing) to get rid of duplication images.

  • Total images: 6385 images (randomly split using 80:10:10 ratios)

    • Bread: 3710 images
      • Train: 2968 images
      • Valid: 371 images
      • Test: 371 images
    • Corgi: 2675 images
      • Train: 2140 images
      • Valid: 268 images
      • Test: 267 images

• ResNet-152

• We already know that in order to benchmark our model performance, we can't just use accuracy and validation_loss value as the only acceptable metrics.

• To reproduce the model, requires our datasets. You can send me an e-mail at [email protected] if you are interested.

• Initial datasets/ directory structure

     └───datasets/
     │     butt/
     │     bread/

• Install dataset-split library

  pip install dataset-split

• Execute dataset-split command with following arguments

  dataset-split dataset/ -r 0.8 0.1 0.1

• Ready-to-go datasets/ directory structure

     └───datasets/
     │   │
     │   └───train
     │   │    │   butt/
     │   │    │   bread/
     │   └───test
     │   │    │   butt/
     │   │    │   bread/
     │   └───valid
     │   │    │   butt/
     │   │    │   bread/

• Clone this repository

  git clone https://github.com/Kawaeee/butt_or_bread.git

• Install dependencies

  pip install -r requirements.txt

• Run the train.py python script

  python train.py --dataset-path datasets/ --model-path buttbread_resnet152_3.h5

• Check jupyter notebook for interactive prediction: predictor.ipynb

• Clone this repository

  git clone https://github.com/Kawaeee/butt_or_bread.git

• Install dependencies

  pip install -r requirements.txt

• Run the streamlit

  streamlit run streamlit_app.py

Streamlit web application will be hosted on https://localhost:8501

• Following instructions below

  # Directly build and run
  docker build -t butt-bread-image .
  docker run --rm --name=butt-bread-container -p 0.0.0.0:8501:8501 butt-bread-image

  # Serve with docker compose
  docker-compose build
  docker-compose up