This is not technically a hot dog... but this network was trained on "sausage, hot dog, frankfurter" and is not too rigorous a model.

Steps towards reproducing results:

Run the download_multiple function inside of download_google_images.py. Example usage:

  download_multiple({"search_words": ["plane", "dog", "hot dog"],
                 "dir": "pictures",
                 "max_count": 1000,
                 "chrome_driver": "/usr/local/custom_bin/chromedriver",
                 "headless": True})

This will create the following directory:

  pictures/
      - dog/
      - plane/
      - /hot\ dog/

We are trying to train a binary classifier, so you need to condense the previous folders into 2 classes: my 2 classes are "hot dog" and "not hot dog" I am using tensorflow ImageDataGenerator() to create datasets ingestible by a CNN, to get data in this form, we need our folder structure to look like:

  data/
      - train/
          - hot\ dog/
          - not\ hot\ dog/
      - val/
          - hot\ dog/
          - not\ hot\ dog/
      - test/
          - hot\ dog/
          - not\ hot\ dog/

Inside of data_prep.py, create_batches will achieve this. Now, from main.py, we will reference this data folder to actually train our CNN. Running this file will create .h5 models at every epoch. You can test this model using pred.py.

Now we must use convert.py to convert it to an .mlmodel, which is the format iOS uses.

ViewController.swift has the necessary swift code for model inference using the native camera. Make sure to add the camera privacy description in your info.plist