CatOrDog is a fun Django application which you can upload an image of a cat or dog. A convolutional neural network is used to predict the type of animal (only cat or dog).
The CNN was developed using keras and trained locally on my laptop (Ubuntu 16.04). Full source code of the CNN can be found in its own repository.
Currently, there is one data model in the app called CatDog. This model has two fields, animal and path. The animal field shows the animal type, cat or dog, whereas the path is the url for the image.
The last 5 images are being rendered in the root page using bootstrap carousel. New predictions can be made by clicking the relevant button. I am using the cloudinary upload widget for uplading images. Using the widget, images are uploaded to cloudinary and the image url is saved to the database to the path field of the newly created CatDog instance. After that, the show page is rendered.
A Celery worker with RabbitMQ broker is being used to asynchronously predict the animal type. The worker file can be found at cat_dog/celery.py. On the initialization of the worker the classifier is loaded from a h5 file. If you are interested in how the h5 file was generated, please click here.
predict method in the task file uses keras.preprocessing.image to convert the uploaded image to an array to be used by the CNN. When the prediction is completed the animal type is saved to animal field.
At this point, the user would most likely be still in the show page. I am making a new request after one second to render the predicted animal type. The predict method should be completed much faster than one second, thus the prediction should be visible.