We are now at the testing stages of a convolutional neural network. This means that we are close to releasing the next version which supports a keras convolutional neural network and pickle support for importing keras models.
So what does this mean? Once the new release is uploaded, a user will be able to import a convolutional neural network object from the Keras branch. As long as there are two folders, one with pictures containing the object that they want to detect, and another file filled with pictures not containing the object. With just one line of code, the neural network object will collect all images, label them, and shuffle them. One more line of code will evaluate the model, showing the F-1 score, recall, and accuracy. And just one more line of code will pickle the model with the metrics and the resizing image dimensions. There is also a loading function that will enable the user the open up the saved model, and feed more images into the neural network.
What’s for the future? Sadly the convolutional neural network only supports binary outcomes for now. This does have some uses. For instance, does a chest x-ray have condition A or not? The reason for this is because deploy-ml is about creating robust, easy to use training wrappers for well established machine learning libraries. There is already a strategy lined out supporting multiple classes and instant labelling based on the name of the files the images are in. However, this has to be considered in the bigger picture of how it fits in the training and deployment objects of this module. As of now, there’s no reason why people cannot benefit from a neural network that predicts binary outcomes of images.