Predicting Battery Lifetime with CNNs

Now we were able start a training job from the command line with the option to modify almost everything on the fly. We could adjust things like number of epochs, batch size, shuffling, checkpoint saving and even switch between model architectures easily, by adding a flag after the command. This allowed us to iterate fast, test different theories, and burn through a lot of (free) credits. We' built our model with tf.Keras using the functional API. We feed the array and scalar features into the model at separate entry points, so we can do different things to them before bringing them back together.