RLearn
RLearn aims to allow Keras and scikit-learn models to be trained remotely on a server with as little configuration required as possible. RLearn takes advantage of the server's existing Keras configuration and H204GPU to accelerate training on the remote server even when the local machine doesnt have the hardware to do so.
NOTE: This package is meant for individual developers who either have an existing desktop with a GPU, or a GPU compute instance on some cloud and wishes to use it to accelerate their machine learning workflow. It could be used in a professional setting but there are no garuntees as to its correctness, security, or robustness so this is not recommended as it could fail at ANY time.
Why not just remotely host a Jupyter notebook for training?
Most of the data pre and postprocessing doesn't need to be done on a GPU accelerated machine. Keeping that computation remotely allows us to free up resources for training or running other applications on the machine.
Usage
Normal Usage :
session = RLearnSession('localhost:8765')
#session.train(model, x, y, compileargs, fitargs)
trained = session.train(model, x_train, y_train, {
'loss': 'categorical_crossentropy',
'optimizer': 'Adadelta',
'metrics': ['accuracy']
}, {
'batch_size': 128,
'epochs': 10
})Advanced Usage :
The underlying methods used by RLearnSession.train() can be used to make management and reuse of datasets, models, and jobs more efficient for your application.
# This is an example from how one can break down the train() method for
# Keras' MNIST CNN model.
session = RLearnSession('localhost:8765')
#session.addData(x, y, name)
session.addData(x_train, y_train, 'mnistdata')
#session.addModel(model, name)
session.addModel(model, 'mnistmodel')
#session.addJob(jobtype, modelname, dataname, compileargs, fitargs)
trained = session.addJob('keras', 'mnistmodel', 'mnistdata',
{
'loss': 'categorical_crossentropy',
'optimizer': 'Adadelta',
'metrics': ['accuracy']
},
{
'batch_size': 128,
'epochs': 10
})