Polyaxon

Deep Learning and Reinforcement learning library for TensorFlow for building end to end models and experiments.

Design Goals

Polyaxon was built with the following goals:

Modularity: The creation of a computation graph based on modular and understandable modules, with the possibility to reuse and share the module in subsequent usage.
Usability: Training a model should be easy enough, and should enable quick experimentations.
Configurable: Models and experiments could be created using a YAML/Json file, but also in python files.
Extensibility: The modularity and the extensive documentation of the code makes it easy to build and extend the set of provided modules.
Performance: Polyaxon is based on internal tensorflow code base and leverage the builtin distributed learning.
Data Preprocessing: Polyaxon provides many pipelines and data processor to support different data inputs.

Quick start

A simple linear regression

X = np.linspace(-1, 1, 100)
y = 2 * X + np.random.randn(*X.shape) * 0.33

# Test a data set
X_val = np.linspace(1, 1.5, 10)
y_val = 2 * X_val + np.random.randn(*X_val.shape) * 0.33


def graph_fn(mode, inputs):
    return plx.layers.SingleUnit(mode)(inputs['X'])


def model_fn(features, labels, mode):
    model = plx.models.Regressor(
        mode, graph_fn=graph_fn, loss_config=plx.configs.LossConfig(module='mean_squared_error'),
        optimizer_config=plx.configs.OptimizerConfig(module='sgd', learning_rate=0.009),
        eval_metrics_config=[],
        summaries='all', name='regressor')
    return model(features, labels)


estimator = plx.estimators.Estimator(model_fn=model_fn, model_dir="/tmp/polyaxon_logs/linear")

estimator.train(input_fn=numpy_input_fn(
    {'X': X}, y, shuffle=False, num_epochs=10000, batch_size=len(X)))

A reinforcement learning problem

env = plx.envs.GymEnvironment('CartPole-v0')

def graph_fn(mode, inputs):
    return plx.layers.FullyConnected(mode, num_units=512)(inputs['state'])

def model_fn(features, labels, mode):
    model = plx.models.DQNModel(
        mode, 
        graph_fn=graph_fn, 
        loss_config=plx.configs.LossConfig(module='huber_loss'),
        num_states=env.num_states, 
        num_actions=env.num_actions,
        optimizer_config=plx.configs.OptimizerConfig(module='sgd', learning_rate=0.01),
        exploration_config=plx.configs.ExplorationConfig(module='decay'),
        target_update_frequency=10, 
        dueling='mean', 
        summaries='all')
    return model(features, labels)

memory = plx.rl.memories.Memory(
    num_states=env.num_states, num_actions=env.num_actions, continuous=env.is_continuous)
agent = plx.estimators.Agent(
    model_fn=model_fn, memory=memory, model_dir="/tmp/polyaxon_logs/dqn_cartpole")

agent.train(env)

A classification problem

X_train, Y_train, X_test, Y_test = load_mnist()

config = {
    'name': 'lenet_mnsit',
    'output_dir': output_dir,
    'eval_every_n_steps': 10,
    'train_steps_per_iteration': 100,
    'run_config': {'save_checkpoints_steps': 100},
    'train_input_data_config': {
        'input_type': plx.configs.InputDataConfig.NUMPY,
        'pipeline_config': {'name': 'train', 'batch_size': 64, 'num_epochs': None,
                            'shuffle': True},
        'x': X_train,
        'y': Y_train
    },
    'eval_input_data_config': {
        'input_type': plx.configs.InputDataConfig.NUMPY,
        'pipeline_config': {'name': 'eval', 'batch_size': 32, 'num_epochs': None,
                            'shuffle': False},
        'x': X_test,
        'y': Y_test
    },
    'estimator_config': {'output_dir': output_dir},
    'model_config': {
        'summaries': 'all',
        'model_type': 'classifier',
        'loss_config': {'name': 'softmax_cross_entropy'},
        'eval_metrics_config': [{'name': 'streaming_accuracy'},
                                {'name': 'streaming_precision'}],
        'optimizer_config': {'name': 'Adam', 'learning_rate': 0.002,
                             'decay_type': 'exponential_decay', 'decay_rate': 0.2},
        'graph_config': {
            'name': 'lenet',
            'definition': [
                (plx.layers.Conv2d, {'num_filter': 32, 'filter_size': 5, 'strides': 1,
                                     'regularizer': 'l2_regularizer'}),
                (plx.layers.MaxPool2d, {'kernel_size': 2}),
                (plx.layers.Conv2d, {'num_filter': 64, 'filter_size': 5,
                                     'regularizer': 'l2_regularizer'}),
                (plx.layers.MaxPool2d, {'kernel_size': 2}),
                (plx.layers.FullyConnected, {'n_units': 1024, 'activation': 'tanh'}),
                (plx.layers.FullyConnected, {'n_units': 10}),
            ]
        }
    }
}
experiment_config = plx.configs.ExperimentConfig.read_configs(config)
xp = plx.experiments.create_experiment(experiment_config)
xp.continuous_train_and_evaluate()

A regression problem

X, y = generate_data(np.sin, np.linspace(0, 100, 10000, dtype=np.float32), time_steps=7)

config = {
    'name': 'time_series',
    'output_dir': output_dir,
    'eval_every_n_steps': 5,
    'run_config': {'save_checkpoints_steps': 100},
    'train_input_data_config': {
        'input_type': plx.configs.InputDataConfig.NUMPY,
        'pipeline_config': {'name': 'train', 'batch_size': 64, 'num_epochs': None,
                            'shuffle': False},
        'x': X['train'],
        'y': y['train']
    },
    'eval_input_data_config': {
        'input_type': plx.configs.InputDataConfig.NUMPY,
        'pipeline_config': {'name': 'eval', 'batch_size': 32, 'num_epochs': None,
                            'shuffle': False},
        'x': X['val'],
        'y': y['val']
    },
    'estimator_config': {'output_dir': output_dir},
    'model_config': {
        'model_type': 'regressor',
        'loss_config': {'name': 'mean_squared_error'},
        'eval_metrics_config': [{'name': 'streaming_root_mean_squared_error'},
                                {'name': 'streaming_mean_absolute_error'}],
        'optimizer_config': {'name': 'Adagrad', 'learning_rate': 0.1},
        'graph_config': {
            'name': 'regressor',
            'definition': [
                (plx.layers.LSTM, {'num_units': 7, 'num_layers': 1}),
                # (Slice, {'begin': [0, 6], 'size': [-1, 1]}),
                (plx.layers.FullyConnected, {'n_units': 1}),
            ]
        }
    }
}
experiment_config = plx.configs.ExperimentConfig.read_configs(config)
xp = plx.experiments.create_experiment(experiment_config)
xp.continuous_train_and_evaluate()

Creating a distributed experiment

def create_experiment(task_type, task_index=0):

    def graph_fn(mode, inputs):
        x = plx.layers.FullyConnected(mode, num_units=32, activation='tanh')(inputs['X'])
        return plx.layers.FullyConnected(mode, num_units=1, activation='sigmoid')(x)

    def model_fn(features, labels, mode):
        model = plx.models.Regressor(
            mode, graph_fn=graph_fn, loss_config=plx.configs.LossConfig(module='absolute_difference'),
            optimizer_config=plx.configs.OptimizerConfig(module='sgd', learning_rate=0.5, decay_type='exponential_decay', decay_steps=10),
            summaries='all', name='xor')
        return model(features, labels)

    os.environ['task_type'] = task_type
    os.environ['task_index'] = str(task_index)

    cluster_config = {
            'master': ['127.0.0.1:9000'],
            'ps': ['127.0.0.1:9001'],
            'worker': ['127.0.0.1:9002'],
            'environment': 'cloud'
        }

    config = plx.configs.RunConfig(cluster_config=cluster_config)

    estimator = plx.estimators.Estimator(model_fn=model_fn, model_dir="/tmp/polyaxon_logs/xor", config=config)

    return plx.experiments.Experiment(estimator, input_fn, input_fn)

Installation

To install the latest version of Polyaxon: pip install polyaxon

Alternatively, you can also install from source by running (from source folder): python setup.py install

Or you can just clone the repo git clone https://github.com/polyaxon/polyaxon.git, and use the commands to do everything in docker:

cmd/rebuild to build the docker containers.
cmd/py to start a python3 shell with all requirements installed.
cmd/jupyter to start a jupyter notebook server.
cmd/tensorboard to start a tensorboard server.
cmd/test to run the tests.

Examples

Some example are provided here, more examples and use case will pushed, a contribution with an example is also appreciated.

Project status

Polyaxon is in a pre-release "alpha" state. All interfaces, programming interfaces, and data structures may be changed without prior notice. We'll do our best to communicate potentially disruptive changes.

Contributions

Please follow the contribution guide line: Contribute to Polyaxon.

License

MIT License

Credit

This work is based and was inspired from different projects, tensorflow.contrib.learn, keras, sonnet, seq2seq and many other great open source projects, see ACKNOWLEDGEMENTS.

The idea behind creating this library is to provide a tool that allow engineers and researchers to develop and experiment with end to end solutions.

The choice of creating a new library was very important to have a complete control over the apis and future design decisions.