Polyaxon with an introductory example

import polyaxon as plx

def graph_fn(mode, inputs):
    inference = plx.layers.FullyConnected(mode=mode, n_units=64, activation='tanh')(inputs)
    return plx.layers.FullyConnected(mode=mode, n_units=10)(inference)

results1 = graph_fn(plx.Modes.TRAIN, dataset1)
results2 = graph_fn(plx.Modes.EVAL, dataset2)

Same thing can be achieved using Subgraph

import polyaxon as plx

config = plx.configs.SubGraphConfig(
    methods=[plx.layers.FullyConnected, plx.layers.FullyConnected],
    kwargs=[{'n_units': 64, 'activation': 'tanh'}, {'n_units': 10}]
modules = plx.libs.SubGraph.build_subgraph_modules(plx.Modes.TRAIN, config)
graph = plx.libs.SubGraph(mode=plx.Modes.TRAIN, name='graph', modules=modules)

results1 = graph(dataset1)
results2 = graph(dataset2)

The difference between the first approach and second is that the second creates a scope for the subgraph and only builds and connects the layers.

Important concepts


Polyaxon make use of tensorflow tf.make_template to easily share variables, and all of Polyaxon module inherits from GraphModule. Each Polyaxon module is python object that represent a part of the computation graph.

Input data

Reading data from a file, set of files, a directory or numpy/pandas objects should be easy and reproducible.

# an example of NUMPY data input configuration
train_data = 'train_input_data_config': {
    'input_type': plx.configs.InputDataConfig.NUMPY,
    'pipeline_config': {'name': 'train', 'batch_size': 64, 'num_epochs': 5,
                        'shuffle': True},
    'x': X_train,
    'y': Y_train


Visualizing the graph is fully customizable and can be defined by providing the level and types of visualization:

The visualization levels: activations, loss, gradients, variables, and learning_rate.

The visualization types currently supported: scalar, histogram, and image.