TfExperiment

A simple library to manage Tensorflow experiments though git and reduce boilerplate. Compatible with tf 1.x

Usage

This library relies of git to manage experiments. Each experiment should be a unique git branch, and the name of the experiment, if not give, will be the current git branch.

experiment = tfExperiment.Experiment(finalizeGraph = False)
experiment.saveGraph()

# output
# > graph location ======================================>
# > tensorboard --logdir  output\experimentName\graph

# with with
with experiment.trainingSession(epochs = 125, saveAfter = 2, testAfter = 2) as ts:
    ts.saveGraph() # function to save the graph
    ts.trainCallback = runTrainingCallback
    ts.testCallback = runTestCallback

# as function
experiment.train(runTrainingCallback)
experiment.test(runTestCallback)

API

__init__(name = None, finalizeGraph = False, location = os.path.join(os.getcwd(), 'output'))

• name: string: Name of the experiment, if no name is provided the name of the current git branch will be used.

• finalizeGraph: bool: Finalizes the graph. Attention I have not tried this feature much.

• location: string: absolute path where the experiment results where saved in a folder with same name as name

train(trainCallback, epochs = 1, saveModelAfter = 2, saveGraph = False, testCallback = None, testAfter = 0)

Runs the training and validates/test the model

• trainCallback: function: Function to be run at each epoch. This should contain your loop with the training actions to execute for each batch. The training callback can take 2 parameters: session (current tf.session), and env (if env is used you should use the exact name) experiment environment with access to functionalities like timer and dataSaver.

• epochs: integer: Number of epochs to run, that is to say the number of times the traininCallbacks will be called. Attention: the experiment object keeps track of the number of epochs run so far, so if you call experiment.train again, the epoch number will continue to grow from the last epoch number.

• saveModelAfter: integer: Save the model after n epochs have run. This only considers the current run.

• saveGraph: bool: If we should save the graph at the current run.

• testCallback: function: Function to call to test/validate the current network. Similar to trainCallback.

• testAfter: integer: test the model after n epochs have run. This only considers the current run.

test(testCallback)

Runs the testing/validation of the model once

• testCallback: function: Function to call to test/validate the current network. Similar to trainCallback.

env: Box object

The env object contains

• env.training.currentEpoch: integer: number of epochs since the instance was initialized.
• env.training.currentEpoch: integer: number of epochs since the instance was initialized.
• env.training.dataSavePath: path string: path in which the data will be used if dataSaver is used during training.
• env.training.dataSaver: dataSaver Instance: dataSaver instance (initialized with env.training.dataSavePath) for training to the training file.
• env.testing.dataSavePath: path string path in which the data will be used if dataSaver is used during testing.
• env.testing.dataSaver: dataSaver Instance: dataSaver instance (initialized with env.testing.dataSavePath) for testing to the training file.

Proposed New API

def TrainExperiment(Experiment):
    def __init__(self, constructor, ...):
        #someconfig
        #self.nrTotEpochs
        #self.epochsToValidateAfter
        #...

    def beforeEpoch
    def afterEpoch

    def beforeSave
    def afterSave

    def beforeTest
    def afterTest

    def beforeIteration
    def afterIteration

    def train(session, data, dataProvider = None):
        return 0 #trainingLoopPerSession

    def validate(session, data, dataProvider = None):
        return 0 #trainingLoopPerSession


experiment(TrainExperiment)