Trainer

This module contains the trainer code and other codes regarding training.

class lenet.trainer.trainer(network, dataset)

Trainer for networks

Parameters:

• network – A network class object
• dataset – A tensorflow dataset object

network

This is the network we initialized with. We pass this as an argument and we add it to the current trainer class.

dataset

This is also the initializer. It comes from the lenet.dataset.mnist module.

session

This is a session created with trainer. This session is used for training.

tensorboard

Is a summary writer tool. This writes things into the tensorboard that is then setup on the tensorboard server. At the end of the trainer, it closes this tensorboard.

accuracy(images, labels)

Return accuracy

Parameters:

• images – images
• labels – labels

Returns:

accuracy

Return type:

float

bp_step(mini_batch_size)

Sample a minibatch of data and run one step of BP.

Parameters:mini_batch_size – Integer Returns:total objective and cost of that step Return type:tuple of tensors

summaries(name='tensorboard')

Just creates a summary merge bufer

Parameters:name – a name for the tensorboard directory

test()

Run validation of the model

Returns:accuracy Return type:float

train(iter=10000, mini_batch_size=500, update_after_iter=1000, training_accuracy=False, summarize=True)

Run backprop for iter iterations

Parameters:

• iter – number of iterations to run
• mini_batch_size – Size of the mini batch to process with
• training_accuracy – if True, will calculate accuracy on training data also.
• update_after_iter – This is the iteration for validation
• summarize – Tensorboard operation

training_accuracy(mini_batch_size=500)

Run validation of the model on training set

Parameters:mini_batch_size – Number of samples in a mini batch Returns:accuracy Return type:float

write_summary(iter=0, mini_batch_size=500)

This method updates the summaries

Parameters:

• iter – iteration number to index values with.
• mini_batch_size – Mini batch to evaluate on.