Dispatcher

Dispatches operations over multiple GPU devices and calculates tower loss

class tflibs.training.dispatcher.Dispatcher(model_cls, model_param, gpus, features, labels=None)

A class for dispatching operations over multiple GPU devices and caculating tower loss

Parameters:

• model_cls (class) – Model class.
• model_param (dict) – A dict of model parameters.
• gpus (list) – A list of gpu ids.
• features (dict) – A dict of tf.Tensor containing feature values.
• labels (tf.Tensor) – A tf.Tensor of labels.

chief

minimize(optimizer, loss_fn, depends=None, global_step=None)

Gets loss_fn which maps model object to loss tensor, caculates tower loss and minimize it.

Parameters:

• optimizer (tflibs.training.Optimizer) – An optimizer.
• loss_fn (function) – A function that maps model object to loss tensor.
• depends (tf.Operation) – An operation that should be run before running an optimization.

Returns:

Train op.

Return type:

tf.Operation

models