train_operation.py

# encoding: utf-
# 8
import tensorflow as tf

import model

import settings
FLAGS = settings.FLAGS

import numpy as np

RMSPROP_DECAY = 0.9                # Decay term for RMSProp.
RMSPROP_MOMENTUM = 0.9             # Momentum in RMSProp.
RMSPROP_EPSILON = 1.0              # Epsilon term for RMSProp.

def train(total_loss, global_step, summaries, batchnorm_updates):
    # Calculate the learning rate schedule.
    num_batches_per_epoch = (FLAGS.num_examples_per_epoch_for_train /
                             FLAGS.batch_size)
    decay_steps = int(num_batches_per_epoch * FLAGS.num_epochs_per_decay)

    # Decay the learning rate exponentially based on the number of steps.
    lr = tf.train.exponential_decay(FLAGS.initial_learning_rate,
                                    global_step,
                                    decay_steps,
                                    FLAGS.learning_rate_decay_factor,
                                    staircase=True)

    # Create an optimizer that performs gradient descent.
    #opt = tf.train.RMSPropOptimizer(lr, RMSPROP_DECAY,
    #                                momentum=RMSPROP_MOMENTUM,
    #                                epsilon=RMSPROP_EPSILON)
    opt = tf.train.AdamOptimizer(lr)
    grads = opt.compute_gradients(total_loss)

    # Add a summary to track the learning rate.
    summaries.append(tf.summary.scalar('learning_rate', lr))

    # Add histograms for gradients.
    for grad, var in grads:
        if grad is not None:
            summaries.append(
                tf.summary.histogram(var.op.name + '/gradients', grad))

    # Apply the gradients to adjust the shared variables.
    apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)

    # Track the moving averages of all trainable variables.
    # Note that we maintain a "double-average" of the BatchNormalization
    # global statistics. This is more complicated then need be but we employ
    # this for backward-compatibility with our previous models.
    variable_averages = tf.train.ExponentialMovingAverage(
        FLAGS.moving_average_decay, global_step)

    # Another possiblility is to use tf.slim.get_variables().
    variables_to_average = (tf.trainable_variables() +
                            tf.moving_average_variables())
    variables_averages_op = variable_averages.apply(variables_to_average)

    # Group all updates to into a single train op.
    batchnorm_updates_op = tf.group(*batchnorm_updates)
    train_op = tf.group(apply_gradient_op, variables_averages_op,
                        batchnorm_updates_op)

    return train_op