common.py

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
import numpy as np

HParams = {'relu_leakiness': 0.1, 'num_classes': 1000, 'learning_rate': 1e-5, 'beta': 1e5, 'codec_dim': 96}

def add_relu(x, leakiness=0.0, scope=None):
  with tf.variable_scope(scope or 'relu'):
    return tf.where(tf.less(x, 0.0), leakiness * x, x, name='leaky_relu')

def add_sigmoid(x, scope=None):
  with tf.variable_scope(scope or 'sigmoid'):
    return tf.nn.sigmoid(x, name='nn_sigmoid')

def add_tanh(x, scope=None):
  with tf.variable_scope(scope or 'tangent'):
    return tf.nn.tanh(x, name='nn_tangent')

def add_conv2d(x, kernel_size, in_channels, out_channels, strides, scope=None):
  with tf.variable_scope(scope or 'conv2d'):
    n = kernel_size * kernel_size * in_channels
    filter = tf.get_variable(
      'filter', [kernel_size, kernel_size, in_channels, out_channels],
      initializer=tf.random_normal_initializer(stddev=np.sqrt(2.0 / n)),
      dtype=tf.float32)
    conv = tf.nn.conv2d(x, filter, strides, padding='SAME')
    biases = tf.get_variable('biases', [out_channels], dtype=tf.float32, initializer=tf.constant_initializer(0.0))
    pre_activation = tf.nn.bias_add(conv, biases)
    return pre_activation