netx4.py

import tensorflow as tf
import numpy as np
'''
Developed by Tiantong for NTIER CVPR 2017 SR Competition
tong.renly@gmail.com
'''

def model(input_tensor):
    with tf.device("/gpu:0"):
        weights = []
        activations_cl = []
        conv_00_w = tf.get_variable("conv_00_w", [5, 5, 4, 64],
                                    initializer=tf.random_normal_initializer(stddev=np.sqrt(2.0/9)))
        conv_00_b = tf.get_variable("conv_00_b", [64], initializer=tf.constant_initializer(0))
        weights.append(conv_00_w)
        weights.append(conv_00_b)
        activations = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(input_tensor, conv_00_w,
                                                             strides=[1, 1, 1, 1], padding='SAME'), conv_00_b),
                                 "conv_00_a")
        activations_cl.append(activations)
        for i in range(10):
            conv_w = tf.get_variable("conv_%02d_w" % (i+1), [3, 3, 64, 64],
                                     initializer=tf.random_normal_initializer(stddev=np.sqrt(2.0/9/64)))
            conv_b = tf.get_variable("conv_%02d_b" % (i+1), [64], initializer=tf.constant_initializer(0))
            weights.append(conv_w)
            weights.append(conv_b)
            activations = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(activations, conv_w,
                                                                 strides=[1, 1, 1, 1], padding='SAME'), conv_b),
                                     "conv_%02d_a" % (i + 1))
            activations_cl.append(activations)

        conv_w = tf.get_variable("conv_20_w", [3, 3, 64, 4],
                                 initializer=tf.random_normal_initializer(stddev=np.sqrt(2.0/9/64)))
        conv_b = tf.get_variable("conv_20_b", [4], initializer=tf.constant_initializer(0))
        weights.append(conv_w)
        weights.append(conv_b)
        activations = tf.nn.bias_add(value=tf.nn.conv2d(activations, conv_w,
                                                        strides=[1, 1, 1, 1], padding='SAME'), bias=conv_b,
                                     name="conv_20_a")
        activations_cl.append(activations)
        return activations, weights, activations_cl