model.py 文件源码

def __reslayer_bottleneck(self, inputs, in_filters, out_filters, stride=1):
        """ A regular resnet block """
        with tf.variable_scope('sub1'):
            kernel = tf.get_variable('weights', [1, 1, in_filters, out_filters / 4],
                                     initializer=xavier_initializer(
                                         dtype=tf.float32),
                                     dtype=tf.float32)
            conv = tf.nn.conv2d(inputs, kernel, [1, stride, stride, 1],
                                padding='SAME',
                                name='conv')
            batch_norm = self.__batch_norm_wrapper(conv, decay=0.9999, shape=[0, 1, 2])
            conv = tf.nn.elu(batch_norm, 'elu')

        with tf.variable_scope('sub2'):
            kernel = tf.get_variable('weights',
                                     [3, 3, out_filters / 4, out_filters / 4],
                                     initializer=xavier_initializer(
                                         dtype=tf.float32),
                                     dtype=tf.float32)
            conv = tf.nn.conv2d(conv, kernel, [1, 1, 1, 1], padding='SAME',
                                name='conv1')
            batch_norm = self.__batch_norm_wrapper(conv, decay=0.9999, shape=[0, 1, 2])
            conv = tf.nn.elu(batch_norm, 'elu')

        with tf.variable_scope('sub3'):
            kernel = tf.get_variable('weights', [1, 1, out_filters / 4, out_filters],
                                     initializer=xavier_initializer(
                                         dtype=tf.float32),
                                     dtype=tf.float32)
            conv = tf.nn.conv2d(conv, kernel, [1, 1, 1, 1],
                                padding='SAME',
                                name='conv')
            batch_norm = self.__batch_norm_wrapper(conv, decay=0.9999, shape=[0, 1, 2])

        with tf.variable_scope('subadd'):
            if in_filters != out_filters:
                kernel = tf.get_variable('weights', [1, 1, in_filters, out_filters],
                                         initializer=xavier_initializer(
                                         dtype=tf.float32),
                                         dtype=tf.float32)
                inputs = tf.nn.conv2d(
                    inputs, kernel, [1, stride, stride, 1], padding='SAME')
            batch_norm += inputs
            conv = tf.nn.elu(batch_norm, 'elu')

            num = np.power(2, np.floor(np.log2(out_filters) / 2))

            grid = self.__put_activations_on_grid(conv, (int(num),
                                                         int(out_filters /
                                                             num)))
            tf.summary.image('sub3/activations', grid, max_outputs=1)

        return conv