SENN.py 文件源码

def inference(self, images, is_train):
        '''Net configuration as the original paper'''
        image_input = tf.reshape(images, [-1, self.N_IN, self.NEFF, 1])
        # ipdb.set_trace()
        with tf.variable_scope('con1') as scope:
            h_conv1 = self._conv_layer_wrapper(image_input, 12, 13, is_train)
        with tf.variable_scope('con2') as scope:
            h_conv2 = self._conv_layer_wrapper(h_conv1, 16, 11, is_train)
        with tf.variable_scope('con3') as scope:
            h_conv3 = self._conv_layer_wrapper(h_conv2, 20, 9, is_train)
        with tf.variable_scope('con4') as scope:
            h_conv4 = self._conv_layer_wrapper(h_conv3, 24, 7, is_train)
        with tf.variable_scope('con5') as scope:
            h_conv5 = self._conv_layer_wrapper(h_conv4, 32, 7, is_train)
        with tf.variable_scope('con6') as scope:
            h_conv6 = self._conv_layer_wrapper(h_conv5, 24, 7, is_train)
        with tf.variable_scope('con7') as scope:
            h_conv7 = self._conv_layer_wrapper(h_conv6, 20, 9, is_train)
        with tf.variable_scope('con8') as scope:
            h_conv8 = self._conv_layer_wrapper(h_conv7, 16, 11, is_train)
        with tf.variable_scope('con9') as scope:
            h_conv9 = self._conv_layer_wrapper(h_conv8, 12, 13, is_train)
        with tf.variable_scope('con10') as scope:
            f_w = h_conv9.get_shape()[1].value
            i_fm = h_conv9.get_shape()[-1].value
            W_con10 = weight_variable(
                [f_w, 129, i_fm, 1])
            b_conv10 = bias_variable([1])
            h_conv10 = conv2d(h_conv9, W_con10) + b_conv10
        return tf.reshape(h_conv10, [-1, self.NEFF])