tensorflow convert fully connected layer to convolutional layer

vgg16_fc_convolution.py

def vgg16_fc_convolution(in_put, out_channel, layer_name, use_relu=True):
    with tf.variable_scope(layer_name):
        input_shape = in_put.get_shape()
        assert len(input_shape) == 4
        height, width, in_channel = input_shape[1:]
        print(height, width, in_channel)
        weights = tf.get_variable(name="weights", shape=[height*width*in_channel, out_channel])
        biases = tf.get_variable(name="biases", shape=[out_channel])
        reshape_weights = tf.reshape(weights,
                                     shape=[tf.to_int32(height), tf.to_int32(width),
                                            tf.to_int32(in_channel), out_channel])
        convolution_output = tf.nn.conv2d(input=in_put, filter=reshape_weights, strides=[1, 1, 1, 1],
                                          padding="VALID")
        output = tf.nn.bias_add(convolution_output, biases)
        if use_relu:
            output = tf.nn.relu(output)
        return output