simple.py 文件源码

def conv2d(x,
           num_outputs,
           kernel_size,
           strides,
           padding='SAME',
           activation=None,
           bn=False,
           post_bn=False,
           phase=None,
           scope=None,
           reuse=None):
    # Convert int to list
    kernel_size = [kernel_size] * 2 if isinstance(kernel_size, int) else kernel_size
    strides = [strides] * 2 if isinstance(strides, int) else strides

    # Convert list to valid list
    kernel_size = list(kernel_size) + [x.get_shape().dims[-1], num_outputs]
    strides = [1] + list(strides) + [1]

    # Conv operation
    with tf.variable_scope(scope, 'conv2d', reuse=reuse):
        kernel = tf.get_variable('weights', kernel_size,
                                 initializer=variance_scaling_initializer())
        biases = tf.get_variable('biases', [num_outputs],
                                 initializer=tf.zeros_initializer)
        output = tf.nn.conv2d(x, kernel, strides, padding, name='conv2d')
        output += biases
        if bn: output = batch_norm(output, phase, scope='bn')
        if activation: output = activation(output)
        if post_bn: output = batch_norm(output, phase, scope='post_bn')

    return output