1回答

正十七

2020-03-24

这是我们resnet block的代码：

def residual_block(x, output_channel):
    """residual connection implementation"""
    input_channel = x.get_shape().as_list()[-1]
    if input_channel * 2 == output_channel:
        increase_dim = True
        strides = (2, 2)
    elif input_channel == output_channel:
        increase_dim = False
        strides = (1, 1)
    else:
        raise Exception("input channel can't match output channel")
    conv1 = tf.layers.conv2d(x,
                             output_channel,
                             (3,3),
                             strides = strides,
                             padding = 'same',
                             activation = tf.nn.relu,
                             name = 'conv1')
    conv2 = tf.layers.conv2d(conv1,
                             output_channel,
                             (3, 3),
                             strides = (1, 1),
                             padding = 'same',
                             activation = tf.nn.relu,
                             name = 'conv2')
    if increase_dim:       
        # [None, image_width, image_height, channel] -> [,,,channel*2]
        pooled_x = tf.layers.average_pooling2d(x,
                                               (2, 2),
                                               (2, 2),
                                               padding = 'valid')
        padded_x = tf.pad(pooled_x,
                          [[0,0],
                           [0,0],
                           [0,0],
                           [input_channel // 2, input_channel // 2]])
    else:
        padded_x = x
    output_x = conv2 + padded_x
    return output_x

在做pooling的时候，我们用了valid。

其实不管是same还是valid，因为做pooling总是要变小的。所以变小之后你得保证做了tf.pad后能跟conv2的shape是一样的。保证了这一点后，用valid还是same都可以的。

0

0