Different behavior for max pooling with padding "VALID"

[Corea]

I translate code below by tensorflow2ncnn.

shape = (1, 4, 7, 1)  # NHWC

images = tf.placeholder(tf.float32, shape=shape, name="input/image")
layer = slim.max_pool2d(images, kernel_size=2, stride=2, padding="VALID")
output = tf.identity(layer, name="output/score")

Output shape is different between ncnn and tensorflow. I think output layer should have 6 elements. (shape is (1, 2, 3, 1)) But ncnn has 8 elements with shape (4, 2, 1).