Same padding, different results.

[rolandgvc]

Was there a change in how padding is being done? I remember passing 'SAME' as an argument to padding was equivalent to [(1,1),(1,1)] when using strides of 2. Can anyone clarify please?

from flax import linen as nn
import numpy as np
x = np.random.rand(1,100,100, 3)
a = nn.max_pool(x, (3, 3), strides=(2, 2), padding='SAME')
b = nn.max_pool(x, (3, 3), strides=(2, 2), padding=[(1,1),(1,1)])
print(np.mean(np.abs(a - b))) # 0.068035774

[marcvanzee]

SAME pads so the input and output shape are the same. Given that 100 - 33*3 = 1, it only pads 1 after each spatial dimension. Right now your padding mask in b pads one before and one after, so I think it is doing too much padding.

>>> b2 = nn.max_pool(x, (3, 3), strides=(2, 2), padding=[(0, 1),(0, 1)])
>>> print(np.mean(np.abs(a - b2)))
0.0

[avital]

Curiously, this means that in your case, whether padding='SAME' is the same as padding=[(1,1),(1,1)] is dependent on whether the input width and height are even or odd

[rolandgvc]

I see now, thanks.