Keras中Conv2D padding='same'的填充值设置及维度计算咨询

Hey there! Let's break down exactly how Keras calculates padding when you set padding='same'—it's totally clear once you walk through the rules and your specific example.

First, the core idea of padding='same': Keras automatically calculates how many pixels to add around your input feature map so that the output's spatial dimensions match your input's spatial dimensions when strides=(1,1). If your strides are larger than 1, it adjusts padding to make the output dimensions equal to ceil(input_size / strides) instead.

Step-by-step calculation rules #

For each spatial dimension (height and width separately):

1. Calculate output size:
   output_size = ceil(input_size / strides)
2. Calculate total padding needed:
   total_pad = max((output_size - 1) * strides + kernel_size - input_size, 0)
3. Split padding evenly (or as evenly as possible):
   • Padding on the top/left side: pad_before = total_pad // 2
   • Padding on the bottom/right side: pad_after = total_pad - pad_before
     Note: If total padding is odd, the extra pixel goes to the bottom/right side.

Applying this to your code example #

Your input X has shape (3, 2, 2, 2)—let's focus on the spatial dimensions: height=2, width=2. You're using kernel_size=(2,2) and strides=(1,1).

1. Output size calculation:
   output_height = ceil(2 / 1) = 2, same for width.
2. Total padding needed:
   For height: (2-1)*1 + 2 - 2 = 1 + 2 - 2 = 1
   Same for width: total padding = 1.
3. Split padding:
   • Height: pad_top = 1//2 = 0, pad_bottom = 1-0 =1
   • Width: pad_left=0, pad_right=1

So Keras adds 0 pixels to the top/left, 1 pixel to the bottom/right of your input (with 0-value padding, the default for Conv2D). After padding, your input becomes 3x3 in spatial dimensions, and applying the 2x2 kernel with stride 1 gives an output of 2x2—matching your original input's spatial size.

After running your code, the output X shape will be (3, 2, 2, 4)—the 4 comes from your filters=4 parameter, and the spatial dimensions stay 2x2 thanks to padding='same'.

Quick note for larger strides #

If you ever use strides=(2,2) with, say, an input height of 5 and kernel size 3:

• output_height = ceil(5/2) =3
• total_pad = (3-1)*2 +3 -5 =4+3-5=2
• pad_top=1, pad_bottom=1 (even padding here)

内容的提问来源于stack exchange，提问作者edn