如何在 tensorflow 2.0 中提供 padding=6？

Question

self.conv6 = tf.keras.layers.Conv2D(1024, kernel_size=3, padding=6,dilation_rate=6) # atrous convolution

this gives error in tensorflow 2.0
I see the padding='same' option
but what if I want to give padding=6

File "/home/jake/venv/lib/python3.7/site-packages/tensorflow_core/python/keras/utils/conv_utils.py", line 202, in normalize_padding padding = value.lower() AttributeError: 'int' object has no attribute 'lower'

Answer 1

According to the tf.keras.layers.Conv2D documentation, padding can only be either 'same' or 'valid' .

'same' ensures that the spatial dimensions are preserved and 'valid' adds the minimum amount of padding required to ensure that the filter "fits" your spatial dimensions (which can be an issue if stride is anything other than one).

You shouldn't need to specify an custom padding size as in most cases it would be unnecessary upsampling. But if you must, you can have a padding layer before the conv layer.

tf.keras.layers.ZeroPadding2D(padding=(6, 6)) # put this before your conv layer