Conv2D 64-> 128 фильтров имеет больше веса, чем Conv2D 128-> 64 фильтра . Почему ? , .
Это пример кода UNET / CAE без пропусков и конкатенаций.
Model: "model_29"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
_________________________________________________________________
conv2d_756 (Conv2D) (None, 32, 32, 64) 18496
_________________________________________________________________
conv2d_757 (Conv2D) (None, 32, 32, 64) 36928
_________________________________________________________________
max_pooling2d_120 (MaxPoolin (None, 16, 16, 64) 0
_________________________________________________________________
conv2d_758 (Conv2D) (None, 16, 16, 128) 73856
_________________________________________________________________
conv2d_759 (Conv2D) (None, 16, 16, 128) 147584
_________________________________________________________________
max_pooling2d_121 (MaxPoolin (None, 8, 8, 128) 0
_________________________________________________________________
conv2d_760 (Conv2D) (None, 8, 8, 256) 295168
_________________________________________________________________
conv2d_761 (Conv2D) (None, 8, 8, 256) 590080
_________________________________________________________________
up_sampling2d_106 (UpSamplin (None, 16, 16, 256) 0
_________________________________________________________________
conv2d_762 (Conv2D) (None, 16, 16, 128) 295040
_________________________________________________________________
conv2d_763 (Conv2D) (None, 16, 16, 128) 147584
_________________________________________________________________
up_sampling2d_107 (UpSamplin (None, 32, 32, 128) 0
_________________________________________________________________
conv2d_764 (Conv2D) (None, 32, 32, 64) 73792
_________________________________________________________________
conv2d_765 (Conv2D) (None, 32, 32, 64) 36928
_________________________________________________________________
up_sampling2d_108 (UpSamplin (None, 64, 64, 64) 0
_________________________________________________________________
conv2d_766 (Conv2D) (None, 64, 64, 32) 18464
_________________________________________________________________
Некоторые примеры:
32 = 18496 - 18464 (conv2d_756 (Conv2D) vs conv2d_766 (Conv2D))
64 = 73856 - 73792 (conv2d_758 (Conv2D) vs conv2d_764 (Conv2D))
128 = 295168 - 295040 (conv2d_760 (Conv2D) vs conv2d_762 (Conv2D))