My encoder-decoder shows asymmetrical architectures (using torchinfo package)

[bobetocalo]

Using a ResNet-18 backbone (ResNetEncoder), I do not undestand why the decoder achieved using a UNet or FPN architecture is not symetrical ... I attach the summary of a UNetDecoder which only consists of 3M parameters, instead of the 11M related to the encoder. Where is the problem? How can I made the decoder completely symmetrical?

The input of the decoder should be the 8x8x512 feature maps ... but I receive the original image as input 256x256x3. Why?

======================================================================================================================================================
Layer (type:depth-idx)                             Input Shape               Output Shape              Param #                   Kernel Shape
======================================================================================================================================================
LitUNet                                            [8, 3, 256, 256]          [8, 98, 256, 256]         --                        --
├─Unet: 1-1                                        [8, 3, 256, 256]          [8, 98, 256, 256]         --                        --
│    └─ResNetEncoder: 2-1                          [8, 3, 256, 256]          [8, 3, 256, 256]          --                        --
│    │    └─Conv2d: 3-1                            [8, 3, 256, 256]          [8, 64, 128, 128]         9,408                     [7, 7]
│    │    └─BatchNorm2d: 3-2                       [8, 64, 128, 128]         [8, 64, 128, 128]         128                       --
│    │    └─ReLU: 3-3                              [8, 64, 128, 128]         [8, 64, 128, 128]         --                        --
│    │    └─MaxPool2d: 3-4                         [8, 64, 128, 128]         [8, 64, 64, 64]           --                        3
│    │    └─Sequential: 3-5                        [8, 64, 64, 64]           [8, 64, 64, 64]           147,968                   --
│    │    └─Sequential: 3-6                        [8, 64, 64, 64]           [8, 128, 32, 32]          525,568                   --
│    │    └─Sequential: 3-7                        [8, 128, 32, 32]          [8, 256, 16, 16]          2,099,712                 --
│    │    └─Sequential: 3-8                        [8, 256, 16, 16]          [8, 512, 8, 8]            8,393,728                 --
│    └─UnetDecoder: 2-2                            [8, 3, 256, 256]          [8, 16, 256, 256]         --                        --
│    │    └─Identity: 3-9                          [8, 512, 8, 8]            [8, 512, 8, 8]            --                        --
│    │    └─ModuleList: 3-10                       --                        --                        3,151,552                 --
│    └─SegmentationHead: 2-3                       [8, 16, 256, 256]         [8, 98, 256, 256]         --                        --
│    │    └─Conv2d: 3-11                           [8, 16, 256, 256]         [8, 98, 256, 256]         14,210                    [3, 3]
│    │    └─Identity: 3-12                         [8, 98, 256, 256]         [8, 98, 256, 256]         --                        --
│    │    └─Activation: 3-13                       [8, 98, 256, 256]         [8, 98, 256, 256]         --                        --
======================================================================================================================================================

[qubvel]

Hi! It's not an original Unet implementation. Orignal one uses VGG encoder as far as I remember, however, newer encoders provide much better results. You can increase the number of channels in decoder to increase the number of parameters with decoder_channels=(...), see https://smp.readthedocs.io/en/latest/models.html#unet