[feat] Adding UPerNet

[brianhou0208]

Description

I tried to add UperNet before. This may not be the best version. Maybe you can have a look.
@GrantorShadow @jawi289o

Issue

#906

Reference

• official repo
• mmseg repo

[qubvel]

Hi @brianhou0208, thanks a lot for working on this! Looks good to me, a few general notes/questions

• Can you please use more verbose variable naming, e.g. ch->channels, h -> height (it is not hold over the repo, but would be nice to have it)
• Please, add the model to the docs/ and README.md
• Did you have a chance to train it with any notebook example?

[brianhou0208]

Hi @qubvel , I’ve updated the code according to the commit, along with the README and documentation.
I’ve also trained the model using the binary_segmentation_intro.ipynb, and the results are shown below.

Due to device limitations, the default parameters caused an OOM error, so I adjusted the decoder_pyramid_channels and decoder_segmentation_channels parameters.

• Valid Dataset

Arch	Backbone	pyramid_channels	segmentation_channels	IoU (Per Image / Dataset)
UPerNet	ResNet34	32	32	90.51 / 91.35
UPerNet	ResNet34	32	16	90.48 / 91.38
UPerNet	ResNet34	32	8	90.01 / 90.92
FPN	ResNet34	-	-	90.56 / 91.39
DeepLabV3Plus	ResNet34	-	-	90.26 / 91.07
UNet	ResNet34	-	-	90.50 / 91.31

• Test Dataset

Arch	Backbone	pyramid_channels	segmentation_channels	IoU (Per Image / Dataset)
UPerNet	ResNet34	32	32	91.11 / 91.75
UPerNet	ResNet34	32	16	91.10 / 91.76
UPerNet	ResNet34	32	8	91.01 / 91.73
FPN	ResNet34	-	-	90.97 / 91.64
DeepLabV3Plus	ResNet34	-	-	90.48 / 91.22
UNet	ResNet34	-	-	91.15 / 91.76

[qubvel]

Thanks for update, I made a quick look at the architecture design, here is main concern

[qubvel]

Looks great, just one nit and we are ready to merge this!

[brianhou0208]

I've finished it and fixed a few small issues.

[qubvel]

@brianhou0208 Thanks a lot for addressing it!