Add MobileNetV3 architecture for Segmentation

[datumbox]

Adding MobileNetV3 models for Semantic Segmentation (resolution 520):

Lite R-ASPP with Dilated MobileNetV3 Large Backbone

Heavily optimized for speed. Good for actual mobile usage.

Weight checkpoint:

PR3276/3rd_training/35354083/model_28.pth

Validate:

python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py --dataset coco\
   --model lraspp_mobilenet_v3_large --test-only --pretrained

Accuracy metrics:

0: global correct: 91.2
0: average row correct: ['94.5', '84.3', '69.5', '72.8', '57.7', '42.0', '77.0', '57.0', '90.4', '36.1', '76.0', '60.8', '81.4', '78.9', '81.0', '87.6', '51.3', '83.9', '62.2', '84.2', '56.1']
0: IoU: ['90.2', '69.2', '57.7', '58.5', '47.8', '35.7', '69.5', '47.1', '79.1', '29.6', '62.6', '34.2', '65.5', '63.4', '70.0', '76.8', '30.1', '61.9', '46.8', '70.6', '49.1']
0: mean IoU: 57.9

Speed Benchmark: 0.3278 sec per image on CPU

DeepLabV3 with Dilated MobileNetV3 Large Backbone

Offers good balance between speed and accuracy, significantly faster than the FCN model with a resnet50 backbone without sacrificing too much accuracy.

Weight checkpoint:

PR3276/3rd_training/35354080/model_28.pth

Validate:

python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py --dataset coco\
   --model deeplabv3_mobilenet_v3_large --test-only --pretrained

Accuracy metrics:

0: global correct: 91.2
0: average row correct: ['93.7', '84.9', '73.6', '74.6', '63.6', '50.6', '80.7', '65.1', '91.3', '42.2', '80.4', '70.6', '82.4', '81.8', '83.7', '88.5', '52.6', '87.8', '65.9', '88.3', '63.3']
0: IoU: ['90.1', '69.7', '58.2', '61.3', '49.7', '37.6', '72.7', '52.7', '79.1', '32.2', '64.6', '36.2', '66.7', '67.4', '70.1', '77.3', '33.1', '67.8', '51.1', '73.3', '54.4']
0: mean IoU: 60.3

Speed Benchmark: 0.5869 sec per image on CPU

[fmassa]

Looks great, thanks!

I only have a couple of minor (non-blocking) comments. The only thing I would really like to see fixed before merge is to have the correct python -m torch.distributed.launch ... commands for reproducibility.

[fmassa]

out of curiosity, what does cn mean in here

[datumbox]

It's from the C0,C1...C5,Cn names used in Object Detection. I use this feature internally to find out where the downsampling was supposed to happen but it's not always done with strides so I had to rename it. If you have any better name for it, happy to change it. I could not think of any...

[fmassa]

thanks for the explanation. Given that this is private I'm fine with this name

[fmassa]

Is this feature used in any of the models? Otherwise we can just remove it

[datumbox]

This is a unique implementation detail from the paper on MobileNetV3 models and it's supposed to produce a further speed optimization on object detection and segmentation. In our training scripts we don't use it because we do transfer learning from ImageNet but if someone really wants to train it from scratch and go smaller I provide a way to do it.

On current master this is public (see reduced_tail param) but here I decide to hide before the release and make it an internal implementation detail for future models. Not quite convinced we will use it but want to provide an implementation very close to the paper.

Personally I would prefer to keep it hidden for now and decide later whether we want this gone. Let me know.

[fmassa]

Sounds good, I'm ok keeping this private for now and maybe removing it from the future.