LightNet++

!!!New Repo.!!! ⇒ EfficientNet.PyTorch: Concise, Modular, Human-friendly PyTorch implementation of EfficientNet with Pre-trained Weights

!!!New Repo.!!! ⇒ MixNet-Pytorch: Concise, Modular, Human-friendly PyTorch implementation of MixNet with Pre-trained Weights

This repository contains the code (PyTorch-1.0+, W.I.P.) for: "LightNet++: Boosted Light-weighted Networks for Real-time Semantic Segmentation" by Huijun Liu.
LightNet++ is an advanced version of LightNet, which purpose to get more concise model design, smaller models, and better performance.

• MobileNetV2Plus: Modified MobileNetV2 (backbone)^[1,8] + DSASPPInPlaceABNBlock^[2,3] + Parallel Bottleneck Channel-Spatial Attention Block (PBCSABlock)^[6] + UnSharp Masking (USM) + Encoder-Decoder Arch.^[3] + InplaceABN^[4].

• ShuffleNetV2Plus: Modified ShuffleNetV2 (backbone)^[1,8] + DSASPPInPlaceABNBlock^[2,3] + Parallel Bottleneck Channel-Spatial Attention Block (PBCSABlock)^[6]+ UnSharp Masking (USM) + Encoder-Decoder Arch.^[3] + InplaceABN^[4].

• MixSeg-MixBiFPN: Modified MixNet (backbone)^[1,8] + MixBiFPNBlock^[2,3] + Encoder-Decoder Arch.^[3]

More about USM(Unsharp Mask)-Operator Block see Repo: SharpPeleeNet

Dependencies

• Python3.6
• PyTorch(1.0.1+)
• inplace_abn
• apex: Tools for easy mixed precision and distributed training in Pytorch
• tensorboard
• tensorboardX
• tqdm

Datasets for Autonomous Driving

• Cityscapes
• Mapillary Vistas Dataset
• Berkeley Deep Drive (BDD100K)
• ApolloScape

Results

Results on Cityscapes (Pixel-level/Semantic Segmentation)

Model	mIoU (S.S* Mixed Precision)	Model Weight
MobileNetV2Plus X1.0	71.5314 (WIP)	cityscapes_mobilenetv2plus_x1.0.pkl (14.3 MB)
ShuffleNetV2Plus X1.0	69.0885-72.5255 (WIP)	cityscapes_shufflenetv2plus_x1.0.pkl (8.59 MB)
MixSeg+MixBiFPN ArchS	72.2321 (WIP)	cityscapes_mixseg_archs_mixbifpn.pkl (16.4 MB)

• S.S.: Single Scale (1024x2048)

Feature Visualization