[Feature] Add DDRNet (PaddlePaddle#1984)

Author: justld

configs/ddrnet/README.md

@@ -0,0 +1,13 @@
+# Deep dual-resolution networks for real-time and accurate semantic segmentation of road scenes
+
+## Reference
+
+> Yuanduo Hong, Huihui Pan, Weichao Sun, Yisong Jia. "Deep dual-resolution networks for real-time and accurate semantic segmentation of road scenes." arXiv preprint arXiv:2101.06085 (2021).
+
+## Performance
+
+### Cityscapes
+
+| Model | Backbone | Resolution | Training Iters | mIoU | mIoU (flip) | mIoU (ms+flip) | Links |
+|-|-|-|-|-|-|-|-|
+|DDRNet_23|-|1024x1024|120000|79.85%|80.11%|80.44%|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/ddrnet23_cityscapes_1024x1024_120k/model.pdparams)\|[log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/ddrnet23_cityscapes_1024x1024_120k/train.log)\|[vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=33c0d5f37e5a708c605e43ef3845ea56)|

configs/ddrnet/ddrnet23_cityscapes_1024x1024_120k.yml

@@ -0,0 +1,24 @@
+_base_: '../_base_/cityscapes_1024x1024.yml'
+
+batch_size: 3
+iters: 120000
+
+model:
+  type: DDRNet_23
+  enable_auxiliary_loss: False
+  pretrained: https://bj.bcebos.com/paddleseg/dygraph/cityscapes/ddrnet23_cityscapes_1024x1024_120k/pretrain/model.pdparams
+
+optimizer:
+  type: sgd
+  weight_decay: 0.0005
+
+loss:
+  types:
+    - type: OhemCrossEntropyLoss
+  coef: [1]
+
+lr_scheduler:
+  type: PolynomialDecay
+  learning_rate: 0.01
+  end_lr: 0.0
+  power: 0.9

paddleseg/models/__init__.py

@@ -56,4 +56,5 @@
 from .bisenetv1 import BiseNetV1
 from .fastfcn import FastFCN
 from .pfpnnet import PFPNNet
-from .glore import GloRe
+from .glore import GloRe
+from .ddrnet import DDRNet_23