MMSegmention系列之三（基本的网络架构和预训练模型）

1、常见设置

•我们默认使用4 gpu的分布式训练。
ImageNet上所有的pytorch style预训练骨干都是我们自己训练的，步骤与paper相同。我们的ResNet风格骨干是基于ResNetV1c变体，其中输入干中的7x7卷积被三个3x3卷积取代。
•对于不同硬件的一致性，我们报告GPU内存为torch.cuda.max_memory_allocated()对于所有4个带有torch.cudnn.benchmark=False的GPU的最大值。注意，这个值通常小于nvidia-smi显示的值
我们报告推理时间为网络转发和后处理的总时间，不包括数据加载时间。结果是通过脚本工具/benchmark.py获得的，该脚本工具使用torch.backends.cudnn计算200幅图像的平均时间 torch.backends.cudnn. benchmark=False.
该框架中有两种推理模式
slide模式:
test_cfg将类似于dict (mode=‘slide’， crop_size=(769,769)， stride=(513,513))。在这种模式下，将输入图像裁剪出多个patch，分别传入网络。 crop size大小和步长由作物大小和步长指定。重叠区域按平均合并整个模式:
test_cfg将类似于dict (mode= ‘whole’)。在这种模式下，整个图像将直接传递到网络中。默认情况下，我们对769x769训练模型使用滑动推理，对其余模型使用整体推理。
•对于输入大小为8x+1(例如769)的情况，传统做法采用align_corner=True。否则，对于8x(例如512,1024)的输入大小，align_corner=False被采用。

2、Baselines（其他模型是通过这些基线模型演变过来的）

目前我们支持下列 EncoderDecoder 类型的方法：

1、FCN

2、PSPNet

3、DeepLabV3

4、PSANet

Please refer to PSANet for details.

5、DeepLabV3+

Please refer to DeepLabV3+ for details.

6、UPerNet

Please refer to UPerNet for details.

7、NonLocal Net

Please refer to NonLocal Net for details.

8、EncNet

Please refer to EncNet for details.

9、 CCNet

Please refer to CCNet for details.

10、DANet

Please refer to DANet for details.

11、APCNet

Please refer to APCNet for details.

12、HRNet

Please refer to HRNet for details.

13、GCNet

Please refer to GCNet for details.

14、DMNet

Please refer to DMNet for details.

15、ANN

Please refer to ANN for details.

16、OCRNet

Please refer to OCRNet for details.

17、Fast-SCNN

Please refer to Fast-SCNN for details.

18 ResNeSt

Please refer to ResNeSt for details.

19、Semantic FPN

Please refer to Semantic FPN for details.

20、 PointRend

Please refer to PointRend for details.

MobileNetV2

Please refer to MobileNetV2 for details.

MobileNetV3

Please refer to MobileNetV3 for details.

EMANet

Please refer to EMANet for details.

DNLNet

Please refer to DNLNet for details.

CGNet

Please refer to CGNet for details.

Mixed Precision (FP16) Training
Please refer Mixed Precision (FP16) Training on BiSeNetV2 for details.

U-Net

Please refer to U-Net for details.

ViT

Please refer to ViT for details.

Swin

Please refer to Swin for details.

SETR

Please refer to SETR for details.

Speed benchmark

3、统计模型

[ALGORITHM] ANN (16 ckpts)

[ALGORITHM] APCNet (12 ckpts)

[BACKBONE] BEiT (2 ckpts)

[ALGORITHM] BiSeNetV1 (11 ckpts)

[ALGORITHM] BiSeNetV2 (4 ckpts)

[ALGORITHM] CCNet (16 ckpts)

[ALGORITHM] CGNet (2 ckpts)

[BACKBONE] ConvNeXt (6 ckpts)

[ALGORITHM] DANet (16 ckpts)

[ALGORITHM] DeepLabV3 (41 ckpts)

[ALGORITHM] DeepLabV3+ (42 ckpts)

[ALGORITHM] DMNet (12 ckpts)

[ALGORITHM] DNLNet (12 ckpts)

[ALGORITHM] DPT (1 ckpts)

[ALGORITHM] EMANet (4 ckpts)

[ALGORITHM] EncNet (12 ckpts)

[ALGORITHM] ERFNet (1 ckpts)

[ALGORITHM] FastFCN (12 ckpts)

[ALGORITHM] Fast-SCNN (1 ckpts)

[ALGORITHM] FCN (41 ckpts)

[ALGORITHM] GCNet (16 ckpts)

[BACKBONE] HRNet (37 ckpts)

[ALGORITHM] ICNet (12 ckpts)

[ALGORITHM] ISANet (16 ckpts)

[ALGORITHM] K-Net (7 ckpts)

[BACKBONE] MAE (1 ckpts)

[BACKBONE] MobileNetV2 (8 ckpts)

[BACKBONE] MobileNetV3 (4 ckpts)

[ALGORITHM] NonLocal Net (16 ckpts)

[ALGORITHM] OCRNet (24 ckpts)

[ALGORITHM] PointRend (4 ckpts)

[ALGORITHM] PSANet (16 ckpts)

[ALGORITHM] PSPNet (54 ckpts)

[BACKBONE] ResNeSt (8 ckpts)

[ALGORITHM] SegFormer (13 ckpts)

[ALGORITHM] Segmenter (5 ckpts)

[ALGORITHM] Semantic FPN (4 ckpts)

[ALGORITHM] SETR (7 ckpts)

[ALGORITHM] STDC (4 ckpts)

[BACKBONE] Swin Transformer (6 ckpts)

[BACKBONE] Twins (12 ckpts)

[ALGORITHM] UNet (25 ckpts)

[ALGORITHM] UPerNet (16 ckpts)

[BACKBONE] Vision Transformer (11 ckpts)