diff --git a/docs/en/device/npu.md b/docs/en/device/npu.md new file mode 100644 index 0000000000..7fc0b899f3 --- /dev/null +++ b/docs/en/device/npu.md @@ -0,0 +1,36 @@ +# NPU (HUAWEI Ascend) + +## Usage + +Please refer to the [building documentation of MMCV](https://mmcv.readthedocs.io/en/latest/get_started/build.html#build-mmcv-full-on-ascend-npu-machine) to install MMCV and [MMEngine](https://mmengine.readthedocs.io/en/latest/get_started/installation.html#build-from-source) on NPU devices. + +Here we use 4 NPUs on your computer to train the model with the following command: + +```shell +bash tools/dist_train.sh configs/deeplabv3/deeplabv3_r50-d8_512x1024_40k_cityscapes.py 4 +``` + +Also, you can use only one NPU to train the model with the following command: + +```shell +python tools/train.py configs/deeplabv3/deeplabv3_r50-d8_512x1024_40k_cityscapes.py +``` + +## Models Results + +| Model | mIoU | Config | Download | +| :-----------------: | :---: | :--------------------------------------------------------------------------------------------------------------------------------------- | :------------------------------------------------------------------------------------------------------------------------- | +| [deeplabv3](<>) | 78.92 | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/deeplabv3/deeplabv3_r50-d8_512x1024_40k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/deeplabv3_r50-d8_512x1024_40k_cityscapes.log.json) | +| [deeplabv3plus](<>) | 79.68 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/deeplabv3plus/deeplabv3plus_r50-d8_512x1024_40k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/deeplabv3plus_r50-d8_512x1024_40k_cityscapes.log.json) | +| [hrnet](<>) | 77.09 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/hrnet/fcn_hr18_512x1024_40k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/fcn_hr18_512x1024_40k_cityscapes.log.json) | +| [fcn](<>) | 72.69 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/fcn/fcn_r50-d8_512x1024_40k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/fcn_r50-d8_512x1024_40k_cityscapes.log.json) | +| [pspnet](<>) | 78.07 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/pspnet_r50-d8_512x1024_80k_cityscapes.log.json) | +| [unet](<>) | 69.00 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/unet/fcn_unet_s5-d16_4x4_512x1024_160k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/fcn_unet_s5-d16_4x4_512x1024_160k_cityscapes.log.json) | +| [apcnet](<>) | 78.07 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/apcnet/apcnet_r50-d8_512x1024_40k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/apcnet_r50-d8_512x1024_40k_cityscapes.log.json) | +| [upernet](<>) | 78.22 | [config](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/upernet/upernet_r50_4xb2-40k_cityscapes-512x1024.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/upernet_r50_512x1024_40k_cityscapes.log.json) | + +**Notes:** + +- If not specially marked, the results on NPU with amp are the basically same as those on the GPU with FP32. + +**All above models are provided by Huawei Ascend group.** diff --git a/docs/en/index.rst b/docs/en/index.rst index ae009bf441..60e6db5d5e 100644 --- a/docs/en/index.rst +++ b/docs/en/index.rst @@ -46,6 +46,11 @@ Welcome to MMSegmentation's documentation! changelog.md faq.md +.. toctree:: + :caption: Device Support + + device/npu.md + .. toctree:: :caption: Switch Language diff --git a/docs/zh_cn/device/npu_zh.md b/docs/zh_cn/device/npu_zh.md new file mode 100644 index 0000000000..506c25b7a6 --- /dev/null +++ b/docs/zh_cn/device/npu_zh.md @@ -0,0 +1,35 @@ +# NPU (华为昇腾) + +## 使用方法 + +首先，请参考[MMCV](https://mmcv.readthedocs.io/zh_CN/latest/get_started/build.html#npu-mmcv-full) 安装带有 NPU 支持的 MMCV与 [MMEngine](https://mmengine.readthedocs.io/en/latest/get_started/installation.html#build-from-source) 。 +使用如下命令，可以利用 4 个 NPU 训练模型（以 deeplabv3为例）： + +```shell +bash tools/dist_train.sh configs/deeplabv3/deeplabv3_r50-d8_512x1024_40k_cityscapes.py 4 +``` + +或者，使用如下命令，在一个 NPU 上训练模型（以 deeplabv3为例）： + +```shell +python tools/train.py configs/deeplabv3/deeplabv3_r50-d8_512x1024_40k_cityscapes.py +``` + +## 经过验证的模型 + +| Model | mIoU | Config | Download | +| :-----------------: | :---: | :--------------------------------------------------------------------------------------------------------------------------------------- | :------------------------------------------------------------------------------------------------------------------------- | +| [deeplabv3](<>) | 78.92 | [config](https://github.com/open-mmlab/mmsegmentation/tree/master/configs/deeplabv3/deeplabv3_r50-d8_512x1024_40k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/deeplabv3_r50-d8_512x1024_40k_cityscapes.log.json) | +| [deeplabv3plus](<>) | 79.68 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/deeplabv3plus/deeplabv3plus_r50-d8_512x1024_40k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/deeplabv3plus_r50-d8_512x1024_40k_cityscapes.log.json) | +| [hrnet](<>) | 77.09 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/hrnet/fcn_hr18_512x1024_40k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/fcn_hr18_512x1024_40k_cityscapes.log.json) | +| [fcn](<>) | 72.69 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/fcn/fcn_r50-d8_512x1024_40k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/fcn_r50-d8_512x1024_40k_cityscapes.log.json) | +| [pspnet](<>) | 78.07 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/pspnet_r50-d8_512x1024_80k_cityscapes.log.json) | +| [unet](<>) | 69.00 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/unet/fcn_unet_s5-d16_4x4_512x1024_160k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/fcn_unet_s5-d16_4x4_512x1024_160k_cityscapes.log.json) | +| [apcnet](<>) | 78.07 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/apcnet/apcnet_r50-d8_512x1024_40k_cityscapes.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/apcnet_r50-d8_512x1024_40k_cityscapes.log.json) | +| [upernet](<>) | 78.22 | [config](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/upernet/upernet_r50_4xb2-40k_cityscapes-512x1024.py) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/device/npu/upernet_r50_512x1024_40k_cityscapes.log.json) | + +**注意:** + +- 如果没有特别标记，NPU 上的结果与使用 FP32 的 GPU 上的结果结果相同。 + +**以上所有模型权重及训练日志均由华为昇腾团队提供** diff --git a/docs/zh_cn/index.rst b/docs/zh_cn/index.rst index 4ac9211b62..c02846ef58 100644 --- a/docs/zh_cn/index.rst +++ b/docs/zh_cn/index.rst @@ -46,6 +46,11 @@ changelog.md faq.md +.. toctree:: + :caption: 设备支持 + + device/npu_zh.md + .. toctree:: :caption: 语言切换 diff --git a/mmseg/utils/util_distribution.py b/mmseg/utils/util_distribution.py index e0f082ed46..a99939ed91 100644 --- a/mmseg/utils/util_distribution.py +++ b/mmseg/utils/util_distribution.py @@ -37,6 +37,7 @@ def build_dp(model, device='cuda', dim=0, *args, **kwargs): assert digit_version(mmcv.__version__) >= digit_version('1.7.0'), \ 'Please use MMCV >= 1.7.0 for NPU training!' from mmcv.device.npu import NPUDataParallel + torch.npu.set_device(kwargs['device_ids'][0]) torch.npu.set_compile_mode(jit_compile=False) dp_factory['npu'] = NPUDataParallel model = model.npu()