mirrors/mmcv
1
0
Fork 0
mirror of https://github.com/open-mmlab/mmcv.git synced 2025-06-03 21:54:52 +08:00
Code Issues Projects Releases Wiki Activity

[Docs] Add npu build.md and update roi_pool_npu adpater (#2812)

Browse Source
This commit is contained in:
sherie 2023-06-09 10:35:41 +08:00 committed by GitHub
parent c9627e867f
commit 36e6b19b78
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 79 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

48
docs/zh_cn/get_started/build.md
View File

@ -354,3 +354,51 @@ y = torch.tensor([1, 5, 3]).mlu()
w = torch.ones(10).float().mlu()
output = sigmoid_focal_loss(x, y, 2.0, 0.25, w, 'none')
```
### 在昇腾 NPU 机器编译 mmcv
在编译 mmcv 前，需要安装 torch_npu，完整安装教程详见 [PyTorch 安装指南](https://gitee.com/ascend/pytorch/blob/master/docs/zh/PyTorch%E5%AE%89%E8%A3%85%E6%8C%87%E5%8D%97/PyTorch%E5%AE%89%E8%A3%85%E6%8C%87%E5%8D%97.md#pytorch%E5%AE%89%E8%A3%85%E6%8C%87%E5%8D%97)
#### 选项 1: 使用 NPU 设备源码编译安装 mmcv (推荐方式)
- 拉取 [MMCV 源码](https://github.com/open-mmlab/mmcv.git)
```bash
git pull https://github.com/open-mmlab/mmcv.git
```
- 编译
```bash
MMCV_WITH_OPS=1 MAX_JOBS=8 FORCE_NPU=1 python setup.py build_ext
```
- 安装
```bash
MMCV_WITH_OPS=1 FORCE_NPU=1 python setup.py develop
```
#### 选项 2: 使用 pip 安装 Ascend 编译版本的 mmcv
Ascend 编译版本的 mmcv 在 mmcv >= 1.7.0 时已经支持直接 pip 安装
```bash
pip install mmcv -f https://download.openmmlab.com/mmcv/dist/ascend/torch1.8.0/index.html
```
#### 验证
```python
import torch
import torch_npu
from mmcv.ops import softmax_focal_loss
# Init tensor to the NPU
x = torch.randn(3, 10).npu()
y = torch.tensor([1, 5, 3]).npu()
w = torch.ones(10).float().npu()
output = softmax_focal_loss(x, y, 2.0, 0.25, w, 'none')
print(output)
```

44
mmcv/ops/csrc/pytorch/npu/roi_pool_npu.cpp
View File

@ -11,19 +11,37 @@ void roi_pool_forward_npu(Tensor input, Tensor rois, Tensor output,
int64_t pooled_channel = 1;
at::Tensor roi_actual_num = at_npu::native::OpPreparation::ApplyTensor(
{}, rois.options().dtype(at::kInt), rois);
OpCommand cmd;
cmd.Name("RoiPoolingWithArgMax")
.Input(input)
.Input(rois)
.Input(roi_actual_num)
.Output(output)
.Output(argmax)
.Attr("pooled_h", pooled_height_64)
.Attr("pooled_w", pooled_width_64)
.Attr("spatial_scale_h", spatial_scale)
.Attr("spatial_scale_w", spatial_scale)
.Attr("pool_channel", pooled_channel)
.Run();
if (input.sizes()[1] % 16 == 0) {
OpCommand cmd;
cmd.Name("RoiPoolingWithArgMax")
.Input(input)
.Input(rois)
.Input(roi_actual_num)
.Output(output)
.Output(argmax)
.Attr("pooled_h", pooled_height_64)
.Attr("pooled_w", pooled_width_64)
.Attr("spatial_scale_h", spatial_scale)
.Attr("spatial_scale_w", spatial_scale)
.Attr("pool_channel", pooled_channel)
.Run();
} else {
OpCommand cmd;
cmd.Name("RoiPoolingWithArgMax")
.Input(input)
.Input(rois)
.Input(roi_actual_num)
.Output(output)
.Output(argmax)
.Attr("pooled_h", pooled_height_64)
.Attr("pooled_w", pooled_width_64)
.Attr("spatial_scale_h", spatial_scale)
.Attr("spatial_scale_w", spatial_scale)
.Attr("pool_channel", pooled_channel)
.Attr("_exclude_engines", (string) "AiCore")
.Run();
}
}
void roi_pool_backward_npu(Tensor grad_output, Tensor rois, Tensor argmax,