Merge pull request #2323 from ckirchhoff2021/npu-dev

Npu-dev: FocalLoss Adapter BugFix

.pre-commit-config.yaml

@@ -29,6 +29,7 @@ repos:
     rev: v2.1.0
     hooks:
       - id: codespell
+        exclude: ^README.md
   - repo: https://github.com/executablebooks/mdformat
     rev: 0.7.9
     hooks:

README.md

@@ -247,6 +247,39 @@ c. Install full version with custom operators for onnxruntime
 
 If you would like to build MMCV from source, please refer to the [guide](https://mmcv.readthedocs.io/en/latest/get_started/build.html).
 
+## NPU build and Installation
+
+You may want to run mmcv on your npu device, then you can build and install mmcv-npu by the following steps.
+
+a. Install the **ascend-toolkit**
+
+```python
+ Ascend-cann-toolkit_{version}_linux-{arch}.run
+```
+
+- You can download the ascend-toolkit package in https://www.hiascend.com/software/cann/community. Choose the **"Ascend-cann-toolkit\_{xxx.xxx}.run"** which fits your develop environment.
+- In order to install **CANN** quickly, you can refer to the documents in https://www.hiascend.com/document/detail/zh/canncommercial/51RC2/envdeployment/instg/instg_000052.html
+
+b. Install the **toch_npu**
+
+- As the dispatch mechanism is based on torch, you have to install torch-npu before running your mmcv.ops on npu device.
+- you can download the torch_npu code from https://gitee.com/ascend/pytorch, and install torch-npu as the steps in README.
+- torch-npu depends on ascend-toolkit. So you have to install the ascend-toolkit, and set the ascend environment.
+- ```python
+  source /usr/local/Ascend/ascned-toolkit/set_env.sh
+  ```
+
+c. build and install mmcv-npu
+
+- ```bash
+  MMCV_WITH_OPS=1 FORCE_NPU=1 python setup.py build_ext
+  MMCV_WITH_OPS=1 FORCE_NPU=1 python setup.py develop
+  ```
+- or
+- ```bash
+  MMCV_WITH_OPS=1 FORCE_NPU=1 python setup.py install
+  ```
+
 ## FAQ
 
 If you face some installation issues, CUDA related issues or RuntimeErrors,

mmcv/ops/csrc/pytorch/npu/focal_loss_npu.cpp

@@ -1,10 +1,20 @@
 #include "pytorch_npu_helper.hpp"
 
 using namespace NPU_NAME_SPACE;
+using namespace std;
 
 void sigmoid_focal_loss_forward_npu(Tensor input, Tensor target, Tensor weight,
                                     Tensor output, float gamma, float alpha) {
-  at::Tensor target_y = at::reshape(target, input.sizes());
+  int64_t n_class = input.size(1);
+  at::Tensor target_y = at::ones_like(input);
+  if(n_class == 1) {
+    target_y = at::reshape(target, input.sizes());
+    target_y = at::mul(target_y, -1.0);
+    target_y = at::add(target_y, 1.0);
+  }
+  else {
+    target_y = at_npu::native::NPUNativeFunctions::one_hot(target, n_class);
+  }
   target_y =
       at_npu::native::NPUNativeFunctions::npu_dtype_cast(target_y, at::kInt);
   int64_t weight_size = weight.size(0);
@@ -14,14 +24,15 @@ void sigmoid_focal_loss_forward_npu(Tensor input, Tensor target, Tensor weight,
                                                                  input.sizes());
   }
   OpCommand cmd;
+  string reduction = "none";
   cmd.Name("SigmoidFocalLoss")
       .Input(input)
       .Input(target_y)
       .Input(weight_y)
       .Output(output)
       .Attr("gamma", gamma)
       .Attr("alpha", alpha)
-      .Attr("reduction", "none")
+      .Attr("reduction", reduction)
       .Run();
 }
 
@@ -31,7 +42,16 @@ void sigmoid_focal_loss_forward_impl(Tensor input, Tensor target, Tensor weight,
 void sigmoid_focal_loss_backward_npu(Tensor input, Tensor target, Tensor weight,
                                      Tensor grad_input, float gamma,
                                      float alpha) {
-  at::Tensor target_y = at::reshape(target, input.sizes());
+  int64_t n_class = input.size(1);
+  at::Tensor target_y = at::ones_like(input);
+  if(n_class == 1) {
+    target_y = at::reshape(target, input.sizes());
+  }
+  else {
+    target_y = at_npu::native::NPUNativeFunctions::one_hot(target, n_class);
+    target_y = at::mul(target_y, -1.0);
+    target_y = at::add(target_y, 1.0);
+  }
   target_y =
       at_npu::native::NPUNativeFunctions::npu_dtype_cast(target_y, at::kInt);
   at::Tensor grad_up = at::ones_like(input);
@@ -42,6 +62,7 @@ void sigmoid_focal_loss_backward_npu(Tensor input, Tensor target, Tensor weight,
                                                                  input.sizes());
   }
   OpCommand cmd;
+  string reduction = "none";
   cmd.Name("SigmoidFocalLossGrad")
       .Input(input)
       .Input(target_y)
@@ -50,7 +71,7 @@ void sigmoid_focal_loss_backward_npu(Tensor input, Tensor target, Tensor weight,
       .Output(grad_input)
       .Attr("gamma", gamma)
       .Attr("alpha", alpha)
-      .Attr("reduction", "none")
+      .Attr("reduction", reduction)
       .Run();
 }
 
@@ -71,16 +92,25 @@ void softmax_focal_loss_forward_npu(Tensor input, Tensor target, Tensor weight,
     weight_y = at_npu::native::NPUNativeFunctions::npu_broadcast(weight,
                                                                  input.sizes());
   }
+  at::Tensor op_output = at::ones_like(input);
   OpCommand cmd;
+  string reduction = "none";
   cmd.Name("SoftmaxFocalLoss")
       .Input(input)
       .Input(target_y)
       .Input(weight_y)
-      .Output(output)
+      .Output(op_output)
       .Attr("gamma", gamma)
       .Attr("alpha", alpha)
-      .Attr("reduction", "none")
+      .Attr("reduction", reduction)
       .Run();
+  int64_t n_batch = input.size(0);
+  c10::SmallVector<int64_t, 2> offsets = {0,0};
+  c10::SmallVector<int64_t, 2> sizes = {n_batch,1};
+  at::IntArrayRef offset = at::IntArrayRef(offsets);
+  at::IntArrayRef size = at::IntArrayRef(sizes);
+  at_npu::native::NPUNativeFunctions::npu_slice_out(op_output, offset,
+                                                    size, output);
 }
 
 void softmax_focal_loss_forward_impl(Tensor input, Tensor target, Tensor weight,
@@ -102,8 +132,8 @@ void softmax_focal_loss_backward_npu(Tensor input, Tensor target, Tensor weight,
     weight_y = at_npu::native::NPUNativeFunctions::npu_broadcast(weight,
                                                                  input.sizes());
   }
-
   OpCommand cmd;
+  string reduction = "none";
   cmd.Name("SoftmaxFocalLossGrad")
       .Input(input)
       .Input(target_y)
@@ -112,7 +142,7 @@ void softmax_focal_loss_backward_npu(Tensor input, Tensor target, Tensor weight,
       .Output(grad_input)
       .Attr("gamma", gamma)
       .Attr("alpha", alpha)
-      .Attr("reduction", "none")
+      .Attr("reduction", reduction)
       .Run();
 }