[Feature]: Support corner_pool related custom operators for onnxruntime in mmcv (#997)

* supports for onnxruntime custom op `mmcv::MMCVTopPool`

* supports for onnxruntime custom op `mmcv::MMCVCornerPool`, involving TopPool, BottomPool, LeftPool and RightPool

* add unittest for corner_pool

* supports mmcv::CornerPool without memcpy

* add docs for mmcv::CornerPool

* re-add docs for mmcv::CornerPool

* fix output dtype doc

* reformat

* format with pre-commit

* format

* fix lint error, by using google clang-format style for c/c++

docs/onnxruntime_custom_ops.md

@@ -27,6 +27,12 @@
     - [Inputs](#inputs-3)
     - [Outputs](#outputs-3)
     - [Type Constraints](#type-constraints-3)
+  - [CornerPool](#cornerpool)
+    - [Description](#description-4)
+    - [Parameters](#parameters-4)
+    - [Inputs](#inputs-4)
+    - [Outputs](#outputs-4)
+    - [Type Constraints](#type-constraints-4)
 
 <!-- TOC -->
 
@@ -171,3 +177,33 @@ Perform sample from `input` with pixel locations from `grid`.
 ### Type Constraints
 
 - T:tensor(float32, Linear)
+
+## CornerPool
+
+### Description
+
+Perform CornerPool on `input` features. Read [CornerNet -- Detecting Objects as Paired Keypoints](https://arxiv.org/abs/1808.01244) for more details.
+
+### Parameters
+
+| Type    | Parameter       | Description                                                      |
+| ------- | --------------- | ---------------------------------------------------------------- |
+| `int`   | `mode`          | corner pool mode, (0: `top`, 1: `bottom`, 2: `left`, 3: `right`) |
+
+### Inputs
+
+<dl>
+<dt><tt>input</tt>: T</dt>
+<dd>Input features. 4-D tensor of shape (N, C, H, W). N is the batch size.</dd>
+</dl>
+
+### Outputs
+
+<dl>
+<dt><tt>output</tt>: T</dt>
+<dd>Output the pooled features. 4-D tensor of shape (N, C, H, W).</dd>
+</dl>
+
+### Type Constraints
+
+- T:tensor(float32)

docs/onnxruntime_op.md

@@ -21,6 +21,7 @@
 |     [RoIAlign](onnxruntime_custom_ops.md#roialign)     |   Y   |   N   |     1.2.5     |
 |          [NMS](onnxruntime_custom_ops.md#nms)          |   Y   |   N   |     1.2.7     |
 | [grid_sampler](onnxruntime_custom_ops.md#grid_sampler) |   Y   |   N   |    master     |
+| [CornerPool](onnxruntime_custom_ops.md#cornerpool)    |   Y   |   N   |    master     |
 
 ## How to build custom operators for ONNX Runtime
 

mmcv/ops/corner_pool.py

@@ -10,9 +10,17 @@ ext_module = ext_loader.load_ext('_ext', [
     'right_pool_forward', 'right_pool_backward'
 ])
 
+_mode_dict = {'top': 0, 'bottom': 1, 'left': 2, 'right': 3}
+
 
 class TopPoolFunction(Function):
 
+    @staticmethod
+    def symbolic(g, input):
+        output = g.op(
+            'mmcv::MMCVCornerPool', input, mode_i=int(_mode_dict['top']))
+        return output
+
     @staticmethod
     def forward(ctx, input):
         output = ext_module.top_pool_forward(input)
@@ -28,6 +36,12 @@ class TopPoolFunction(Function):
 
 class BottomPoolFunction(Function):
 
+    @staticmethod
+    def symbolic(g, input):
+        output = g.op(
+            'mmcv::MMCVCornerPool', input, mode_i=int(_mode_dict['bottom']))
+        return output
+
     @staticmethod
     def forward(ctx, input):
         output = ext_module.bottom_pool_forward(input)
@@ -43,6 +57,12 @@ class BottomPoolFunction(Function):
 
 class LeftPoolFunction(Function):
 
+    @staticmethod
+    def symbolic(g, input):
+        output = g.op(
+            'mmcv::MMCVCornerPool', input, mode_i=int(_mode_dict['left']))
+        return output
+
     @staticmethod
     def forward(ctx, input):
         output = ext_module.left_pool_forward(input)
@@ -58,6 +78,12 @@ class LeftPoolFunction(Function):
 
 class RightPoolFunction(Function):
 
+    @staticmethod
+    def symbolic(g, input):
+        output = g.op(
+            'mmcv::MMCVCornerPool', input, mode_i=int(_mode_dict['right']))
+        return output
+
     @staticmethod
     def forward(ctx, input):
         output = ext_module.right_pool_forward(input)

mmcv/ops/csrc/onnxruntime/corner_pool.h

@@ -0,0 +1,45 @@
+#ifndef ONNXRUNTIME_CORNER_POOL_H
+#define ONNXRUNTIME_CORNER_POOL_H
+
+#include <assert.h>
+#include <onnxruntime_cxx_api.h>
+
+struct MMCVCornerPoolKernel {
+ public:
+  MMCVCornerPoolKernel(Ort::CustomOpApi ort, const OrtKernelInfo* info)
+      : ort_(ort) {
+    mode_ = ort_.KernelInfoGetAttribute<int64_t>(info, "mode");
+  }
+
+  void Compute(OrtKernelContext* context);
+
+ private:
+  Ort::CustomOpApi ort_;
+
+  int64_t mode_;
+};
+
+struct MMCVCornerPoolCustomOp
+    : Ort::CustomOpBase<MMCVCornerPoolCustomOp, MMCVCornerPoolKernel> {
+  void* CreateKernel(Ort::CustomOpApi api, const OrtKernelInfo* info) {
+    return new MMCVCornerPoolKernel(api, info);
+  }
+
+  const char* GetName() const { return "MMCVCornerPool"; }
+
+  size_t GetInputTypeCount() const { return 1; }
+  ONNXTensorElementDataType GetInputType(size_t) const {
+    return ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT;
+  }
+
+  size_t GetOutputTypeCount() const { return 1; }
+  ONNXTensorElementDataType GetOutputType(size_t) const {
+    return ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT;
+  }
+
+  // force cpu
+  const char* GetExecutionProviderType() const {
+    return "CPUExecutionProvider";
+  }
+};
+#endif  // ONNXRUNTIME_CORNER_POOL_H

mmcv/ops/csrc/onnxruntime/cpu/corner_pool.cpp

@@ -0,0 +1,122 @@
+#include "corner_pool.h"
+
+#include "../ort_mmcv_utils.h"
+
+void TopPoolForwardCPU(const float *input, float *output, const int batch_size,
+                       const int channels, const int height, const int width) {
+  for (int n = 0; n < batch_size; n++) {
+    int index_n = n * channels * width * height;
+    for (int c = 0; c < channels; c++) {
+      int index_n_c = index_n + c * width * height;
+      for (int w = 0; w < width; w++) {
+        // directly copy the most bottom value from input to output
+        output[index_n_c + (height - 1) * width + w] =
+            input[index_n_c + (height - 1) * width + w];
+        // do top_pool
+        for (int h = height - 2; h >= 0; h--) {
+          output[index_n_c + h * width + w] =
+              std::max(output[index_n_c + (h + 1) * width + w],
+                       input[index_n_c + h * width + w]);
+        }  // for h
+      }    // for w
+    }      // for c
+  }        // for n
+}
+
+void BottomPoolForwardCPU(const float *input, float *output,
+                          const int batch_size, const int channels,
+                          const int height, const int width) {
+  for (int n = 0; n < batch_size; n++) {
+    int index_n = n * channels * width * height;
+    for (int c = 0; c < channels; c++) {
+      int index_n_c = index_n + c * width * height;
+      for (int w = 0; w < width; w++) {
+        // directly copy the most top value from input to output
+        output[index_n_c + w] = input[index_n_c + w];
+        // do top_pool
+        for (int h = 1; h < height; h++) {
+          output[index_n_c + h * width + w] =
+              std::max(output[index_n_c + (h - 1) * width + w],
+                       input[index_n_c + h * width + w]);
+        }  // for h
+      }    // for w
+    }      // for c
+  }        // for n
+}
+
+void LeftPoolForwardCPU(const float *input, float *output, const int batch_size,
+                        const int channels, const int height, const int width) {
+  for (int n = 0; n < batch_size; n++) {
+    int index_n = n * channels * width * height;
+    for (int c = 0; c < channels; c++) {
+      int index_n_c = index_n + c * width * height;
+      for (int h = 0; h < height; h++) {
+        // directly copy the most right value from input to output
+        output[index_n_c + h * width + width - 1] =
+            input[index_n_c + h * width + width - 1];
+        // do left_pool
+        for (int w = width - 2; w >= 0; w--) {
+          output[index_n_c + h * width + w] =
+              std::max(output[index_n_c + h * width + w + 1],
+                       input[index_n_c + h * width + w]);
+        }  // for w
+      }    // for h
+    }      // for c
+  }        // for n
+}
+
+void RightPoolForwardCPU(const float *input, float *output,
+                         const int batch_size, const int channels,
+                         const int height, const int width) {
+  for (int n = 0; n < batch_size; n++) {
+    int index_n = n * channels * width * height;
+    for (int c = 0; c < channels; c++) {
+      int index_n_c = index_n + c * width * height;
+      for (int h = 0; h < height; h++) {
+        // directly copy the most left value from input to output
+        output[index_n_c + h * width] = input[index_n_c + h * width];
+        // do right_pool
+        for (int w = 1; w < width; w++) {
+          output[index_n_c + h * width + w] =
+              std::max(output[index_n_c + h * width + w - 1],
+                       input[index_n_c + h * width + w]);
+        }  // for w
+      }    // for h
+    }      // for c
+  }        // for n
+}
+
+void MMCVCornerPoolKernel::Compute(OrtKernelContext *context) {
+  const int mode = int(mode_);
+  typedef float T;
+  const OrtValue *input = ort_.KernelContext_GetInput(context, 0);
+  const T *input_data =
+      reinterpret_cast<const float *>(ort_.GetTensorData<T>(input));
+
+  // get output memory
+  OrtTensorDimensions out_dimensions(ort_, input);
+  OrtValue *output = ort_.KernelContext_GetOutput(
+      context, 0, out_dimensions.data(), out_dimensions.size());
+  T *output_data = ort_.GetTensorMutableData<T>(output);
+
+  // 'top': 0, 'bottom': 1, 'left': 2, 'right':3
+  assert(mode == 0 || mode == 1 || mode == 2 || mode == 3);
+
+  // do corner_pool
+  int batch_size = out_dimensions.data()[0];
+  int input_channels = out_dimensions.data()[1];
+  int input_height = out_dimensions.data()[2];
+  int input_width = out_dimensions.data()[3];
+  if (mode == 0)
+    TopPoolForwardCPU(input_data, output_data, batch_size, input_channels,
+                      input_height, input_width);
+  else if (mode == 1)
+    BottomPoolForwardCPU(input_data, output_data, batch_size, input_channels,
+                         input_height, input_width);
+  else if (mode == 2)
+    LeftPoolForwardCPU(input_data, output_data, batch_size, input_channels,
+                       input_height, input_width);
+  else
+    RightPoolForwardCPU(input_data, output_data, batch_size, input_channels,
+                        input_height, input_width);
+}

mmcv/ops/csrc/onnxruntime/cpu/onnxruntime_register.cpp

@@ -1,5 +1,6 @@
 #include "onnxruntime_register.h"
 
+#include "corner_pool.h"
 #include "grid_sample.h"
 #include "nms.h"
 #include "ort_mmcv_utils.h"
@@ -13,6 +14,7 @@ NmsOp c_NmsOp;
 MMCVRoiAlignCustomOp c_MMCVRoiAlignCustomOp;
 MMCVRoIAlignRotatedCustomOp c_MMCVRoIAlignRotatedCustomOp;
 GridSampleOp c_GridSampleOp;
+MMCVCornerPoolCustomOp c_MMCVCornerPoolCustomOp;
 
 OrtStatus *ORT_API_CALL RegisterCustomOps(OrtSessionOptions *options,
                                           const OrtApiBase *api) {
@@ -45,5 +47,10 @@ OrtStatus *ORT_API_CALL RegisterCustomOps(OrtSessionOptions *options,
     return status;
   }
 
+  if (auto status =
+          ortApi->CustomOpDomain_Add(domain, &c_MMCVCornerPoolCustomOp)) {
+    return status;
+  }
+
   return ortApi->AddCustomOpDomain(options, domain);
 }

tests/test_ops/test_onnx.py

@@ -448,3 +448,49 @@ def test_interpolate():
     if os.path.exists(onnx_file):
         os.remove(onnx_file)
     assert np.allclose(pytorch_result, onnx_result, atol=1e-3)
+
+
+@pytest.mark.parametrize('mode', ['top', 'bottom', 'left', 'right'])
+def test_corner_pool(mode, opset=11):
+    if torch.__version__ == 'parrots':
+        pytest.skip('onnx is not supported in parrots directly')
+
+    from mmcv.ops import get_onnxruntime_op_path
+    ort_custom_op_path = get_onnxruntime_op_path()
+    if not os.path.exists(ort_custom_op_path):
+        pytest.skip('custom ops for onnxruntime are not compiled.')
+
+    from mmcv.ops.corner_pool import CornerPool
+
+    def corner_pool_func(input):
+        corner_pool_module = CornerPool(mode)
+        return corner_pool_module.corner_pool.apply(input)
+
+    wrapped_model = WrapFunction(corner_pool_func).eval()
+
+    input = torch.rand((2, 3, 9, 12))  # (n,c,h,w)
+
+    with torch.no_grad():
+        torch.onnx.export(
+            wrapped_model,
+            input,
+            onnx_file,
+            export_params=True,
+            keep_initializers_as_inputs=True,
+            input_names=['input'],
+            output_names=['output'],
+            opset_version=opset)
+
+    onnx_model = onnx.load(onnx_file)
+    input_all = [node.name for node in onnx_model.graph.input]
+    input_initializer = [node.name for node in onnx_model.graph.initializer]
+    net_feed_input = list(set(input_all) - set(input_initializer))
+    assert (len(net_feed_input) == 1)
+
+    session_options = rt.SessionOptions()
+    session_options.register_custom_ops_library(ort_custom_op_path)
+    sess = rt.InferenceSession(onnx_file, session_options)
+    ort_result = sess.run(None, {'input': input.detach().numpy()})
+    pytorch_results = wrapped_model(input.clone())
+    os.remove(onnx_file)
+    assert np.allclose(pytorch_results, ort_result, atol=1e-5)