[Feature]: add custom op roialign for onnxruntime (#724)

* add ort custom op roialign

* Update roi_align.py

* Update test_onnx.py

* Update test_onnx.py

* Update test_onnx.py

* Update test_onnx.py

* Update onnxruntime_register.cpp

* Update roiAlign.h

* Update roiAlign.cpp

* lint modification

* update roiAlign.cpp

* lint check

* lint check

* lint fix

* lint fix

* fix lint

* add link to commit
Co-authored-by: maningsheng <maningsheng@sensetime.com>

docs/onnxruntime_op.md

@@ -15,9 +15,10 @@
 
 ## List of operators for ONNX Runtime supported in MMCV
 
-| Operator |  CPU  |  GPU  | Note  |
-| :------: | :---: | :---: | :---: |
-| SoftNMS  |   Y   |   N   | None  |
+| Operator |  CPU  |  GPU  |                                                Note                                                 |
+| :------: | :---: | :---: | :-------------------------------------------------------------------------------------------------: |
+| SoftNMS  |   Y   |   N   | commit [94810f](https://github.com/open-mmlab/mmcv/commit/94810f2297871d0ea3ca650dcb2e842f5374d998) |
+| RoiAlign |   Y   |   N   |                                                None                                                 |
 
 ## How to build custom operators for ONNX Runtime
 

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

 #include "onnxruntime_register.h"
 
 #include "ort_mmcv_utils.h"
+#include "roi_align.h"
 #include "soft_nms.h"
 
 const char *c_MMCVOpDomain = "mmcv";
 SoftNmsOp c_SoftNmsOp;
+MMCVRoiAlignCustomOp c_MMCVRoiAlignCustomOp;
 
 OrtStatus *ORT_API_CALL RegisterCustomOps(OrtSessionOptions *options,
                                           const OrtApiBase *api) {
@@ -19,5 +21,10 @@ OrtStatus *ORT_API_CALL RegisterCustomOps(OrtSessionOptions *options,
     return status;
   }
 
+  if (auto status =
+          ortApi->CustomOpDomain_Add(domain, &c_MMCVRoiAlignCustomOp)) {
+    return status;
+  }
+
   return ortApi->AddCustomOpDomain(options, domain);
 }

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

+#include "roi_align.h"
+
+#include "../ort_mmcv_utils.h"
+
+// implementation taken from Caffe2
+struct PreCalc {
+  int pos1;
+  int pos2;
+  int pos3;
+  int pos4;
+  float w1;
+  float w2;
+  float w3;
+  float w4;
+};
+
+void pre_calc_for_bilinear_interpolate(
+    const int height, const int width, const int pooled_height,
+    const int pooled_width, const int iy_upper, const int ix_upper,
+    float roi_start_h, float roi_start_w, float bin_size_h, float bin_size_w,
+    int roi_bin_grid_h, int roi_bin_grid_w, std::vector<PreCalc> &pre_calc) {
+  int pre_calc_index = 0;
+  for (int ph = 0; ph < pooled_height; ph++) {
+    for (int pw = 0; pw < pooled_width; pw++) {
+      for (int iy = 0; iy < iy_upper; iy++) {
+        const float yy =
+            roi_start_h + ph * bin_size_h +
+            static_cast<float>(iy + .5f) * bin_size_h /
+                static_cast<float>(roi_bin_grid_h);  // e.g., 0.5, 1.5
+        for (int ix = 0; ix < ix_upper; ix++) {
+          const float xx = roi_start_w + pw * bin_size_w +
+                           static_cast<float>(ix + .5f) * bin_size_w /
+                               static_cast<float>(roi_bin_grid_w);
+
+          float x = xx;
+          float y = yy;
+          // deal with: inverse elements are out of feature map boundary
+          if (y < -1.0 || y > height || x < -1.0 || x > width) {
+            // empty
+            PreCalc pc;
+            pc.pos1 = 0;
+            pc.pos2 = 0;
+            pc.pos3 = 0;
+            pc.pos4 = 0;
+            pc.w1 = 0;
+            pc.w2 = 0;
+            pc.w3 = 0;
+            pc.w4 = 0;
+            pre_calc[pre_calc_index] = pc;
+            pre_calc_index += 1;
+            continue;
+          }
+
+          if (y <= 0) {
+            y = 0;
+          }
+          if (x <= 0) {
+            x = 0;
+          }
+
+          int y_low = (int)y;
+          int x_low = (int)x;
+          int y_high;
+          int x_high;
+
+          if (y_low >= height - 1) {
+            y_high = y_low = height - 1;
+            y = (float)y_low;
+          } else {
+            y_high = y_low + 1;
+          }
+
+          if (x_low >= width - 1) {
+            x_high = x_low = width - 1;
+            x = (float)x_low;
+          } else {
+            x_high = x_low + 1;
+          }
+
+          float ly = y - y_low;
+          float lx = x - x_low;
+          float hy = 1. - ly, hx = 1. - lx;
+          float w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx;
+
+          // save weights and indices
+          PreCalc pc;
+          pc.pos1 = y_low * width + x_low;
+          pc.pos2 = y_low * width + x_high;
+          pc.pos3 = y_high * width + x_low;
+          pc.pos4 = y_high * width + x_high;
+          pc.w1 = w1;
+          pc.w2 = w2;
+          pc.w3 = w3;
+          pc.w4 = w4;
+          pre_calc[pre_calc_index] = pc;
+
+          pre_calc_index += 1;
+        }
+      }
+    }
+  }
+}
+
+void ROIAlignForwardCPU(const int nthreads, const float *input,
+                        const float *rois, float *output, float *argmax_y,
+                        float *argmax_x, const int pooled_height,
+                        const int pooled_width, const float spatial_scale,
+                        const int sampling_ratio,
+                        const int pool_mode,  // 0 - max pool, 1 - avg pool
+                        const bool aligned, const int channels,
+                        const int height, const int width) {
+  int n_rois = nthreads / channels / pooled_width / pooled_height;
+  // (n, c, ph, pw) is an element in the pooled output
+  // can be parallelized using omp
+  // #pragma omp parallel for num_threads(32)
+  for (int n = 0; n < n_rois; n++) {
+    int index_n = n * channels * pooled_width * pooled_height;
+
+    const float *offset_rois = rois + n * 5;
+    int roi_batch_ind = offset_rois[0];
+
+    // Do not use rounding; this implementation detail is critical
+    float offset = aligned ? (float)0.5 : (float)0.0;
+    float roi_start_w = offset_rois[1] * spatial_scale - offset;
+    float roi_start_h = offset_rois[2] * spatial_scale - offset;
+    float roi_end_w = offset_rois[3] * spatial_scale - offset;
+    float roi_end_h = offset_rois[4] * spatial_scale - offset;
+
+    float roi_width = roi_end_w - roi_start_w;
+    float roi_height = roi_end_h - roi_start_h;
+    if (aligned) {
+      /*AT_ASSERTM(roi_width >= 0 && roi_height >= 0,
+                 "ROIs in ROIAlign cannot have non-negative size!");*/
+      assert(roi_width >= 0 && roi_height >= 0);
+    } else {  // for backward-compatibility only
+      roi_width = std::max(roi_width, (float)1.);
+      roi_height = std::max(roi_height, (float)1.);
+    }
+    float bin_size_h =
+        static_cast<float>(roi_height) / static_cast<float>(pooled_height);
+    float bin_size_w =
+        static_cast<float>(roi_width) / static_cast<float>(pooled_width);
+
+    // We use roi_bin_grid to sample the grid and mimic integral
+    int roi_bin_grid_h = (sampling_ratio > 0)
+                             ? sampling_ratio
+                             : ceil(roi_height / pooled_height);  // e.g., = 2
+    int roi_bin_grid_w =
+        (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width);
+
+    // When the grid is empty, output zeros == 0/1, instead of NaN.
+    const float count =
+        std::max(roi_bin_grid_h * roi_bin_grid_w, 1);  // e.g. = 4
+
+    // we want to precalculate indices and weights shared by all channels,
+    // this is the key point of optimization
+    std::vector<PreCalc> pre_calc(roi_bin_grid_h * roi_bin_grid_w *
+                                  pooled_width * pooled_height);
+    pre_calc_for_bilinear_interpolate(
+        height, width, pooled_height, pooled_width, roi_bin_grid_h,
+        roi_bin_grid_w, roi_start_h, roi_start_w, bin_size_h, bin_size_w,
+        roi_bin_grid_h, roi_bin_grid_w, pre_calc);
+
+    for (int c = 0; c < channels; c++) {
+      int index_n_c = index_n + c * pooled_width * pooled_height;
+      const float *offset_input =
+          input + (roi_batch_ind * channels + c) * height * width;
+      int pre_calc_index = 0;
+
+      for (int ph = 0; ph < pooled_height; ph++) {
+        for (int pw = 0; pw < pooled_width; pw++) {
+          int index = index_n_c + ph * pooled_width + pw;
+
+          float output_val = 0.;
+          float maxval = -10000;
+          float maxidx_y = -1.f, maxidx_x = -1.f;
+          for (int iy = 0; iy < roi_bin_grid_h; iy++) {
+            const float y = roi_start_h + ph * bin_size_h +
+                            static_cast<float>(iy + .5f) * bin_size_h /
+                                static_cast<float>(roi_bin_grid_h);
+            for (int ix = 0; ix < roi_bin_grid_w; ix++) {
+              const float x = roi_start_w + pw * bin_size_w +
+                              static_cast<float>(ix + .5f) * bin_size_w /
+                                  static_cast<float>(roi_bin_grid_w);
+              PreCalc pc = pre_calc[pre_calc_index];
+              float val = pc.w1 * offset_input[pc.pos1] +
+                          pc.w2 * offset_input[pc.pos2] +
+                          pc.w3 * offset_input[pc.pos3] +
+                          pc.w4 * offset_input[pc.pos4];
+              if (val > maxval) {
+                maxval = val;
+                maxidx_y = y;
+                maxidx_x = x;
+              }
+              output_val += val;
+              pre_calc_index += 1;
+            }
+          }
+          if (pool_mode == 0) {
+            // We do max pooling inside a bin
+            output[index] = maxval;
+            argmax_y[index] = maxidx_y;
+            argmax_x[index] = maxidx_x;
+          } else if (pool_mode == 1) {
+            // We do average (integral) pooling inside a bin
+            output[index] = output_val / count;
+          }  // if
+        }    // for pw
+      }      // for ph
+    }        // for c
+  }          // for n
+}
+
+void MMCVRoiAlignKernel::Compute(OrtKernelContext *context) {
+  // Setup inputs
+  const OrtValue *input_X = ort_.KernelContext_GetInput(context, 0);
+  const float *X_data =
+      reinterpret_cast<const float *>(ort_.GetTensorData<float>(input_X));
+  const OrtValue *input_rois = ort_.KernelContext_GetInput(context, 1);
+  const float *rois = reinterpret_cast<const float *>(
+      ort_.GetTensorData<const float *>(input_rois));
+
+  // Setup output
+  OrtTensorDimensions out_dimensions(ort_, input_X);
+  OrtTensorDimensions roi_dimensions(ort_, input_rois);
+
+  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];
+
+  out_dimensions.data()[0] = roi_dimensions.data()[0];
+  out_dimensions.data()[2] = aligned_height_;
+  out_dimensions.data()[3] = aligned_width_;
+
+  OrtValue *output = ort_.KernelContext_GetOutput(
+      context, 0, out_dimensions.data(), out_dimensions.size());
+  float *out = ort_.GetTensorMutableData<float>(output);
+  OrtTensorTypeAndShapeInfo *output_info = ort_.GetTensorTypeAndShape(output);
+  ort_.ReleaseTensorTypeAndShapeInfo(output_info);
+
+  // TODO: forward here
+  int output_size = out_dimensions.data()[0];
+  for (auto i = 1; i < out_dimensions.size(); ++i) {
+    output_size *= out_dimensions.data()[i];
+  }
+
+  int poolMod = 1;
+  if (pool_mode_ == "max") poolMod = 0;
+
+  float *argmax_x = nullptr, *argmax_y = nullptr;
+  if (poolMod == 0) {
+    argmax_y = new float[output_size];
+    argmax_x = new float[output_size];
+  }
+
+  ROIAlignForwardCPU(output_size, X_data, rois, out, argmax_y, argmax_x,
+                     aligned_height_, aligned_width_, spatial_scale_,
+                     sampling_ratio_, poolMod, aligned_, input_channels,
+                     input_height, input_width);
+
+  if (argmax_x) delete argmax_x;
+  if (argmax_y) delete argmax_y;
+}

mmcv/ops/csrc/onnxruntime/roi_align.h

+#ifndef ONNXRUNTIME_ROI_ALIGN_H
+#define ONNXRUNTIME_ROI_ALIGN_H
+
+#include <assert.h>
+#include <onnxruntime_cxx_api.h>
+
+#include <cmath>
+#include <mutex>
+#include <string>
+#include <vector>
+
+struct MMCVRoiAlignKernel {
+ public:
+  MMCVRoiAlignKernel(Ort::CustomOpApi ort, const OrtKernelInfo *info)
+      : ort_(ort) {
+    aligned_ = ort_.KernelInfoGetAttribute<int64_t>(info, "aligned");
+    aligned_height_ =
+        ort_.KernelInfoGetAttribute<int64_t>(info, "aligned_height");
+    aligned_width_ =
+        ort_.KernelInfoGetAttribute<int64_t>(info, "aligned_width");
+    pool_mode_ = ort_.KernelInfoGetAttribute<std::string>(info, "pool_mode");
+    sampling_ratio_ =
+        ort_.KernelInfoGetAttribute<int64_t>(info, "sampling_ratio");
+    spatial_scale_ = ort_.KernelInfoGetAttribute<float>(info, "spatial_scale");
+  }
+
+  void Compute(OrtKernelContext *context);
+
+ private:
+  Ort::CustomOpApi ort_;
+
+  int aligned_height_;
+  int aligned_width_;
+  float spatial_scale_;
+  int sampling_ratio_;
+  std::string pool_mode_;
+  int aligned_;
+};
+
+struct MMCVRoiAlignCustomOp
+    : Ort::CustomOpBase<MMCVRoiAlignCustomOp, MMCVRoiAlignKernel> {
+  void *CreateKernel(Ort::CustomOpApi api, const OrtKernelInfo *info) {
+    return new MMCVRoiAlignKernel(api, info);
+  }
+  const char *GetName() const { return "MMCVRoiAlign"; }
+
+  size_t GetInputTypeCount() const { return 2; }
+  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_ROI_ALIGN_H

mmcv/ops/roi_align.py

@@ -15,6 +15,27 @@ class RoIAlignFunction(Function):
     @staticmethod
     def symbolic(g, input, rois, output_size, spatial_scale, sampling_ratio,
                  pool_mode, aligned):
+        has_custom_op = False
+        try:
+            import os.path as osp
+
+            from mmcv.ops import get_onnxruntime_op_path
+            ort_op_path = get_onnxruntime_op_path()
+            has_custom_op = osp.exists(ort_op_path)
+        except ImportError:
+            pass
+        if has_custom_op:
+            return g.op(
+                'mmcv::MMCVRoiAlign',
+                input,
+                rois,
+                aligned_height_i=output_size[0],
+                aligned_width_i=output_size[1],
+                spatial_scale_f=spatial_scale,
+                sampling_ratio_i=max(0, sampling_ratio),
+                pool_mode_s=pool_mode,
+                aligned_i=aligned)
+
         from torch.onnx.symbolic_opset9 import sub, squeeze
         from torch.onnx.symbolic_helper import _slice_helper
         from torch.onnx import TensorProtoDataType

tests/test_ops/test_onnx.py

@@ -139,7 +139,12 @@ def test_softnms():
 
 def test_roialign():
     from mmcv.ops import roi_align
-
+    ort_custom_op_path = ''
+    try:
+        from mmcv.ops import get_onnxruntime_op_path
+        ort_custom_op_path = get_onnxruntime_op_path()
+    except ImportError:
+        pass
     # roi align config
     pool_h = 2
     pool_w = 2
@@ -178,7 +183,11 @@ def test_roialign():
                 keep_initializers_as_inputs=True,
                 input_names=['input', 'rois'],
                 opset_version=11)
+
         onnx_model = onnx.load(onnx_file)
+        session_options = rt.SessionOptions()
+        if os.path.exists(ort_custom_op_path):
+            session_options.register_custom_ops_library(ort_custom_op_path)
 
         # compute onnx_output
         input_all = [node.name for node in onnx_model.graph.input]
@@ -187,7 +196,7 @@ def test_roialign():
         ]
         net_feed_input = list(set(input_all) - set(input_initializer))
         assert (len(net_feed_input) == 2)
-        sess = rt.InferenceSession(onnx_file)
+        sess = rt.InferenceSession(onnx_file, session_options)
         onnx_output = sess.run(None, {
             'input': input.detach().numpy(),
             'rois': rois.detach().numpy()