Fix roi_perspective_transform_op bug (#18522)

* fix transform matrix bug, test=develop

* modify API.spec

paddle/fluid/API.spec

@@ -382,7 +382,7 @@ paddle.fluid.layers.rpn_target_assign (ArgSpec(args=['bbox_pred', 'cls_logits',
 paddle.fluid.layers.retinanet_target_assign (ArgSpec(args=['bbox_pred', 'cls_logits', 'anchor_box', 'anchor_var', 'gt_boxes', 'gt_labels', 'is_crowd', 'im_info', 'num_classes', 'positive_overlap', 'negative_overlap'], varargs=None, keywords=None, defaults=(1, 0.5, 0.4)), ('document', 'fa1d1c9d5e0111684c0db705f86a2595'))
 paddle.fluid.layers.sigmoid_focal_loss (ArgSpec(args=['x', 'label', 'fg_num', 'gamma', 'alpha'], varargs=None, keywords=None, defaults=(2, 0.25)), ('document', 'aeac6aae100173b3fc7f102cf3023a3d'))
 paddle.fluid.layers.anchor_generator (ArgSpec(args=['input', 'anchor_sizes', 'aspect_ratios', 'variance', 'stride', 'offset', 'name'], varargs=None, keywords=None, defaults=(None, None, [0.1, 0.1, 0.2, 0.2], None, 0.5, None)), ('document', '0aaacaf9858b8270a8ab5b0aacdd94b7'))
-paddle.fluid.layers.roi_perspective_transform (ArgSpec(args=['input', 'rois', 'transformed_height', 'transformed_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1.0,)), ('document', '54e3bf70e3bdbd58b3b9b65b3c69a854'))
+paddle.fluid.layers.roi_perspective_transform (ArgSpec(args=['input', 'rois', 'transformed_height', 'transformed_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1.0,)), ('document', 'a82016342789ba9d85737e405f824ff1'))
 paddle.fluid.layers.generate_proposal_labels (ArgSpec(args=['rpn_rois', 'gt_classes', 'is_crowd', 'gt_boxes', 'im_info', 'batch_size_per_im', 'fg_fraction', 'fg_thresh', 'bg_thresh_hi', 'bg_thresh_lo', 'bbox_reg_weights', 'class_nums', 'use_random', 'is_cls_agnostic', 'is_cascade_rcnn'], varargs=None, keywords=None, defaults=(256, 0.25, 0.25, 0.5, 0.0, [0.1, 0.1, 0.2, 0.2], None, True, False, False)), ('document', '69def376b42ef0681d0cc7f53a2dac4b'))
 paddle.fluid.layers.generate_proposals (ArgSpec(args=['scores', 'bbox_deltas', 'im_info', 'anchors', 'variances', 'pre_nms_top_n', 'post_nms_top_n', 'nms_thresh', 'min_size', 'eta', 'name'], varargs=None, keywords=None, defaults=(6000, 1000, 0.5, 0.1, 1.0, None)), ('document', 'b7d707822b6af2a586bce608040235b1'))
 paddle.fluid.layers.generate_mask_labels (ArgSpec(args=['im_info', 'gt_classes', 'is_crowd', 'gt_segms', 'rois', 'labels_int32', 'num_classes', 'resolution'], varargs=None, keywords=None, defaults=None), ('document', 'b319b10ddaf17fb4ddf03518685a17ef'))

paddle/fluid/operators/detection/roi_perspective_transform_op.cc

@@ -272,6 +272,9 @@ class CPUROIPerspectiveTransformOpKernel : public framework::OpKernel<T> {
     T* output_data = out->mutable_data<T>(ctx.GetPlace());
     const T* rois_data = rois->data<T>();
 
+    T* transform_matrix =
+        out_transform_matrix->mutable_data<T>({rois_num, 9}, ctx.GetPlace());
+
     for (int n = 0; n < rois_num; ++n) {
       const T* n_rois = rois_data + n * 8;
       T roi_x[4];
@@ -282,11 +285,12 @@ class CPUROIPerspectiveTransformOpKernel : public framework::OpKernel<T> {
       }
       int image_id = roi2image_data[n];
       // Get transform matrix
-      T* transform_matrix =
-          out_transform_matrix->mutable_data<T>({9}, ctx.GetPlace());
+      T matrix[9];
       get_transform_matrix<T>(transformed_width, transformed_height, roi_x,
-                              roi_y, transform_matrix);
-
+                              roi_y, matrix);
+      for (int i = 0; i < 9; i++) {
+        transform_matrix[n * 9 + i] = matrix[i];
+      }
       for (int c = 0; c < channels; ++c) {
         for (int out_h = 0; out_h < transformed_height; ++out_h) {
           for (int out_w = 0; out_w < transformed_width; ++out_w) {
@@ -295,7 +299,7 @@ class CPUROIPerspectiveTransformOpKernel : public framework::OpKernel<T> {
                 c * transformed_height * transformed_width +
                 out_h * transformed_width + out_w;
             T in_w, in_h;
-            get_source_coords<T>(transform_matrix, out_w, out_h, &in_w, &in_h);
+            get_source_coords<T>(matrix, out_w, out_h, &in_w, &in_h);
             if (in_quad<T>(in_w, in_h, roi_x, roi_y)) {
               if (GT<T>(-0.5, in_w) ||
                   GT<T>(in_w, static_cast<T>(in_width - 0.5)) ||
@@ -507,7 +511,7 @@ class ROIPerspectiveTransformOp : public framework::OperatorWithKernel {
                                       static_cast<int64_t>(transformed_width)});
     auto mask_dims = framework::make_ddim(mask_dims_v);
 
-    std::vector<int64_t> matrix_dims_v(9);
+    std::vector<int64_t> matrix_dims_v({rois_dims[0], 9});
     auto matrix_dims = framework::make_ddim(matrix_dims_v);
 
     ctx->SetOutputDim("Out", out_dims);
@@ -580,7 +584,7 @@ class ROIPerspectiveTransformOpMaker
               "(Tensor), "
               "The output transform matrix of ROIPerspectiveTransformOp is a "
               "1-D tensor with shape "
-              "(9,).");
+              "(num_rois, 9).");
     AddOutput("Out2InIdx",
               "(Tensor), "
               "An intermediate tensor used to map indexes of input feature map "

paddle/fluid/operators/detection/roi_perspective_transform_op.cu

@@ -274,11 +274,14 @@ __device__ void get_transform_matrix(const int transformed_width,
 }
 
 template <typename T>
-__global__ void RoiTransformKernel(
-    const float* input_data, const float* rois_data, const int* roi2image_data,
-    int num_rois, int in_height, int in_width, int channels,
-    int transformed_height, int transformed_width, float spatial_scale,
-    T* output_data, int* out2in_idx, T* out2in_w, int* mask, T* matrix) {
+__global__ void RoiTransformKernel(const float* input_data,
+                                   const float* rois_data,
+                                   const int* roi2image_data, int num_rois,
+                                   int in_height, int in_width, int channels,
+                                   int transformed_height,
+                                   int transformed_width, float spatial_scale,
+                                   T* output_data, int* out2in_idx, T* out2in_w,
+                                   int* mask, T* transform_matrix) {
   int output_size =
       num_rois * transformed_height * transformed_width * channels;
 
@@ -303,9 +306,12 @@ __global__ void RoiTransformKernel(
     }
 
     // Get transform matrix
+    T matrix[9];
     get_transform_matrix<T>(transformed_width, transformed_height, roi_x, roi_y,
                             matrix);
-
+    for (int i = 0; i < 9; i++) {
+      transform_matrix[n * 9 + i] = matrix[i];
+    }
     // Get source coords
     T in_w;
     T in_h;
@@ -389,7 +395,8 @@ class CUDAROIPerspectiveTransformOpKernel : public framework::OpKernel<T> {
     int grid = (out_size + block - 1) / block;
 
     // Get transform matrix
-    T* matrix = out_transform_matrix->mutable_data<T>({9}, ctx.GetPlace());
+    T* matrix =
+        out_transform_matrix->mutable_data<T>({rois_num, 9}, ctx.GetPlace());
 
     RoiTransformKernel<T><<<grid, block, 0, stream>>>(
         input_data, rois_data, roi2image_dev.data<int>(), rois_num, in_height,

python/paddle/fluid/layers/detection.py

@@ -2108,7 +2108,7 @@ def roi_perspective_transform(input,
             (num_rois, 1, transformed_h, transformed_w).
 
             transform_matrix: The transform matrix of ROIPerspectiveTransformOp which is
-            a 1-D tensor with shape (9,).
+            a 2-D tensor with shape (num_rois, 9).
 
     Examples:
         .. code-block:: python

python/paddle/fluid/tests/unittests/test_roi_perspective_transform_op.py

@@ -200,6 +200,7 @@ def roi_transform(in_data, rois, rois_lod, transformed_height,
     out = np.zeros([rois_num, channels, transformed_height, transformed_width])
     mask = np.zeros(
         [rois_num, 1, transformed_height, transformed_width]).astype('int')
+    matrix = np.zeros([rois_num, 9], dtype=in_data.dtype)
     for n in range(rois_num):
         roi_x = []
         roi_y = []
@@ -209,7 +210,7 @@ def roi_transform(in_data, rois, rois_lod, transformed_height,
         image_id = roi2image[n]
         transform_matrix = get_transform_matrix(
             transformed_width, transformed_height, roi_x, roi_y)
-
+        matrix[n] = transform_matrix
         for c in range(channels):
             for out_h in range(transformed_height):
                 for out_w in range(transformed_width):
@@ -224,7 +225,7 @@ def roi_transform(in_data, rois, rois_lod, transformed_height,
                     else:
                         out[n][c][out_h][out_w] = 0.0
                         mask[n][0][out_h][out_w] = 0
-    return out.astype("float32"), mask, transform_matrix
+    return out.astype("float32"), mask, matrix
 
 
 class TestROIPoolOp(OpTest):