[0-size Tensor No.358] Add 0-size Tensor support for paddle.vision.ops.generate_proposals [fluid_ops] (#74261)

* Fix

* ci

Author: co63oc

paddle/phi/kernels/cpu/generate_proposals_kernel.cc

@@ -15,6 +15,7 @@
 #include "paddle/phi/kernels/generate_proposals_kernel.h"
 #include "paddle/phi/backends/cpu/cpu_context.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/funcs/detection/nms_util.h"
 #include "paddle/phi/kernels/funcs/gather.h"
 
@@ -314,6 +315,18 @@ void GenerateProposalsKernel(const Context& dev_ctx,
 
   rpn_roi_probs->Resize(common::make_ddim({scores.numel(), 1}));
   dev_ctx.template Alloc<T>(rpn_roi_probs);
+  if (scores.numel() == 0) {
+    rpn_rois->Resize(common::make_ddim({0, 4}));
+    if (rpn_rois_num != nullptr) {
+      rpn_rois_num->Resize(common::make_ddim({}));
+      phi::Full<int64_t, Context>(
+          dev_ctx,
+          phi::IntArray(common::vectorize(rpn_rois_num->dims())),
+          0,
+          rpn_rois_num);
+    }
+    return;
+  }
 
   DenseTensor bbox_deltas_swap, scores_swap;
   bbox_deltas_swap.Resize(common::make_ddim({num, h_bbox, w_bbox, c_bbox}));

paddle/phi/kernels/gpu/generate_proposals_kernel.cu

@@ -27,6 +27,7 @@ namespace cub = hipcub;
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/backends/gpu/gpu_primitives.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/funcs/detection/bbox_util.h"
 #include "paddle/phi/kernels/funcs/for_range.h"
 #include "paddle/phi/kernels/funcs/gather.cu.h"
@@ -490,6 +491,24 @@ void GenerateProposalsKernel(const Context &dev_ctx,
   int64_t h_bbox = bbox_dim[2];
   int64_t w_bbox = bbox_dim[3];
 
+  rpn_rois->Resize(common::make_ddim({bbox_deltas.numel() / 4, 4}));
+  dev_ctx.template Alloc<T>(rpn_rois);
+  rpn_roi_probs->Resize(common::make_ddim({scores.numel(), 1}));
+  dev_ctx.template Alloc<T>(rpn_roi_probs);
+
+  if (scores.numel() == 0) {
+    rpn_rois->Resize(common::make_ddim({0, 4}));
+    if (rpn_rois_num != nullptr) {
+      rpn_rois_num->Resize(common::make_ddim({}));
+      phi::Full<int64_t, Context>(
+          dev_ctx,
+          phi::IntArray(common::vectorize(rpn_rois_num->dims())),
+          0,
+          rpn_rois_num);
+    }
+    return;
+  }
+
   DenseTensor bbox_deltas_swap, scores_swap;
   bbox_deltas_swap.Resize(common::make_ddim({num, h_bbox, w_bbox, c_bbox}));
   dev_ctx.template Alloc<T>(&bbox_deltas_swap);
@@ -506,11 +525,6 @@ void GenerateProposalsKernel(const Context &dev_ctx,
   tmp_anchors.Resize(common::make_ddim({tmp_anchors.numel() / 4, 4}));
   tmp_variances.Resize(common::make_ddim({tmp_variances.numel() / 4, 4}));
 
-  rpn_rois->Resize(common::make_ddim({bbox_deltas.numel() / 4, 4}));
-  dev_ctx.template Alloc<T>(rpn_rois);
-  rpn_roi_probs->Resize(common::make_ddim({scores.numel(), 1}));
-  dev_ctx.template Alloc<T>(rpn_roi_probs);
-
   T *rpn_rois_data = rpn_rois->data<T>();
   T *rpn_roi_probs_data = rpn_roi_probs->data<T>();
 

paddle/phi/kernels/xpu/generate_proposals_kernel.cc

@@ -18,6 +18,7 @@
 #include "paddle/phi/backends/xpu/xpu_context.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/funcs/math_function_impl.h"
 
 #include "paddle/phi/common/memory_utils.h"
@@ -306,6 +307,26 @@ void GenerateProposalsKernel(const Context& dev_ctx,
   int h_bbox = bbox_dim[2];
   int w_bbox = bbox_dim[3];
 
+  // output
+  rpn_rois->Resize(common::make_ddim({bbox_deltas.numel() / 4, 4}));
+  dev_ctx.template Alloc<T>(rpn_rois);
+
+  rpn_roi_probs->Resize(common::make_ddim({scores.numel(), 1}));
+  dev_ctx.template Alloc<T>(rpn_roi_probs);
+
+  if (scores.numel() == 0) {
+    rpn_rois->Resize(common::make_ddim({0, 4}));
+    if (rpn_rois_num != nullptr) {
+      rpn_rois_num->Resize(common::make_ddim({}));
+      phi::Full<int64_t, Context>(
+          dev_ctx,
+          phi::IntArray(common::vectorize(rpn_rois_num->dims())),
+          0,
+          rpn_rois_num);
+    }
+    return;
+  }
+
   DenseTensor bbox_deltas_swap, scores_swap;
   bbox_deltas_swap.Resize(common::make_ddim({num, h_bbox, w_bbox, c_bbox}));
   dev_ctx.template Alloc<T>(&bbox_deltas_swap);
@@ -333,13 +354,6 @@ void GenerateProposalsKernel(const Context& dev_ctx,
   tmp_anchors.Resize(common::make_ddim({tmp_anchors.numel() / 4, 4}));
   tmp_variances.Resize(common::make_ddim({tmp_variances.numel() / 4, 4}));
 
-  // output
-  rpn_rois->Resize(common::make_ddim({bbox_deltas.numel() / 4, 4}));
-  dev_ctx.template Alloc<T>(rpn_rois);
-
-  rpn_roi_probs->Resize(common::make_ddim({scores.numel(), 1}));
-  dev_ctx.template Alloc<T>(rpn_roi_probs);
-
   auto place = dev_ctx.GetPlace();
   auto cpu_place = phi::CPUPlace();
 

test/legacy_test/test_generate_proposals_v2_op.py

@@ -560,6 +560,95 @@ def init_test_output(self):
 #         np.testing.assert_allclose(self.roi_probs_expected, np.array(roi_probs), rtol=1e-5)
 #         np.testing.assert_allclose(self.rois_num_expected, np.array(rois_num), rtol=1e-5)
 
+
+class TestGenerateProposalsV2Op_ZeroSize(OpTest):
+    def set_data(self):
+        self.init_test_params()
+        self.init_test_input()
+        self.init_test_output()
+        self.inputs = {
+            'Scores': self.scores,
+            'BboxDeltas': self.bbox_deltas,
+            'ImShape': self.im_shape.astype(np.float32),
+            'Anchors': self.anchors,
+            'Variances': self.variances,
+        }
+
+        self.attrs = {
+            'pre_nms_topN': self.pre_nms_topN,
+            'post_nms_topN': self.post_nms_topN,
+            'nms_thresh': self.nms_thresh,
+            'min_size': self.min_size,
+            'eta': self.eta,
+            'pixel_offset': self.pixel_offset,
+        }
+
+        self.outputs = {
+            'RpnRois': self.rpn_rois[0],
+            'RpnRoiProbs': self.rpn_roi_probs[0],
+        }
+
+    def test_check_output(self):
+        self.check_output(check_pir=True, check_symbol_infer=False)
+
+    def setUp(self):
+        self.op_type = "generate_proposals_v2"
+        self.python_api = python_generate_proposals_v2
+        self.set_data()
+
+    def init_test_params(self):
+        self.pre_nms_topN = 12000  # train 12000, test 2000
+        self.post_nms_topN = 5000  # train 6000, test 1000
+        self.nms_thresh = 0.7
+        self.min_size = 3.0
+        self.eta = 1.0
+        self.pixel_offset = True
+
+    def init_test_input(self):
+        batch_size = 1
+        input_channels = 0
+        layer_h = 16
+        layer_w = 16
+        input_feat = np.random.random(
+            (batch_size, input_channels, layer_h, layer_w)
+        ).astype('float32')
+        self.anchors, self.variances = anchor_generator_in_python(
+            input_feat=input_feat,
+            anchor_sizes=[16.0, 32.0],
+            aspect_ratios=[0.5, 1.0],
+            variances=[1.0, 1.0, 1.0, 1.0],
+            stride=[16.0, 16.0],
+            offset=0.5,
+        )
+        self.im_shape = np.array([[64, 64]]).astype('float32')
+        num_anchors = self.anchors.shape[2]
+        self.scores = np.random.random(
+            (batch_size, num_anchors, layer_h, layer_w)
+        ).astype('float32')
+        self.bbox_deltas = np.random.random(
+            (batch_size, num_anchors * 4, layer_h, layer_w)
+        ).astype('float32')
+
+    def init_test_output(self):
+        (
+            self.rpn_rois,
+            self.rpn_roi_probs,
+            self.rois_num,
+        ) = generate_proposals_v2_in_python(
+            self.scores,
+            self.bbox_deltas,
+            self.im_shape,
+            self.anchors,
+            self.variances,
+            self.pre_nms_topN,
+            self.post_nms_topN,
+            self.nms_thresh,
+            self.min_size,
+            self.eta,
+            self.pixel_offset,
+        )
+
+
 if __name__ == '__main__':
     paddle.enable_static()
     unittest.main()