[cherry-pick] Fix proposals CPU zero-size bugs

paddle/phi/kernels/cpu/generate_proposals_kernel.cc

@@ -328,6 +328,17 @@ void GenerateProposalsKernel(const Context& dev_ctx,
     return;
   }
 
+  if (bbox_deltas.numel() == 0 || im_shape.numel() == 0) {
+    rpn_rois->Resize(make_ddim({0, 4}));
+    rpn_roi_probs->Resize(make_ddim({0, 1}));
+    if (rpn_rois_num != nullptr) {
+      rpn_rois_num->Resize(make_ddim({num}));
+      int64_t* num_data = dev_ctx.template Alloc<int64_t>(rpn_rois_num);
+      std::fill_n(num_data, num, 0);
+    }
+    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);

test/legacy_test/test_generate_proposals_v2_zero_size_input.py

@@ -0,0 +1,366 @@
+#   Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Unit tests for generate_proposals with zero-size inputs.
+
+Covers two crash scenarios discovered in a.config:
+  1. im_shape has zero second dimension (shape [N, 0], numel=0):
+       -> ClipTiledBoxes / FilterBoxes would dereference empty data pointer -> SIGSEGV
+  2. bbox_deltas has zero batch dimension (shape [0, C, H, W], numel=0)
+       while scores batch > 0:
+       -> rpn_rois pre-allocated as [0,4], AppendProposals writes to nullptr -> SIGSEGV
+
+Fix: GenerateProposalsKernel now returns empty tensors immediately when
+     bbox_deltas.numel() == 0 or im_shape.numel() == 0.
+"""
+
+import unittest
+
+import numpy as np
+
+import paddle
+
+# Force CPU execution: these tests target the CPU kernel fix.
+paddle.device.set_device('cpu')
+
+
+def _call_generate_proposals(
+    scores_shape,
+    bbox_deltas_shape,
+    im_shape_val,
+    anchors_shape,
+    variances_shape,
+    pre_nms_top_n,
+    post_nms_top_n,
+    nms_thresh=0.7,
+    min_size=0.0,
+    eta=1.0,
+    pixel_offset=False,
+    return_rois_num=True,
+):
+    """Helper: construct tensors on CPU and call generate_proposals."""
+    cpu = paddle.CPUPlace()
+    np.random.seed(0)
+    scores = paddle.to_tensor(
+        np.random.rand(*scores_shape).astype('float32'), place=cpu
+    )
+    bbox_deltas = paddle.to_tensor(
+        np.random.rand(*bbox_deltas_shape).astype('float32')
+        if 0 not in bbox_deltas_shape
+        else np.empty(bbox_deltas_shape, dtype='float32'),
+        place=cpu,
+    )
+    im_shape = paddle.to_tensor(
+        np.array(im_shape_val, dtype='float32'), place=cpu
+    )
+    anchors = paddle.to_tensor(
+        np.random.rand(*anchors_shape).astype('float32'), place=cpu
+    )
+    variances = paddle.to_tensor(
+        np.ones(variances_shape, dtype='float32'), place=cpu
+    )
+
+    return paddle.vision.ops.generate_proposals(
+        scores,
+        bbox_deltas,
+        im_shape,
+        anchors,
+        variances,
+        pre_nms_top_n=pre_nms_top_n,
+        post_nms_top_n=post_nms_top_n,
+        nms_thresh=nms_thresh,
+        min_size=min_size,
+        eta=eta,
+        pixel_offset=pixel_offset,
+        return_rois_num=return_rois_num,
+    )
+
+
+class TestGenerateProposalsZeroImShape(unittest.TestCase):
+    """
+    im_shape second dimension = 0  (e.g. shape [1, 0] or [2, 0]).
+
+    Corresponds to a.config lines with Tensor([1, 0],"float32") or
+    Tensor([2, 0],"float32") as im_shape argument.
+    Expected: no crash; rpn_rois shape [0,4], rpn_roi_probs shape [0,1].
+    """
+
+    def setUp(self):
+        paddle.disable_static()
+
+    def _check_zero_output(self, result, return_rois_num):
+        if return_rois_num:
+            rpn_rois, rpn_roi_probs, _ = result
+        else:
+            rpn_rois, rpn_roi_probs = result
+        self.assertEqual(
+            list(rpn_rois.shape),
+            [0, 4],
+            f"rpn_rois shape {rpn_rois.shape} != [0, 4]",
+        )
+        self.assertEqual(
+            list(rpn_roi_probs.shape),
+            [0, 1],
+            f"rpn_roi_probs shape {rpn_roi_probs.shape} != [0, 1]",
+        )
+
+    def test_im_shape_zero_dim1_batch1_c15_h40_w60(self):
+        """a.config line 2: scores[1,15,40,60], bbox_deltas[1,60,40,60], im_shape[1,0]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 15, 40, 60],
+            bbox_deltas_shape=[1, 60, 40, 60],
+            im_shape_val=np.empty((1, 0), dtype='float32'),
+            anchors_shape=[36000, 4],
+            variances_shape=[36000, 4],
+            pre_nms_top_n=12000,
+            post_nms_top_n=2000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_im_shape_zero_dim1_batch1_c15_h42_w63(self):
+        """a.config line 4: scores[1,15,42,63], bbox_deltas[1,60,42,63], im_shape[1,0]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 15, 42, 63],
+            bbox_deltas_shape=[1, 60, 42, 63],
+            im_shape_val=np.empty((1, 0), dtype='float32'),
+            anchors_shape=[39690, 4],
+            variances_shape=[39690, 4],
+            pre_nms_top_n=12000,
+            post_nms_top_n=2000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_im_shape_zero_dim1_batch1_c3_h10_w14(self):
+        """a.config line 6: scores[1,3,10,14], bbox_deltas[1,12,10,14], im_shape[1,0]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 3, 10, 14],
+            bbox_deltas_shape=[1, 12, 10, 14],
+            im_shape_val=np.empty((1, 0), dtype='float32'),
+            anchors_shape=[420, 4],
+            variances_shape=[420, 4],
+            pre_nms_top_n=2000,
+            post_nms_top_n=2000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_im_shape_zero_dim1_batch1_c3_h10_w15(self):
+        """a.config line 8: scores[1,3,10,15], bbox_deltas[1,12,10,15], im_shape[1,0]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 3, 10, 15],
+            bbox_deltas_shape=[1, 12, 10, 15],
+            im_shape_val=np.empty((1, 0), dtype='float32'),
+            anchors_shape=[450, 4],
+            variances_shape=[450, 4],
+            pre_nms_top_n=2000,
+            post_nms_top_n=2000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_im_shape_zero_dim1_batch1_c4_h16_w16_pixel_offset(self):
+        """a.config line 10: scores[1,4,16,16], bbox_deltas[1,16,16,16], im_shape[1,0], pixel_offset=True"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 4, 16, 16],
+            bbox_deltas_shape=[1, 16, 16, 16],
+            im_shape_val=np.empty((1, 0), dtype='float32'),
+            anchors_shape=[16, 16, 4, 4],
+            variances_shape=[16, 16, 4, 4],
+            pre_nms_top_n=12000,
+            post_nms_top_n=5000,
+            nms_thresh=0.7,
+            min_size=3.0,
+            pixel_offset=True,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_im_shape_zero_dim1_batch1_c9_h10_w8(self):
+        """a.config line 12: scores[1,9,10,8], bbox_deltas[1,36,10,8], im_shape[1,0]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 9, 10, 8],
+            bbox_deltas_shape=[1, 36, 10, 8],
+            im_shape_val=np.empty((1, 0), dtype='float32'),
+            anchors_shape=[720, 4],
+            variances_shape=[720, 4],
+            pre_nms_top_n=4000,
+            post_nms_top_n=4000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_im_shape_zero_dim1_batch1_c9_h12_w9(self):
+        """a.config line 14: scores[1,9,12,9], bbox_deltas[1,36,12,9], im_shape[1,0]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 9, 12, 9],
+            bbox_deltas_shape=[1, 36, 12, 9],
+            im_shape_val=np.empty((1, 0), dtype='float32'),
+            anchors_shape=[972, 4],
+            variances_shape=[972, 4],
+            pre_nms_top_n=4000,
+            post_nms_top_n=4000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_im_shape_zero_dim1_batch2_c3_h4_w4(self):
+        """a.config line 16: scores[2,3,4,4], bbox_deltas[2,12,4,4], im_shape[2,0]"""
+        result = _call_generate_proposals(
+            scores_shape=[2, 3, 4, 4],
+            bbox_deltas_shape=[2, 12, 4, 4],
+            im_shape_val=np.empty((2, 0), dtype='float32'),
+            anchors_shape=[4, 4, 3, 4],
+            variances_shape=[4, 4, 3, 4],
+            pre_nms_top_n=10,
+            post_nms_top_n=5,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+
+class TestGenerateProposalsZeroBboxDeltas(unittest.TestCase):
+    """
+    bbox_deltas batch dimension = 0 (e.g. shape [0, C, H, W], numel=0)
+    while scores batch > 0.
+
+    Corresponds to a.config lines with Tensor([0, C, H, W],"float32") as
+    bbox_deltas argument (odd-numbered pairs in a.config).
+    Expected: no crash; rpn_rois shape [0,4], rpn_roi_probs shape [0,1].
+    """
+
+    def setUp(self):
+        paddle.disable_static()
+
+    def _check_zero_output(self, result, return_rois_num):
+        if return_rois_num:
+            rpn_rois, rpn_roi_probs, _ = result
+        else:
+            rpn_rois, rpn_roi_probs = result
+        self.assertEqual(
+            list(rpn_rois.shape),
+            [0, 4],
+            f"rpn_rois shape {rpn_rois.shape} != [0, 4]",
+        )
+        self.assertEqual(
+            list(rpn_roi_probs.shape),
+            [0, 1],
+            f"rpn_roi_probs shape {rpn_roi_probs.shape} != [0, 1]",
+        )
+
+    def test_bbox_deltas_zero_batch_c15_h40_w60(self):
+        """a.config line 1: scores[1,15,40,60], bbox_deltas[0,60,40,60], im_shape[1,2]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 15, 40, 60],
+            bbox_deltas_shape=[0, 60, 40, 60],
+            im_shape_val=[[60.0, 40.0]],
+            anchors_shape=[36000, 4],
+            variances_shape=[36000, 4],
+            pre_nms_top_n=12000,
+            post_nms_top_n=2000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_bbox_deltas_zero_batch_c15_h42_w63(self):
+        """a.config line 3: scores[1,15,42,63], bbox_deltas[0,60,42,63], im_shape[1,2]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 15, 42, 63],
+            bbox_deltas_shape=[0, 60, 42, 63],
+            im_shape_val=[[60.0, 42.0]],
+            anchors_shape=[39690, 4],
+            variances_shape=[39690, 4],
+            pre_nms_top_n=12000,
+            post_nms_top_n=2000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_bbox_deltas_zero_batch_c3_h10_w14(self):
+        """a.config line 5: scores[1,3,10,14], bbox_deltas[0,12,10,14], im_shape[1,2]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 3, 10, 14],
+            bbox_deltas_shape=[0, 12, 10, 14],
+            im_shape_val=[[12.0, 10.0]],
+            anchors_shape=[420, 4],
+            variances_shape=[420, 4],
+            pre_nms_top_n=2000,
+            post_nms_top_n=2000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_bbox_deltas_zero_batch_c3_h10_w15(self):
+        """a.config line 7: scores[1,3,10,15], bbox_deltas[0,12,10,15], im_shape[1,2]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 3, 10, 15],
+            bbox_deltas_shape=[0, 12, 10, 15],
+            im_shape_val=[[12.0, 10.0]],
+            anchors_shape=[450, 4],
+            variances_shape=[450, 4],
+            pre_nms_top_n=2000,
+            post_nms_top_n=2000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_bbox_deltas_zero_batch_c4_h16_w16_pixel_offset(self):
+        """a.config line 9: scores[1,4,16,16], bbox_deltas[0,16,16,16], im_shape[1,2], pixel_offset=True"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 4, 16, 16],
+            bbox_deltas_shape=[0, 16, 16, 16],
+            im_shape_val=[[16.0, 16.0]],
+            anchors_shape=[16, 16, 4, 4],
+            variances_shape=[16, 16, 4, 4],
+            pre_nms_top_n=12000,
+            post_nms_top_n=5000,
+            nms_thresh=0.7,
+            min_size=3.0,
+            pixel_offset=True,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_bbox_deltas_zero_batch_c9_h10_w8(self):
+        """a.config line 11: scores[1,9,10,8], bbox_deltas[0,36,10,8], im_shape[1,2]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 9, 10, 8],
+            bbox_deltas_shape=[0, 36, 10, 8],
+            im_shape_val=[[36.0, 10.0]],
+            anchors_shape=[720, 4],
+            variances_shape=[720, 4],
+            pre_nms_top_n=4000,
+            post_nms_top_n=4000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_bbox_deltas_zero_batch_c9_h12_w9(self):
+        """a.config line 13: scores[1,9,12,9], bbox_deltas[0,36,12,9], im_shape[1,2]"""
+        result = _call_generate_proposals(
+            scores_shape=[1, 9, 12, 9],
+            bbox_deltas_shape=[0, 36, 12, 9],
+            im_shape_val=[[36.0, 12.0]],
+            anchors_shape=[972, 4],
+            variances_shape=[972, 4],
+            pre_nms_top_n=4000,
+            post_nms_top_n=4000,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+    def test_bbox_deltas_zero_batch_multibatch_c3_h4_w4(self):
+        """a.config line 15: scores[2,3,4,4], bbox_deltas[0,12,4,4], im_shape[2,3]"""
+        result = _call_generate_proposals(
+            scores_shape=[2, 3, 4, 4],
+            bbox_deltas_shape=[0, 12, 4, 4],
+            im_shape_val=[[4.0, 4.0, 1.0], [4.0, 4.0, 1.0]],
+            anchors_shape=[4, 4, 3, 4],
+            variances_shape=[4, 4, 3, 4],
+            pre_nms_top_n=10,
+            post_nms_top_n=5,
+        )
+        self._check_zero_output(result, return_rois_num=True)
+
+
+if __name__ == '__main__':
+    paddle.disable_static()
+    unittest.main()