[phi] add yolov3_loss yaml and unittest (#44476)

* add yaml and unittest

* update yaml

* update backward yaml and unittest

* update yaml

* add Yolov3LossGradInferMeta

* update yolov3_loss_op.cc

* fix bug

* code format

paddle/fluid/operators/detection/yolov3_loss_op.cc

@@ -15,6 +15,7 @@
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/imperative/type_defs.h"
 #include "paddle/phi/core/infermeta_utils.h"
+#include "paddle/phi/infermeta/backward.h"
 #include "paddle/phi/infermeta/multiary.h"
 
 namespace paddle {
@@ -178,20 +179,6 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker {
 class Yolov3LossOpGrad : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
-  void InferShape(framework::InferShapeContext* ctx) const override {
-    PADDLE_ENFORCE_EQ(
-        ctx->HasInput("X"),
-        true,
-        platform::errors::NotFound("Input(X) should not be null"));
-    PADDLE_ENFORCE_EQ(
-        ctx->HasInput(framework::GradVarName("Loss")),
-        true,
-        platform::errors::NotFound("Input(Loss@GRAD) should not be null"));
-    auto dim_x = ctx->GetInputDim("X");
-    if (ctx->HasOutput(framework::GradVarName("X"))) {
-      ctx->SetOutputDim(framework::GradVarName("X"), dim_x);
-    }
-  }
 
  protected:
   framework::OpKernelType GetExpectedKernelType(
@@ -234,10 +221,15 @@ namespace ops = paddle::operators;
 DECLARE_INFER_SHAPE_FUNCTOR(yolov3_loss,
                             Yolov3LossInferShapeFunctor,
                             PD_INFER_META(phi::Yolov3LossInferMeta));
+DECLARE_INFER_SHAPE_FUNCTOR(yolov3_loss_grad,
+                            Yolov3LossGradInferShapeFunctor,
+                            PD_INFER_META(phi::Yolov3LossGradInferMeta));
 REGISTER_OPERATOR(yolov3_loss,
                   ops::Yolov3LossOp,
                   ops::Yolov3LossOpMaker,
                   ops::Yolov3LossGradMaker<paddle::framework::OpDesc>,
                   ops::Yolov3LossGradMaker<paddle::imperative::OpBase>,
                   Yolov3LossInferShapeFunctor);
-REGISTER_OPERATOR(yolov3_loss_grad, ops::Yolov3LossOpGrad);
+REGISTER_OPERATOR(yolov3_loss_grad,
+                  ops::Yolov3LossOpGrad,
+                  Yolov3LossGradInferShapeFunctor);

paddle/phi/api/yaml/legacy_api.yaml

@@ -2626,6 +2626,18 @@
     func : yolo_box
     data_type : x
 
+# yolov3_loss
+- api : yolov3_loss
+  args : (Tensor x, Tensor gt_box, Tensor gt_label, Tensor gt_score, int[] anchors, int[] anchor_mask, int class_num, float ignore_thresh, int downsample_ratio, bool use_label_smooth=true, float scale_x_y=1.0)
+  output : Tensor(loss), Tensor(objectness_mask), Tensor(gt_match_mask)
+  infer_meta :
+    func : Yolov3LossInferMeta
+  kernel :
+    func : yolov3_loss
+    data_type : x
+  optional : gt_score
+  backward : yolov3_loss_grad
+
 - api : zeros_like
   args : (Tensor x, DataType dtype=DataType::UNDEFINED, Place place = {})
   output : Tensor

paddle/phi/api/yaml/legacy_backward.yaml

@@ -2531,3 +2531,13 @@
   kernel :
     func : where_grad
   no_need_buffer : x, y
+
+- backward_api : yolov3_loss_grad
+  forward : yolov3_loss(Tensor x, Tensor gt_box, Tensor gt_label, Tensor gt_score, int[] anchors, int[] anchor_mask, int class_num, float ignore_thresh, int downsample_ratio, bool use_label_smooth=true, float scale_x_y=1.0) -> Tensor(loss), Tensor(objectness_mask), Tensor(gt_match_mask)
+  args : (Tensor x, Tensor gt_box, Tensor gt_label, Tensor gt_score, Tensor objectness_mask, Tensor gt_match_mask, Tensor loss_grad, int[] anchors, int[] anchor_mask, int class_num, float ignore_thresh, int downsample_ratio, bool use_label_smooth=true, float scale_x_y=1.0)
+  output : Tensor(x_grad), Tensor(gt_box_grad), Tensor(gt_label_grad), Tensor(gt_score_grad)
+  infer_meta :
+    func : Yolov3LossGradInferMeta
+  kernel :
+    func : yolov3_loss_grad
+  optional : gt_score

paddle/phi/infermeta/backward.cc

@@ -828,4 +828,28 @@ void UnStackGradInferMeta(const std::vector<const MetaTensor*>& out_grad,
   x_grad->set_dtype(out_grad[0]->dtype());
 }
 
+void Yolov3LossGradInferMeta(const MetaTensor& x,
+                             const MetaTensor& gt_box,
+                             const MetaTensor& gt_label,
+                             const MetaTensor& gt_score,
+                             const MetaTensor& objectness_mask,
+                             const MetaTensor& gt_match_mask,
+                             const MetaTensor& loss_grad,
+                             const std::vector<int>& anchors,
+                             const std::vector<int>& anchor_mask,
+                             int class_num,
+                             float ignore_thresh,
+                             int downsample_ratio,
+                             bool use_label_smooth,
+                             float scale_x_y,
+                             MetaTensor* x_grad,
+                             MetaTensor* gt_box_grad,
+                             MetaTensor* gt_label_grad,
+                             MetaTensor* gt_score_grad) {
+  if (x_grad) {
+    x_grad->set_dims(x.dims());
+    x_grad->set_dtype(x.dtype());
+  }
+}
+
 }  // namespace phi

paddle/phi/infermeta/backward.h

@@ -328,4 +328,23 @@ void UnStackGradInferMeta(const std::vector<const MetaTensor*>& out_grad,
                           int axis,
                           MetaTensor* x_grad);
 
+void Yolov3LossGradInferMeta(const MetaTensor& x,
+                             const MetaTensor& gt_box,
+                             const MetaTensor& gt_label,
+                             const MetaTensor& gt_score,
+                             const MetaTensor& objectness_mask,
+                             const MetaTensor& gt_match_mask,
+                             const MetaTensor& loss_grad,
+                             const std::vector<int>& anchors,
+                             const std::vector<int>& anchor_mask,
+                             int class_num,
+                             float ignore_thresh,
+                             int downsample_ratio,
+                             bool use_label_smooth,
+                             float scale_x_y,
+                             MetaTensor* x_grad,
+                             MetaTensor* gt_box_grad,
+                             MetaTensor* gt_label_grad,
+                             MetaTensor* gt_score_grad);
+
 }  // namespace phi

paddle/phi/kernels/cpu/yolov3_loss_grad_kernel.cc

@@ -122,9 +122,9 @@ void Yolov3LossGradKernel(const Context& dev_ctx,
                           const DenseTensor& gt_box,
                           const DenseTensor& gt_label,
                           const paddle::optional<DenseTensor>& gt_score,
-                          const DenseTensor& loss_grad,
                           const DenseTensor& objectness_mask,
                           const DenseTensor& gt_match_mask,
+                          const DenseTensor& loss_grad,
                           const std::vector<int>& anchors,
                           const std::vector<int>& anchor_mask,
                           int class_num,

paddle/phi/kernels/yolov3_loss_grad_kernel.h

@@ -24,9 +24,9 @@ void Yolov3LossGradKernel(const Context& dev_ctx,
                           const DenseTensor& gt_box,
                           const DenseTensor& gt_label,
                           const paddle::optional<DenseTensor>& gt_score,
-                          const DenseTensor& loss_grad,
                           const DenseTensor& objectness_mask,
                           const DenseTensor& gt_match_mask,
+                          const DenseTensor& loss_grad,
                           const std::vector<int>& anchors,
                           const std::vector<int>& anchor_mask,
                           int class_num,

paddle/phi/ops/compat/yolov3_loss_sig.cc

@@ -37,9 +37,9 @@ KernelSignature Yolov3LossGradOpArgumentMapping(
        "GTBox",
        "GTLabel",
        "GTScore",
-       "Loss@GRAD",
        "ObjectnessMask",
-       "GTMatchMask"},
+       "GTMatchMask",
+       "Loss@GRAD"},
       {"anchors",
        "anchor_mask",
        "class_num",

python/paddle/fluid/tests/unittests/test_yolov3_loss_op.py

@@ -169,11 +169,41 @@ def YOLOv3Loss(x, gtbox, gtlabel, gtscore, attrs):
             gt_matches.astype('int32'))
 
 
+def yolo_loss_wrapper(x,
+                      gt_box,
+                      gt_label,
+                      gt_score=None,
+                      anchors=[
+                          10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116,
+                          90, 156, 198, 373, 326
+                      ],
+                      anchor_mask=[0, 1, 2],
+                      class_num=5,
+                      ignore_thresh=0.7,
+                      downsample_ratio=32,
+                      use_label_smooth=True,
+                      scale_x_y=1.):
+    loss = paddle.vision.ops.yolo_loss(x,
+                                       gt_box=gt_box,
+                                       gt_label=gt_label,
+                                       anchors=anchors,
+                                       anchor_mask=anchor_mask,
+                                       class_num=class_num,
+                                       ignore_thresh=ignore_thresh,
+                                       downsample_ratio=downsample_ratio,
+                                       gt_score=gt_score,
+                                       use_label_smooth=use_label_smooth,
+                                       scale_x_y=scale_x_y)
+    return loss
+
+
 class TestYolov3LossOp(OpTest):
 
     def setUp(self):
         self.initTestCase()
         self.op_type = 'yolov3_loss'
+        self.python_api = yolo_loss_wrapper
+        self.python_out_sig = ['Loss']
         x = logit(np.random.uniform(0, 1, self.x_shape).astype('float64'))
         gtbox = np.random.random(size=self.gtbox_shape).astype('float64')
         gtlabel = np.random.randint(0, self.class_num, self.gtbox_shape[:2])
@@ -212,11 +242,14 @@ class TestYolov3LossOp(OpTest):
 
     def test_check_output(self):
         place = core.CPUPlace()
-        self.check_output_with_place(place, atol=2e-3)
+        self.check_output_with_place(place, atol=2e-3, check_eager=True)
 
     def test_check_grad_ignore_gtbox(self):
         place = core.CPUPlace()
-        self.check_grad_with_place(place, ['X'], 'Loss', max_relative_error=0.2)
+        self.check_grad_with_place(place, ['X'],
+                                   'Loss',
+                                   max_relative_error=0.2,
+                                   check_eager=True)
 
     def initTestCase(self):
         self.anchors = [

python/paddle/vision/ops.py

@@ -186,6 +186,12 @@ def yolo_loss(x,
                                              scale_x_y=1.)
     """
 
+    if in_dygraph_mode():
+        loss, _, _ = _C_ops.final_state_yolov3_loss(
+            x, gt_box, gt_label, gt_score, anchors, anchor_mask, class_num,
+            ignore_thresh, downsample_ratio, use_label_smooth, scale_x_y)
+        return loss
+
     if _non_static_mode():
         loss, _, _ = _C_ops.yolov3_loss(
             x, gt_box, gt_label, gt_score, 'anchors', anchors, 'anchor_mask',