[XPU] add pos_weight for sigmoid_cross_entropy_with_logits. (#55001)

houj04 · web-flow · commit 4315bc4c8051 · 2023-08-09T11:41:47.000+08:00
* [XPU] add pos_weight for sigmoid_cross_entropy_with_logits.

* update xdnn version.
diff --git a/paddle/phi/kernels/xpu/sigmoid_cross_entropy_with_logits_grad_kernel.cc b/paddle/phi/kernels/xpu/sigmoid_cross_entropy_with_logits_grad_kernel.cc
@@ -46,7 +46,12 @@ void SigmoidCrossEntropyWithLogitsGradKernel(
   int* hit = RAII_GUARD.alloc_l3_or_gm<int>(x.numel());
   PADDLE_ENFORCE_NOT_NULL(
       hit, errors::External("XPU alloc_l3_or_gm returns nullptr"));
-
+  auto pos_weight_data =
+      (pos_weight.get_ptr() == nullptr ? nullptr
+                                       : pos_weight.get_ptr()->data<T>());
+  // int sigmoid_cross_entropy_with_logits_grad(Context* ctx, const T* x, const
+  // T* label, const T* dy, T* dx, int64_t m, int64_t n, TH* hit = nullptr,
+  // int64_t ignore_index = -100, const T* pos_weight = nullptr);
   int r = xpu::sigmoid_cross_entropy_with_logits_grad(
       dev_ctx.x_context(),
       reinterpret_cast<const XPUType*>(x.data<T>()),
@@ -56,7 +61,8 @@ void SigmoidCrossEntropyWithLogitsGradKernel(
       1,
       x.numel(),
       hit,
-      ignore_index);
+      ignore_index,
+      reinterpret_cast<const XPUType*>(pos_weight_data));
   PADDLE_ENFORCE_XDNN_SUCCESS(r, "sigmoid_cross_entropy_with_logits");
   if (normalize) {
     int* non_zero = RAII_GUARD.alloc_l3_or_gm<int>(1);
diff --git a/paddle/phi/kernels/xpu/sigmoid_cross_entropy_with_logits_kernel.cc b/paddle/phi/kernels/xpu/sigmoid_cross_entropy_with_logits_kernel.cc
@@ -43,7 +43,12 @@ void SigmoidCrossEntropyWithLogitsKernel(
   int* hit = RAII_GUARD.alloc_l3_or_gm<int>(x.numel());
   PADDLE_ENFORCE_NOT_NULL(
       hit, errors::External("XPU alloc_l3_or_gm returns nullptr"));
-
+  auto pos_weight_data =
+      (pos_weight.get_ptr() == nullptr ? nullptr
+                                       : pos_weight.get_ptr()->data<T>());
+  // int sigmoid_cross_entropy_with_logits(Context* ctx, const T* x, const T*
+  // label, T* y, int64_t m, int64_t n, TH* hit = nullptr, int64_t ignore_index
+  // = -100, const T* pos_weight = nullptr);
   int r = xpu::sigmoid_cross_entropy_with_logits(
       dev_ctx.x_context(),
       reinterpret_cast<const XPUType*>(x.data<T>()),
@@ -52,7 +57,8 @@ void SigmoidCrossEntropyWithLogitsKernel(
       1,
       x.numel(),
       hit,
-      ignore_index);
+      ignore_index,
+      reinterpret_cast<const XPUType*>(pos_weight_data));
   PADDLE_ENFORCE_XDNN_SUCCESS(r, "sigmoid_cross_entropy_with_logits");
   if (normalize) {
     int* non_zero = RAII_GUARD.alloc_l3_or_gm<int>(1);
diff --git a/test/xpu/test_sigmoid_cross_entropy_with_logits_op_xpu.py b/test/xpu/test_sigmoid_cross_entropy_with_logits_op_xpu.py
@@ -240,6 +240,41 @@ def set_output(self):
             term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
             self.outputs = {'Out': -term1 - term2}
 
+    class TestSigmoidCrossEntropyWithLogitsOp7(
+        TestSigmoidCrossEntropyWithLogitsOp
+    ):
+        """Test sigmoid_cross_entropy_with_logit_op with binary label"""
+
+        def set_inputs(self):
+            batch_size = [10, 10]
+            num_classes = 20
+            self.inputs = {
+                'X': logit(
+                    np.random.uniform(
+                        0, 1, tuple(batch_size + [num_classes])
+                    ).astype(self.dtype)
+                ),
+                'Label': np.random.randint(
+                    0, 2, tuple(batch_size + [num_classes])
+                ).astype(self.dtype),
+                'pos_weight': np.random.uniform(
+                    0, 1, tuple(batch_size + [num_classes])
+                ).astype(self.dtype),
+            }
+            self.attrs = {'num_classes': num_classes, 'batch_size': batch_size}
+
+        def set_output(self):
+            # Fw Pass is implemented as elementwise sigmoid followed by
+            # elementwise logistic loss
+            # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
+            term1 = np.maximum(self.inputs['X'], 0)
+            term2 = self.inputs['X'] * self.inputs['Label']
+            term3 = (
+                np.log(1 + np.exp(-1 * np.abs(self.inputs['X'])))
+                * self.inputs['pos_weight']
+            )
+            self.outputs = {'Out': term1 - term2 + term3}
+
     class TestSigmoidCrossEntropyWithLogitsNorm(
         TestSigmoidCrossEntropyWithLogitsOp
     ):
