[0-size Tensor Job2 No.61] Add 0-size Tensor support for log_loss (#73401)

co63oc · web-flow · commit 114e1d3562d8 · 2025-06-20T16:18:53.000+08:00
* Fix

* Fix

* Fix

* Fix
diff --git a/paddle/phi/kernels/impl/log_loss_grad_kernel_impl.h b/paddle/phi/kernels/impl/log_loss_grad_kernel_impl.h
@@ -26,6 +26,10 @@ void LogLossGradKernel(const Context& dev_ctx,
                        const DenseTensor& out_grad,
                        float epsilon,
                        DenseTensor* in_grad) {
+  if (in_grad && in_grad->numel() == 0) {
+    dev_ctx.template Alloc<T>(in_grad);
+    return;
+  }
   auto prediction = EigenVector<T>::Flatten(input);
   auto label_out = EigenVector<T>::Flatten(label);
 
diff --git a/paddle/phi/kernels/impl/log_loss_kernel_impl.h b/paddle/phi/kernels/impl/log_loss_kernel_impl.h
@@ -26,6 +26,7 @@ void LogLossKernel(const Context& dev_ctx,
                    float epsilon,
                    DenseTensor* out) {
   dev_ctx.template Alloc<T>(out);
+  if (out && out->numel() == 0) return;
 
   auto prediction = EigenVector<T>::Flatten(input);
   auto label_out = EigenVector<T>::Flatten(label);
diff --git a/paddle/phi/kernels/xpu/log_loss_grad_kernel.cc b/paddle/phi/kernels/xpu/log_loss_grad_kernel.cc
@@ -25,6 +25,10 @@ void LogLossGradXPUKernel(const Context& dev_ctx,
                           const DenseTensor& out_grad,
                           float epsilon_in,
                           DenseTensor* in_grad) {
+  if (in_grad && in_grad->numel() == 0) {
+    dev_ctx.template Alloc<T>(in_grad);
+    return;
+  }
   auto* predict = &input;
   auto* labels = &label;
   auto* dloss = &out_grad;
diff --git a/paddle/phi/kernels/xpu/log_loss_kernel.cc b/paddle/phi/kernels/xpu/log_loss_kernel.cc
@@ -29,6 +29,8 @@ void LogLossXPUKernel(const Context& dev_ctx,
   auto* loss = out;
   auto epsilon = static_cast<T>(epsilon_in);
   dev_ctx.template Alloc<T>(loss);
+  if (out && out->numel() == 0) return;
+
   int64_t n = predict->numel();
   int r = xpu::log_loss(dev_ctx.x_context(),
                         predict->data<T>(),
diff --git a/test/legacy_test/test_log_loss_op.py b/test/legacy_test/test_log_loss_op.py
@@ -60,5 +60,42 @@ def test_check_grad(self):
         )
 
 
+class TestLogLossOp_ZeroSize(OpTest):
+    def init_shape(self):
+        self.x_shape = (0, 1)
+        self.label_shape = (0, 1)
+
+    def setUp(self):
+        self.init_shape()
+        self.op_type = 'log_loss'
+        self.python_api = F.log_loss
+        self.public_python_api = F.log_loss
+
+        x = np.random.random(self.x_shape).astype("float32")
+        predicted = sigmoid_array(x)
+        labels = np.random.randint(0, 2, self.label_shape).astype("float32")
+        epsilon = 1e-7
+        self.inputs = {
+            'Predicted': predicted,
+            'Labels': labels,
+        }
+
+        self.attrs = {'epsilon': epsilon}
+        loss = -labels * np.log(predicted + epsilon) - (1 - labels) * np.log(
+            1 - predicted + epsilon
+        )
+        self.outputs = {'Loss': loss}
+
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
+    def test_check_grad(self):
+        self.check_grad(
+            ['Predicted'],
+            'Loss',
+            check_pir=True,
+        )
+
+
 if __name__ == '__main__':
     unittest.main()
