[Accuracy diff No.34-35、63-64] Fix accuracy diff for logit API (#72973)

ooooo-create · web-flow · commit f6316d4f7243 · 2025-06-03T11:18:47.000+08:00
* logit

* fix

* fix
diff --git a/paddle/phi/kernels/funcs/activation_functor.h b/paddle/phi/kernels/funcs/activation_functor.h
@@ -1213,10 +1213,17 @@ struct LogitGradFunctor {
   template <typename Device, typename X, typename dOut, typename dX, typename P>
   void operator()(Device d, X x, dOut dout, dX dx, P p, float eps) const {
     // logit(x)' = 1/(x*(1-x))
-    dx.device(d) =
-        (x < static_cast<T>(eps) || x > static_cast<T>(1.0 - eps))
-            .select(p.constant(static_cast<T>(0)),
-                    dout * (static_cast<T>(1) / ((static_cast<T>(1) - x) * x)));
+    if (!eps) {
+      dx.device(d) = (x < static_cast<T>(0.0) || x > static_cast<T>(1.0))
+                         .select(p.constant(static_cast<T>(NAN)),
+                                 dout * (static_cast<T>(1) /
+                                         ((static_cast<T>(1) - x) * x)));
+    } else {
+      dx.device(d) = (x < static_cast<T>(eps) || x > static_cast<T>(1.0 - eps))
+                         .select(p.constant(static_cast<T>(0)),
+                                 dout * (static_cast<T>(1) /
+                                         ((static_cast<T>(1) - x) * x)));
+    }
   }
 };
 
@@ -3359,9 +3366,15 @@ struct CudaLogitGradFunctor : public BaseActivationFunctor<T> {
   // logit(x)' = 1/(x*(1-x))
   __device__ __forceinline__ T operator()(const T dout, const T arg_x) const {
     MT x = static_cast<MT>(arg_x);
-    MT dx = (x < static_cast<MT>(eps) || x > one - static_cast<MT>(eps))
-                ? zero
-                : (static_cast<MT>(dout) / (x * (one - x)));
+    MT dx;
+    if (!eps) {
+      dx = (x < zero || x > one) ? static_cast<T>(NAN)
+                                 : (static_cast<MT>(dout) / (x * (one - x)));
+    } else {
+      dx = (x < static_cast<MT>(eps) || x > one - static_cast<MT>(eps))
+               ? zero
+               : (static_cast<MT>(dout) / (x * (one - x)));
+    }
     return static_cast<T>(dx);
   }
   static constexpr ActBwdOpFwdDeps FwdDeps() {
diff --git a/test/legacy_test/test_logit_op.py b/test/legacy_test/test_logit_op.py
@@ -24,17 +24,38 @@
 
 
 def logit(x, eps):
-    x_min = np.minimum(x, 1.0 - eps)
-    x_max = np.maximum(x_min, eps)
-    return np.log(x_max / (1.0 - x_max))
+    if eps:
+        x_min = np.minimum(x, 1.0 - eps)
+        x_max = np.maximum(x_min, eps)
+        return np.log(x_max / (1.0 - x_max))
+    else:
+        return np.where(
+            (x < 0.0) | (x > 1.0),
+            np.array(np.nan, dtype=x.dtype),
+            np.log(x / (1.0 - x)),
+        )
 
 
 def logit_grad(x, eps=1e-8):
-    tmp_x = np.select(
-        [x < eps, x > (1.0 - eps)], [x * 0.0, x * 0.0], default=-1.0
-    )
-    x_1 = 1.0 - x
-    _x = np.select([tmp_x == -1.0], [np.reciprocal(x * x_1)], default=0.0)
+    if eps:
+        tmp_x = np.select(
+            [x < eps, x > (1.0 - eps)], [x * 0.0, x * 0.0], default=-1.0
+        )
+        x_1 = 1.0 - x
+        _x = np.select([tmp_x == -1.0], [np.reciprocal(x * x_1)], default=0.0)
+    else:
+        tmp_x = np.select(
+            [x < 0.0, x > 1.0],
+            [np.array(np.nan, dtype=x.dtype), np.array(np.nan, dtype=x.dtype)],
+            default=-1.0,
+        )
+        x_1 = 1.0 - x
+        _x = np.select(
+            [tmp_x == -1.0],
+            [np.reciprocal(x * x_1)],
+            default=np.array(np.nan, dtype=x.dtype),
+        )
+
     if _x.size == 0:
         dout = np.full_like(x, fill_value=0.0)
     else:
@@ -162,9 +183,13 @@ def set_attrs(self):
 
 
 class TestLogitAPI(unittest.TestCase):
-    def setUp(self):
+    def init_data(self):
         self.x_shape = [120]
-        self.x = np.random.uniform(0.0, 1.0, self.x_shape).astype(np.float32)
+        self.x_dtype = "float32"
+
+    def setUp(self):
+        self.init_data()
+        self.x = np.random.uniform(-1.0, 1.0, self.x_shape).astype(self.x_dtype)
         self.place = (
             paddle.CUDAPlace(0)
             if paddle.base.core.is_compiled_with_cuda()
@@ -175,22 +200,38 @@ def check_api(self, eps=1e-8):
         ref_out = logit(self.x, eps)
         # test static api
         with paddle.static.program_guard(paddle.static.Program()):
-            x = paddle.static.data(name='x', shape=self.x_shape)
+            x = paddle.static.data(
+                name='x', shape=self.x_shape, dtype=self.x_dtype
+            )
             y = paddle.logit(x, eps)
             exe = paddle.static.Executor(self.place)
             out = exe.run(feed={'x': self.x}, fetch_list=[y])
         np.testing.assert_allclose(out[0], ref_out, rtol=1e-05)
         # test dygrapg api
         paddle.disable_static()
-        x = paddle.to_tensor(self.x)
-        y = paddle.logit(x, 1e-8)
+        x = paddle.to_tensor(self.x, dtype=self.x_dtype)
+        y = paddle.logit(x, eps)
         np.testing.assert_allclose(y.numpy(), ref_out, rtol=1e-05)
         paddle.enable_static()
 
+    def check_api_grad(self, eps=1e-8):
+        ref_grad = logit_grad(self.x, eps)
+        numpy_tensor = np.ones(self.x_shape).astype(self.x_dtype)
+        # test dygrapg api
+        paddle.disable_static()
+        paddle_outgrad = paddle.to_tensor(numpy_tensor / numpy_tensor.size)
+        x = paddle.to_tensor(self.x, dtype=self.x_dtype)
+        x.stop_gradient = False
+        y = paddle.logit(x, eps)
+        x_grad = paddle.grad([y], [x], [paddle_outgrad])
+        np.testing.assert_allclose(x_grad[0].numpy(), ref_grad, rtol=1e-05)
+        paddle.enable_static()
+
     def test_check_api(self):
         paddle.enable_static()
         for eps in [1e-6, 0.0]:
             self.check_api(eps)
+            self.check_api_grad(eps)
 
     def test_errors(self):
         paddle.enable_static()
@@ -202,5 +243,17 @@ def test_errors(self):
             self.assertRaises(TypeError, paddle.logit, x, dtype='int32')
 
 
+class TestLogitAPICase1(unittest.TestCase):
+    def init_data(self):
+        self.x_shape = [120]
+        self.x_dtype = "float64"
+
+
+class TestLogitAPICase2(unittest.TestCase):
+    def init_data(self):
+        self.x_shape = [120]
+        self.x_dtype = "float16"
+
+
 if __name__ == "__main__":
     unittest.main()
