[autodiff] Clear all dual fields when exiting context manager

python/taichi/ad/_ad.py

@@ -283,6 +283,11 @@ def shape_flatten(shape):
         else:
             assert parameters_shape_flatten == len(self.seed)
 
+        # Clear gradients
+        if self.clear_gradients:
+            # TODO: the clear gradients should be controlled to clear adjoint/dual/adjoint_visited respectively
+            clear_all_gradients()
+
         # Set seed for each variable
         if len(self.seed) == 1:
             if len(self.param.shape) == 0:
@@ -294,11 +299,6 @@ def shape_flatten(shape):
         else:
             self.param.dual.from_numpy(np.array(self.seed, dtype=np.float32))
 
-        # Clear gradients
-        if self.clear_gradients:
-            for ls in self.loss:
-                ls.dual.fill(0)
-
         # Attach the context manager to the runtime
         self.runtime.fwd_mode_manager = self
 

tests/python/test_ad_basics_fwd.py

@@ -103,3 +103,23 @@ def func():
 
     with ti.ad.FwdMode(loss=d, param=c):
         func()
+
+
+@test_utils.test()
+def test_clear_all_dual_field():
+    x = ti.field(float, shape=(), needs_dual=True)
+    y = ti.field(float, shape=(), needs_dual=True)
+    loss = ti.field(float, shape=(), needs_dual=True)
+
+    x[None] = 2.0
+    y[None] = 3.0
+
+    @ti.kernel
+    def clear_dual_test():
+        y[None] = x[None]**2
+        loss[None] += y[None]
+
+    for _ in range(5):
+        with ti.ad.FwdMode(loss=loss, param=x):
+            clear_dual_test()
+        assert y.dual[None] == 4.0