[Cherry-Pick] Clear 'BasicEngine' when an exception occurs in the backward. (#32546) (#32615)

* clear 'BasicEngine' when an exception occurs in the backward. (#32546)

* clear 'BasicEngine' when an exception occurs in the backward.

* deal with conflict.

* deal with conflict.

* forward return any type. (#32661)

paddle/fluid/imperative/basic_engine.cc

@@ -471,12 +471,20 @@ void BasicEngine::Execute() {
 
       {
         VLOG(3) << "Start to execute grad op " << cur_op.Type();
+        try {
           if (tmp_ins_ptr == nullptr) {
             OpBase::Run(cur_op.InnerOp(), bwd_ins, tmp_outs, cur_op.Attrs(),
                         cur_op.place());
           } else {
-          OpBase::Run(cur_op.InnerOp(), *tmp_ins_ptr, tmp_outs, cur_op.Attrs(),
-                      cur_op.place());
+            OpBase::Run(cur_op.InnerOp(), *tmp_ins_ptr, tmp_outs,
+                        cur_op.Attrs(), cur_op.place());
+          }
+        } catch (platform::EnforceNotMet& exception) {
+          Clear();
+          throw std::move(exception);
+        } catch (std::exception& ex) {
+          Clear();
+          PADDLE_THROW(platform::errors::External("%s", ex.what()));
         }
       }
 

paddle/fluid/imperative/py_layer_fwd.h

@@ -115,12 +115,12 @@ py::object PyLayerApply(const platform::Place& place, const py::object& cls,
               tuple_result[i].cast<std::shared_ptr<imperative::VarBase>>();
           output_vars.push_back(temp_out);
         } catch (py::cast_error&) {
-          PADDLE_THROW(platform::errors::Unimplemented(
-              "The output of `PyLayer.forward` should be `Tensor`."));
+          // Only collect Tensor type in 'kwargs' and pass them to backward.
+          // Ignore other types of input temporarily.
         }
       } else {
-        PADDLE_THROW(platform::errors::Unimplemented(
-            "The output of `PyLayer.forward` can not be `None`."));
+        // Only collect Tensor type in 'kwargs' and pass them to backward.
+        // Ignore other types of input temporarily.
       }
     }
   } else {
@@ -130,13 +130,17 @@ py::object PyLayerApply(const platform::Place& place, const py::object& cls,
             result_forward.cast<std::shared_ptr<imperative::VarBase>>();
         output_vars.push_back(temp_out);
       } catch (py::cast_error&) {
-        PADDLE_THROW(platform::errors::Unimplemented(
-            "The output of `PyLayer.forward` should be `Tensor`."));
+        // Only collect Tensor type in 'kwargs' and pass them to backward.
+        // Ignore other types of input temporarily.
       }
     } else {
-      PADDLE_THROW(platform::errors::Unimplemented(
-          "The output of `PyLayer.forward` can not be `None`."));
+      // Only collect Tensor type in 'kwargs' and pass them to backward.
+      // Ignore other types of input temporarily.
+    }
   }
+  if (output_vars.size() == 0) {
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "At least one output of `PyLayer.forward` is a `Tensor`."));
   }
 
   NameVarBaseMap outs = {{"Out", output_vars}};

paddle/fluid/operators/py_layer_op.cc

@@ -86,6 +86,12 @@ void RunPyObject(py::object *py_object,
       }
     }
   } else {
+    if (1 != outs->size()) {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "The number of outputs of `PyLayer.backward` should be %d, but "
+          "received 1.",
+          outs->size()));
+    }
     if ((*outs)[0] != nullptr) {
       if (Py_None != py_result.ptr()) {
         try {

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

@@ -30,7 +30,7 @@ class TestPyLayer(unittest.TestCase):
                 y1 = func1(x1)
                 y2 = func1(x2)
                 ctx.save_for_backward(y1, y2)
-                return y1, y2
+                return y1, 1, y2, None
 
             @staticmethod
             def backward(ctx, dy1, dy2):
@@ -44,7 +44,7 @@ class TestPyLayer(unittest.TestCase):
         input1.stop_gradient = False
         input2.stop_gradient = False
         z = tanh.apply(input1, input1, paddle.tanh, paddle.square)
-        z = z[0] + z[1]
+        z = z[0] + z[2]
         z.mean().backward()
 
         z2 = paddle.tanh(input2) + paddle.tanh(input2)
@@ -61,7 +61,7 @@ class TestPyLayer(unittest.TestCase):
                 y1 = func1(x1)
                 y2 = func1(x2)
                 ctx.save_for_backward(y1, y2)
-                return y1, y2
+                return 1, None, y1, y2, ''
 
             @staticmethod
             def backward(ctx, dy1, dy2):
@@ -79,7 +79,7 @@ class TestPyLayer(unittest.TestCase):
         input3.stop_gradient = True
         input4.stop_gradient = True
         z = tanh.apply(input1, input3, paddle.tanh, paddle.square)
-        z = z[0] + z[1]
+        z = z[2] + z[3]
         z.mean().backward()
 
         z2 = paddle.tanh(input2) + paddle.tanh(input4)
@@ -115,6 +115,27 @@ class TestPyLayer(unittest.TestCase):
         self.assertTrue(
             np.max(np.abs((input1.grad.numpy() - input2.grad.numpy()))) < 1e-10)
 
+    def test_pylayer_num_output_match(self):
+        class tanh(PyLayer):
+            @staticmethod
+            def forward(
+                    ctx,
+                    x1,
+                    x2, ):
+                return x1 + x2
+
+            @staticmethod
+            def backward(ctx, dy1):
+                return dy1 + 1
+
+        input1 = paddle.randn([2, 3]).astype("float64")
+        input2 = input1.detach().clone()
+        input1.stop_gradient = False
+        input2.stop_gradient = False
+        z = tanh.apply(input1, input2)
+        with self.assertRaises(ValueError):
+            z.mean().backward()
+
     def test_pylayer_dtype(self):
         class tanh(PyLayer):
             @staticmethod
@@ -150,20 +171,20 @@ class TestPyLayer(unittest.TestCase):
                 return args
 
         input1 = paddle.randn([2, 3]).astype("float64")
-        with self.assertRaises(NotImplementedError):
+        with self.assertRaises(ValueError):
             z = Layer_None1.apply(input1)
 
         class Layer_None2(PyLayer):
             @staticmethod
             def forward(ctx, *args):
-                return [None, None]
+                return [None, args[0]]
 
             @staticmethod
             def backward(ctx, *args):
                 return args
 
         input1 = paddle.randn([2, 3]).astype("float64")
-        with self.assertRaises(NotImplementedError):
+        # return None
         z = Layer_None2.apply(input1)
 
         class Layer_one1(PyLayer):
@@ -176,20 +197,21 @@ class TestPyLayer(unittest.TestCase):
                 return args
 
         input1 = paddle.randn([2, 3]).astype("float64")
-        with self.assertRaises(NotImplementedError):
+        # At least one output of `PyLayer.backward` is a `Tensor`
+        with self.assertRaises(ValueError):
             z = Layer_one1.apply(input1)
 
         class Layer_one2(PyLayer):
             @staticmethod
             def forward(ctx, *args):
-                return [1, 2]
+                return [1, 2, args[0]]
 
             @staticmethod
             def backward(ctx, *args):
                 return args
 
         input1 = paddle.randn([2, 3]).astype("float64")
-        with self.assertRaises(NotImplementedError):
+        # return int 
         z = Layer_one2.apply(input1)
 
         class Layer_no_fw(PyLayer):
@@ -234,7 +256,6 @@ class TestPyLayer(unittest.TestCase):
         z = Layer_bk_none1.apply(input2)
 
         with self.assertRaises(ValueError):
-            with paddle.fluid.dygraph.guard():
             z.sum().backward()
 
         class Layer_bk_none2(PyLayer):
@@ -249,8 +270,8 @@ class TestPyLayer(unittest.TestCase):
         input1 = paddle.randn([2, 3]).astype("float64")
         input1.stop_gradient = False
         z = Layer_bk_none2.apply(input1, input1)
+
         with self.assertRaises(ValueError):
-            with paddle.fluid.dygraph.guard():
             z.mean().backward()
 
         class Layer_bk_one1(PyLayer):
@@ -265,8 +286,8 @@ class TestPyLayer(unittest.TestCase):
         input1 = paddle.randn([2, 3]).astype("float64")
         input1.stop_gradient = False
         z = Layer_bk_one1.apply(input1)
+
         with self.assertRaises(ValueError):
-            with paddle.fluid.dygraph.guard():
             z.mean().backward()
 
         class Layer_bk_one2(PyLayer):
@@ -280,10 +301,10 @@ class TestPyLayer(unittest.TestCase):
 
         input1 = paddle.randn([2, 3]).astype("float64")
         input1.stop_gradient = False
+
         y = Layer_bk_one2.apply(input1, input1)
         z = y[0] + y[1]
         with self.assertRaises(ValueError):
-            with paddle.fluid.dygraph.guard():
             z.mean().backward()
 
         class Layer_no_bk(PyLayer):
@@ -295,8 +316,7 @@ class TestPyLayer(unittest.TestCase):
         input1.stop_gradient = False
         z = Layer_no_bk.apply(input1)
 
-        with self.assertRaises(NotImplementedError):
-            with paddle.fluid.dygraph.guard():
+        with self.assertRaises(OSError):
             z = z[0] + z[1]
             z.mean().backward()
 
@@ -313,7 +333,6 @@ class TestPyLayer(unittest.TestCase):
         input1.stop_gradient = False
         z = Layer_bk_match.apply(input1)
         with self.assertRaises(ValueError):
-            with paddle.fluid.dygraph.guard():
             z = z[0] + z[1]
             z.mean().backward()
 
@@ -334,7 +353,6 @@ class TestPyLayer(unittest.TestCase):
         z = Layer_bk_none1.apply(input1, input2)
 
         with self.assertRaises(ValueError):
-            with paddle.fluid.dygraph.guard():
             z.mean().backward()
 
         class Layer_bk_none2(PyLayer):
@@ -353,7 +371,6 @@ class TestPyLayer(unittest.TestCase):
         z = Layer_bk_none2.apply(input1, input2)
         z = z[0] + z[1]
         with self.assertRaises(ValueError):
-            with paddle.fluid.dygraph.guard():
             z.mean().backward()
 
     def test_pylayer_inplace(self):