add check for sparse parameters with weight_decay

Author: MRXLT

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

@@ -448,7 +448,6 @@ def test_adam_op_dygraph(self):
 
     def test_adam_op_with_state_dict(self):
 
-        import paddle
         paddle.disable_static()
         emb = paddle.nn.Embedding(10, 10)
 
@@ -517,6 +516,20 @@ def test_adam_op_invalid_input(self):
             adam = paddle.optimizer.Adam(
                 0.1, epsilon=-1, parameters=linear.parameters())
 
+    def test_adam_op_with_sparse_input_and_weight_decay(self):
+
+        paddle.disable_static()
+        x_data = np.arange(0, 10).reshape((10, 1)).astype(np.int64)
+        x = paddle.to_tensor(x_data, stop_gradient=False)
+        emb = paddle.nn.Embedding(10, 10, sparse=True)
+        adam = paddle.optimizer.Adam(
+            0.001, parameters=emb.parameters(), weight_decay=0.01)
+
+        with self.assertRaises(RuntimeError):
+            out = emb(x)
+            out.backward()
+            adam.step()
+
 
 if __name__ == "__main__":
     unittest.main()

python/paddle/optimizer/optimizer.py

@@ -913,6 +913,12 @@ def step(self):
         for param in self._parameter_list:
             if not param.trainable:
                 continue
+            if hasattr(
+                    param, "_is_sparse"
+            ) and param._is_sparse and self.regularization is not None:
+                raise RuntimeError(
+                    "Optimizer don't support weight_decay with sparse parameters, please set it to None."
+                )
             if param._grad_ivar() is not None:
                 grad_var = param._grad_ivar()
                 params_grads.append((param, grad_var))