diff --git a/paddleseg/models/decoupled_segnet.py b/paddleseg/models/decoupled_segnet.py
index 8386d6b74d..0f7b26e41d 100644
--- a/paddleseg/models/decoupled_segnet.py
+++ b/paddleseg/models/decoupled_segnet.py
@@ -226,7 +226,7 @@ def flow_warp(self, input, flow, size):
         w_grid = w_grid.tile([size[0]]).transpose([1, 0])
         grid = paddle.concat([w_grid.unsqueeze(2), h_grid.unsqueeze(2)], axis=2)
         grid.unsqueeze(0).tile([input_shape[0], 1, 1, 1])
-        grid = grid + paddle.transpose(flow, (0, 2, 3, 1)) / norm
+        grid = grid + paddle.transpose(flow, (0, 2, 3, 1)) / norm.astype(flow.dtype)
 
         output = F.grid_sample(input, grid)
         return output
diff --git a/paddleseg/models/losses/cross_entropy_loss.py b/paddleseg/models/losses/cross_entropy_loss.py
index b1cfb3a624..b563ded03b 100644
--- a/paddleseg/models/losses/cross_entropy_loss.py
+++ b/paddleseg/models/losses/cross_entropy_loss.py
@@ -124,13 +124,13 @@ def _post_process_loss(self, logit, label, semantic_weights, loss):
             loss = loss * semantic_weights
 
         if self.weight is not None:
-            _one_hot = F.one_hot(label * mask, logit.shape[-1])
+            _one_hot = F.one_hot(label * mask.astype(label.dtype), logit.shape[-1])
             coef = paddle.sum(_one_hot * self.weight, axis=-1)
         else:
             coef = paddle.ones_like(label)
 
         if self.top_k_percent_pixels == 1.0:
-            avg_loss = paddle.mean(loss) / (paddle.mean(mask * coef) + self.EPS)
+            avg_loss = paddle.mean(loss) / (paddle.mean(mask * coef.astype(mask.dtype)) + self.EPS)
         else:
             loss = loss.reshape((-1, ))
             top_k_pixels = int(self.top_k_percent_pixels * loss.numel())
diff --git a/paddleseg/models/losses/ohem_cross_entropy_loss.py b/paddleseg/models/losses/ohem_cross_entropy_loss.py
index 56bb49c0e2..aefeb3ac3c 100644
--- a/paddleseg/models/losses/ohem_cross_entropy_loss.py
+++ b/paddleseg/models/losses/ohem_cross_entropy_loss.py
@@ -65,7 +65,7 @@ def forward(self, logit, label):
 
         if self.min_kept < num_valid and num_valid > 0:
             # let the value which ignored greater than 1
-            prob = prob + (1 - valid_mask)
+            prob = prob + (1 - valid_mask).astype(prob.dtype)
 
             # get the prob of relevant label
             label_onehot = F.one_hot(label, c)
diff --git a/paddleseg/models/sfnet.py b/paddleseg/models/sfnet.py
index d04dfbe5d7..4859f576a9 100644
--- a/paddleseg/models/sfnet.py
+++ b/paddleseg/models/sfnet.py
@@ -219,7 +219,7 @@ def flow_warp(self, input, flow, size):
         w_grid = w_grid.tile([size[0]]).transpose([1, 0])
         grid = paddle.concat([w_grid.unsqueeze(2), h_grid.unsqueeze(2)], axis=2)
         grid.unsqueeze(0).tile([input_shape[0], 1, 1, 1])
-        grid = grid + paddle.transpose(flow, (0, 2, 3, 1)) / norm
+        grid = grid + paddle.transpose(flow, (0, 2, 3, 1)) / norm.astype(flow.dtype)
 
         output = F.grid_sample(input, grid)
         return output