Fix valid ratio for Deformable Detr (#20958)

* fix: valid ratio has right value

* chore: remove unnecessary line
Co-authored-by: Jeongyeon Nam <jy.nam@navercorp.com>

src/transformers/models/deformable_detr/modeling_deformable_detr.py

@@ -1509,8 +1509,8 @@ class DeformableDetrModel(DeformableDetrPreTrainedModel):
         """Get the valid ratio of all feature maps."""
 
         _, height, width = mask.shape
-        valid_height = torch.sum(~mask[:, :, 0], 1)
-        valid_width = torch.sum(~mask[:, 0, :], 1)
+        valid_height = torch.sum(mask[:, :, 0], 1)
+        valid_width = torch.sum(mask[:, 0, :], 1)
         valid_ratio_heigth = valid_height.float() / height
         valid_ratio_width = valid_width.float() / width
         valid_ratio = torch.stack([valid_ratio_width, valid_ratio_heigth], -1)
@@ -1687,9 +1687,6 @@ class DeformableDetrModel(DeformableDetrPreTrainedModel):
         spatial_shapes = torch.as_tensor(spatial_shapes, dtype=torch.long, device=source_flatten.device)
         level_start_index = torch.cat((spatial_shapes.new_zeros((1,)), spatial_shapes.prod(1).cumsum(0)[:-1]))
         valid_ratios = torch.stack([self.get_valid_ratio(m) for m in masks], 1)
-
-        # revert valid_ratios
-        valid_ratios = ~valid_ratios.bool()
         valid_ratios = valid_ratios.float()
 
         # Fourth, sent source_flatten + mask_flatten + lvl_pos_embed_flatten (backbone + proj layer output) through encoder