Spaces:

laolida-w
/

1

Sleeping

laolida-w commited on 18 days ago

Commit

5bb2d2a

verified ·

1 Parent(s): 697a4fe

Update eval_dataset.py

Files changed (1) hide show

eval_dataset.py CHANGED Viewed

@@ -134,12 +134,12 @@ class SingleRegionCaptionDataset(Dataset):
         aspect_ratio = model_inputs["aspect_ratio"]
         ret = dict(
-            input_ids=input_ids.cuda().unsqueeze(0),
-            attention_mask=attention_mask.cuda().to(self.data_dtype).unsqueeze(0),
-            pixel_values=pixel_values.cuda().to(self.data_dtype).flatten(0, 1),
-            global_mask_values=mask_values.cuda().to(self.data_dtype).squeeze(),
             bboxes=[data_dict["bboxes"]],
-            aspect_ratios=aspect_ratio.unsqueeze(0).cuda(),
         )
         return ret
@@ -273,11 +273,11 @@ class MultiRegionDataset(Dataset):
         aspect_ratio = model_inputs["aspect_ratio"]
         ret = dict(
-            input_ids=input_ids.cuda().unsqueeze(0),
-            attention_mask=attention_mask.cuda().to(self.data_dtype).unsqueeze(0),
-            pixel_values=pixel_values.cuda().to(self.data_dtype).flatten(0, 1),
-            global_mask_values=mask_values.cuda().to(self.data_dtype).squeeze(),
             bboxes=[data_dict["bboxes"]],
-            aspect_ratios=aspect_ratio.unsqueeze(0).cuda(),
         )
         return ret

         aspect_ratio = model_inputs["aspect_ratio"]
         ret = dict(
+            input_ids=input_ids.unsqueeze(0),
+            attention_mask=attention_mask.to(self.data_dtype).unsqueeze(0),
+            pixel_values=pixel_values.to(self.data_dtype).flatten(0, 1),
+            global_mask_values=mask_values.to(self.data_dtype).squeeze(),
             bboxes=[data_dict["bboxes"]],
+            aspect_ratios=aspect_ratio.unsqueeze(0),
         )
         return ret
         aspect_ratio = model_inputs["aspect_ratio"]
         ret = dict(
+            input_ids=input_ids.unsqueeze(0),
+            attention_mask=attention_mask.to(self.data_dtype).unsqueeze(0),
+            pixel_values=pixel_values.to(self.data_dtype).flatten(0, 1),
+            global_mask_values=mask_values.to(self.data_dtype).squeeze(),
             bboxes=[data_dict["bboxes"]],
+            aspect_ratios=aspect_ratio.unsqueeze(0),
         )
         return ret