Use tensor in window_reverse to avoid precision issue (#617)

暂时避免触发 scale OP，将 `window_size` 转为 Tensor 以触发

deploy/groundingdino/predict.py

@@ -123,7 +123,7 @@ def plot_boxes_to_image(image_pil, tgt):
     # draw boxes and masks
     for box, label in zip(boxes, labels):
         # from 0..1 to 0..W, 0..H
-        box = box * paddle.to_tensor([W, H, W, H])
+        box = box * paddle.to_tensor([W, H, W, H]).astype(paddle.float32)
         # from xywh to xyxy
         box[:2] -= box[2:] / 2
         box[2:] += box[:2]

paddlemix/examples/groundingdino/run_predict.py

@@ -40,7 +40,7 @@ def plot_boxes_to_image(image_pil, tgt):
     # draw boxes and masks
     for box, label in zip(boxes, labels):
         # from 0..1 to 0..W, 0..H
-        box = box * paddle.to_tensor([W, H, W, H])
+        box = box * paddle.to_tensor([W, H, W, H]).astype(paddle.float32)
         # from xywh to xyxy
         box[:2] -= box[2:] / 2
         box[2:] += box[:2]

paddlemix/models/audioldm2/clap_module/htsat_model.py

@@ -214,7 +214,7 @@ def window_reverse(windows, window_size, H, W):
     Returns:
         x: (B, H, W, C)
     """
-    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    B = int(windows.shape[0] / (H * W / paddle.to_tensor(window_size * window_size)))
     x = windows.reshape([B, H // window_size, W // window_size, window_size, window_size, -1])
     x = x.transpose([0, 1, 3, 2, 4, 5]).reshape([B, H, W, -1])
     return x

paddlemix/models/groundingdino/backbone/swin_transformer.py

@@ -164,7 +164,7 @@ def window_reverse(windows, window_size, H, W):
     Returns:
         x: (B, H, W, C)
     """
-    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    B = int(windows.shape[0] / (H * W / paddle.to_tensor(window_size * window_size)))
     x = windows.reshape([B, H // window_size, W // window_size, window_size, window_size, -1])
     x = x.transpose([0, 1, 3, 2, 4, 5]).reshape([B, H, W, -1])
     return x