Fix RoIPool reference implementation in Python 2

Also fixes a bug in the clip_boxes_to_image -- this function needs a test!

Author: fmassa

test/test_ops.py

@@ -26,7 +26,7 @@ def slow_roi_pooling(self, x, rois, pool_h, pool_w, spatial_scale=1,
                     start_h, end_h = int(roi[2].item()), int(roi[4].item()) + 1
                     start_w, end_w = int(roi[1].item()), int(roi[3].item()) + 1
                     roi_x = x[roi[0].long(), :, start_h:end_h, start_w:end_w]
-                    bin_h, bin_w = roi_x.size(-2) / pool_h, roi_x.size(-1) / pool_w
+                    bin_h, bin_w = roi_x.size(-2) / float(pool_h), roi_x.size(-1) / float(pool_w)
 
                     for j in range(0, pool_h):
                         cj = slice(int(np.floor(j * bin_h)), int(np.ceil((j + 1) * bin_h)))

torchvision/ops/boxes.py

@@ -72,13 +72,14 @@ def clip_boxes_to_image(boxes, size):
     Returns:
         clipped_boxes (Tensor[N, 4])
     """
-    boxes_x = boxes[:, 0::2]
-    boxes_y = boxes[:, 1::2]
+    dim = boxes.dim()
+    boxes_x = boxes[..., 0::2]
+    boxes_y = boxes[..., 1::2]
     height, width = size
     boxes_x = boxes_x.clamp(min=0, max=width)
-    boxes_y = boxes_x.clamp(min=0, max=height)
-    clipped_boxes = torch.stack((boxes_x, boxes_y), dim=2)
-    return clipped_boxes.reshape(-1, 4)
+    boxes_y = boxes_y.clamp(min=0, max=height)
+    clipped_boxes = torch.stack((boxes_x, boxes_y), dim=dim)
+    return clipped_boxes.reshape(boxes.shape)
 
 
 def box_area(boxes):