Change reshape to support empty batches. (pytorch#3031)

Author: datumbox

torchvision/models/detection/roi_heads.py

@@ -40,7 +40,7 @@ def fastrcnn_loss(class_logits, box_regression, labels, regression_targets):
     sampled_pos_inds_subset = torch.where(labels > 0)[0]
     labels_pos = labels[sampled_pos_inds_subset]
     N, num_classes = class_logits.shape
-    box_regression = box_regression.reshape(N, -1, 4)
+    box_regression = box_regression.reshape(N, box_regression.size(-1) // 4, 4)
 
     box_loss = det_utils.smooth_l1_loss(
         box_regression[sampled_pos_inds_subset, labels_pos],