[fbsync] Adding ciou and diou support in _box_loss() (#5984)

Summary:
* Adding ciou and diou support in `_box_loss()`

* Fix linter

* Addressing comments for nits

Reviewed By: datumbox

Differential Revision: D36413352

fbshipit-source-id: 8a53c68ff2a58966f02edae7f210c8db5240a4d2

Author: NicolasHug

torchvision/models/detection/_utils.py

@@ -5,7 +5,7 @@
 import torch
 from torch import Tensor, nn
 from torch.nn import functional as F
-from torchvision.ops import FrozenBatchNorm2d, generalized_box_iou_loss
+from torchvision.ops import FrozenBatchNorm2d, complete_box_iou_loss, distance_box_iou_loss, generalized_box_iou_loss
 
 
 class BalancedPositiveNegativeSampler:
@@ -518,7 +518,7 @@ def _box_loss(
     bbox_regression_per_image: Tensor,
     cnf: Optional[Dict[str, float]] = None,
 ) -> Tensor:
-    torch._assert(type in ["l1", "smooth_l1", "giou"], f"Unsupported loss: {type}")
+    torch._assert(type in ["l1", "smooth_l1", "ciou", "diou", "giou"], f"Unsupported loss: {type}")
 
     if type == "l1":
         target_regression = box_coder.encode_single(matched_gt_boxes_per_image, anchors_per_image)
@@ -527,7 +527,12 @@ def _box_loss(
         target_regression = box_coder.encode_single(matched_gt_boxes_per_image, anchors_per_image)
         beta = cnf["beta"] if cnf is not None and "beta" in cnf else 1.0
         return F.smooth_l1_loss(bbox_regression_per_image, target_regression, reduction="sum", beta=beta)
-    else:  # giou
+    else:
         bbox_per_image = box_coder.decode_single(bbox_regression_per_image, anchors_per_image)
         eps = cnf["eps"] if cnf is not None and "eps" in cnf else 1e-7
+        if type == "ciou":
+            return complete_box_iou_loss(bbox_per_image, matched_gt_boxes_per_image, reduction="sum", eps=eps)
+        if type == "diou":
+            return distance_box_iou_loss(bbox_per_image, matched_gt_boxes_per_image, reduction="sum", eps=eps)
+        # otherwise giou
         return generalized_box_iou_loss(bbox_per_image, matched_gt_boxes_per_image, reduction="sum", eps=eps)