[fbsync] fix bug when the target is empty in FCOS (#5267)

Summary:
* fix bug when the target is empty

* Add unittest for empty instance training

Reviewed By: kazhang

Differential Revision: D33927512

fbshipit-source-id: e92355380948d9181e135b7612596c5309afeeda

Co-authored-by: Zhiqiang Wang <[email protected]>
Co-authored-by: Joao Gomes <[email protected]>

Author: 3 people

test/test_models_detection_negative_samples.py

@@ -143,6 +143,17 @@ def test_forward_negative_sample_retinanet(self):
 
         assert_equal(loss_dict["bbox_regression"], torch.tensor(0.0))
 
+    def test_forward_negative_sample_fcos(self):
+        model = torchvision.models.detection.fcos_resnet50_fpn(
+            num_classes=2, min_size=100, max_size=100, pretrained_backbone=False
+        )
+
+        images, targets = self._make_empty_sample()
+        loss_dict = model(images, targets)
+
+        assert_equal(loss_dict["bbox_regression"], torch.tensor(0.0))
+        assert_equal(loss_dict["bbox_ctrness"], torch.tensor(0.0))
+
     def test_forward_negative_sample_ssd(self):
         model = torchvision.models.detection.ssd300_vgg16(num_classes=2, pretrained_backbone=False)
 

torchvision/models/detection/fcos.py

@@ -59,9 +59,13 @@ def compute_loss(
         all_gt_classes_targets = []
         all_gt_boxes_targets = []
         for targets_per_image, matched_idxs_per_image in zip(targets, matched_idxs):
-            gt_classes_targets = targets_per_image["labels"][matched_idxs_per_image.clip(min=0)]
+            if len(targets_per_image["labels"]) == 0:
+                gt_classes_targets = targets_per_image["labels"].new_zeros((len(matched_idxs_per_image),))
+                gt_boxes_targets = targets_per_image["boxes"].new_zeros((len(matched_idxs_per_image), 4))
+            else:
+                gt_classes_targets = targets_per_image["labels"][matched_idxs_per_image.clip(min=0)]
+                gt_boxes_targets = targets_per_image["boxes"][matched_idxs_per_image.clip(min=0)]
             gt_classes_targets[matched_idxs_per_image < 0] = -1  # backgroud
-            gt_boxes_targets = targets_per_image["boxes"][matched_idxs_per_image.clip(min=0)]
             all_gt_classes_targets.append(gt_classes_targets)
             all_gt_boxes_targets.append(gt_boxes_targets)
 
@@ -95,13 +99,14 @@ def compute_loss(
         ]
         bbox_reg_targets = torch.stack(bbox_reg_targets, dim=0)
         if len(bbox_reg_targets) == 0:
-            bbox_reg_targets.new_zeros(len(bbox_reg_targets))
-        left_right = bbox_reg_targets[:, :, [0, 2]]
-        top_bottom = bbox_reg_targets[:, :, [1, 3]]
-        gt_ctrness_targets = torch.sqrt(
-            (left_right.min(dim=-1)[0] / left_right.max(dim=-1)[0])
-            * (top_bottom.min(dim=-1)[0] / top_bottom.max(dim=-1)[0])
-        )
+            gt_ctrness_targets = bbox_reg_targets.new_zeros(bbox_reg_targets.size()[:-1])
+        else:
+            left_right = bbox_reg_targets[:, :, [0, 2]]
+            top_bottom = bbox_reg_targets[:, :, [1, 3]]
+            gt_ctrness_targets = torch.sqrt(
+                (left_right.min(dim=-1)[0] / left_right.max(dim=-1)[0])
+                * (top_bottom.min(dim=-1)[0] / top_bottom.max(dim=-1)[0])
+            )
         pred_centerness = bbox_ctrness.squeeze(dim=2)
         loss_bbox_ctrness = nn.functional.binary_cross_entropy_with_logits(
             pred_centerness[foregroud_mask], gt_ctrness_targets[foregroud_mask], reduction="sum"