FIX: trainig fails if targets list is empty (#198)

* FIX: trainig fails if targets list is empty

* Update utils.py

utils/utils.py

@@ -288,22 +288,23 @@ def build_targets(model, targets):
     if type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel):
         model = model.module
 
+    nt = len(targets)
     txy, twh, tcls, indices = [], [], [], []
     for i in model.yolo_layers:
         layer = model.module_list[i][0]
 
         # iou of targets-anchors
+        t, a = targets, []
         gwh = targets[:, 4:6] * layer.nG
-        iou = [wh_iou(x, gwh) for x in layer.anchor_vec]
-        iou, a = torch.stack(iou, 0).max(0)  # best iou and anchor
+        if nt:
+            iou = [wh_iou(x, gwh) for x in layer.anchor_vec]
+            iou, a = torch.stack(iou, 0).max(0)  # best iou and anchor
 
-        # reject below threshold ious (OPTIONAL, increases P, lowers R)
-        reject = True
-        if reject:
-            j = iou > 0.10
-            t, a, gwh = targets[j], a[j], gwh[j]
-        else:
-            t = targets
+            # reject below threshold ious (OPTIONAL, increases P, lowers R)
+            reject = True
+            if reject:
+                j = iou > 0.10
+                t, a, gwh = targets[j], a[j], gwh[j]
 
         # Indices
         b, c = t[:, :2].long().t()  # target image, class
@@ -320,7 +321,7 @@ def build_targets(model, targets):
 
         # Class
         tcls.append(c)
-        if c.shape[0]:
+        if nt:
             assert c.max() <= layer.nC, 'Target classes exceed model classes'
 
     return txy, twh, tcls, indices