updates

models.py

@@ -156,7 +156,8 @@ class YOLOLayer(nn.Module):
             io[..., :4] *= self.stride
 
             io[..., 4:] = torch.sigmoid(io[..., 4:])  # p_conf, p_cls
-            # io[..., 5:] = F.softmax(io[..., 5:], dim=4)  # p_cls
+            # io[..., 4:] = F.softmax(io[..., 4:], dim=4)  # unified detection CE
+            # io[..., 4] = io[..., 5:].max(4)[0]  # unified detection BCE
 
             if self.nc == 1:
                 io[..., 5] = 1  # single-class model https://github.com/ultralytics/yolov3/issues/235

train.py

@@ -37,7 +37,7 @@ except:
 hyp = {'giou': 1.582,  # giou loss gain
        'xy': 4.688,  # xy loss gain
        'wh': 0.1857,  # wh loss gain
-       'cls': 27.76,  # cls loss gain  (CE should be around ~1.0)
+       'cls': 27.76,  # cls loss gain  (CE=~1.0, uCE=~20, uBCE=~200,~30)
        'cls_pw': 1.446,  # cls BCELoss positive_weight
        'obj': 21.35,  # obj loss gain
        'obj_pw': 3.941,  # obj BCELoss positive_weight

utils/utils.py

@@ -324,6 +324,14 @@ def compute_loss(p, targets, model, giou_loss=True):  # predictions, targets, mo
                 lcls += (k * h['cls']) * BCEcls(pi[..., 5:], tclsm)  # BCE
                 # lcls += (k * h['cls']) * CE(pi[..., 5:], tcls[i])  # CE
 
+            # udm_ce = torch.zeros_like(pi0[..., 0]).long()  # unified detection matrix for CE
+            # udm_ce[b, a, gj, gi] = tcls[i] + 1
+            # lcls += (k * h['cls']) * CE(pi0[..., 4:].view(-1, model.nc + 1), udm_ce.view(-1))  # unified CE
+
+            # udm = torch.zeros_like(pi0[..., 5:])  # unified detection matrix for BCE
+            # udm[b, a, gj, gi, tcls[i]] = 1.0
+            # lcls += (k * h['cls']) * BCEcls(pi0[..., 5:], udm)  # unified BCE (hyps 200-30)
+
             # Append targets to text file
             # with open('targets.txt', 'a') as file:
             #     [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]