updates

train.py

@@ -11,19 +11,33 @@ from models import *
 from utils.datasets import *
 from utils.utils import *
 
-# Hyperparameters: train.py --evolve --epochs 2 --img-size 320, Metrics: 0.204      0.302      0.175      0.234 (square smart)
+# Hyperparameters: train.py --data data/coco.data --img-size 320 --single-scale --batch-size 64 --accumulate 1 --epochs 1 --evolve     0.087      0.281      0.109      0.121
 hyp = {'giou': .035,  # giou loss gain
        'xy': 0.20,  # xy loss gain
        'wh': 0.10,  # wh loss gain
        'cls': 0.035,  # cls loss gain
+       'cls_pw': 79.0,  # cls BCELoss positive_weight
        'conf': 1.61,  # conf loss gain
-       'conf_bpw': 3.53,  # conf BCELoss positive_weight
+       'conf_pw': 3.53,  # conf BCELoss positive_weight
        'iou_t': 0.29,  # iou target-anchor training threshold
        'lr0': 0.001,  # initial learning rate
        'lrf': -4.,  # final learning rate = lr0 * (10 ** lrf)
        'momentum': 0.90,  # SGD momentum
        'weight_decay': 0.0005}  # optimizer weight decay
 
+# hyp = {'giou': 1.0,  # giou loss gain
+#        'xy': 1.0,  # xy loss gain
+#        'wh': 1.0,  # wh loss gain
+#        'cls': 1.0,  # cls loss gain
+#        'cls_pw': 79.0,  # cls BCELoss positive_weight
+#        'conf': 1.0,  # conf loss gain
+#        'conf_pw': 6.0,  # conf BCELoss positive_weight
+#        'iou_t': 0.29,  # iou target-anchor training threshold
+#        'lr0': 0.001,  # initial learning rate
+#        'lrf': -4.,  # final learning rate = lr0 * (10 ** lrf)
+#        'momentum': 0.90,  # SGD momentum
+#        'weight_decay': 0.0005}  # optimizer weight decay
+
 
 def train(
         cfg,
@@ -329,7 +343,7 @@ if __name__ == '__main__':
             # Mutate hyperparameters
             old_hyp = hyp.copy()
             init_seeds(seed=int(time.time()))
-            s = [.4, .4, .4, .4, .4, .4, .4, .4 * 0, .4 * 0, .04 * 0, .4 * 0]  # fractional sigmas
+            s = [.4, .4, .4, .4, .4, .4, .4, .4, .4 * 0, .4 * 0, .04 * 0, .4 * 0]  # fractional sigmas
             for i, k in enumerate(hyp.keys()):
                 x = (np.random.randn(1) * s[i] + 1) ** 1.1  # plt.hist(x.ravel(), 100)
                 hyp[k] = hyp[k] * float(x)  # vary by about 30% 1sigma

utils/utils.py

@@ -279,7 +279,8 @@ def compute_loss(p, targets, model, giou_loss=False):  # predictions, targets, m
     # Define criteria
     MSE = nn.MSELoss()
     CE = nn.CrossEntropyLoss()  # (weight=model.class_weights)
-    BCE = nn.BCEWithLogitsLoss(pos_weight=ft([h['conf_bpw']]))
+    BCEcls = nn.BCEWithLogitsLoss(pos_weight=ft([h['cls_pw']]))
+    BCEconf = nn.BCEWithLogitsLoss(pos_weight=ft([h['conf_pw']]))
 
     # Compute losses
     bs = p[0].shape[0]  # batch size
@@ -301,13 +302,17 @@ def compute_loss(p, targets, model, giou_loss=False):  # predictions, targets, m
             else:
                 lxy += (k * h['xy']) * MSE(torch.sigmoid(pi[..., 0:2]), txy[i])  # xy loss
                 lwh += (k * h['wh']) * MSE(pi[..., 2:4], twh[i])  # wh yolo loss
+
+            # tclsm = torch.zeros_like(pi[..., 5:])
+            # tclsm[range(len(b)), tcls[i]] = 1.0
+            # lcls += (k * h['cls']) * BCEcls(pi[..., 5:], tclsm)  # class_conf loss
             lcls += (k * h['cls']) * CE(pi[..., 5:], tcls[i])  # class_conf loss
 
             # # Append 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)]
 
-        lconf += (k * h['conf']) * BCE(pi0[..., 4], tconf)  # obj_conf loss
+        lconf += (k * h['conf']) * BCEconf(pi0[..., 4], tconf)  # obj_conf loss
     loss = lxy + lwh + lconf + lcls
 
     return loss, torch.cat((lxy, lwh, lconf, lcls, loss)).detach()