updates

test.py

@@ -93,7 +93,7 @@ def test(
                 # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ...
                 image_id = int(Path(paths[si]).stem.split('_')[-1])
                 box = pred[:, :4].clone()  # xyxy
-                scale_coords(img_size, box, shapes[si])  # to original shape
+                scale_coords(imgs[si].shape, box, shapes[si])  # to original shape
                 box = xyxy2xywh(box)  # xywh
                 box[:, :2] -= box[:, 2:] / 2  # xy center to top-left corner
                 for di, d in enumerate(pred):

train.py

@@ -11,20 +11,34 @@ from utils.datasets import *
 from utils.utils import *
 
 # Hyperparameters
-# 0.852       0.94      0.924      0.883       1.33       8.52    0.06833    0.01524    0.01509     0.9013     0.1003   0.001325     -3.853     0.8948  0.0004053  # hyp
+# 0.861      0.956      0.936      0.897       1.51      10.39     0.1367    0.01057    0.01181     0.8409     0.1287   0.001028     -3.441     0.9127  0.0004841
-hyp = {'k': 8.52,  # loss multiple
+hyp = {'k': 10.39,  # loss multiple
-       'xy': 0.06833,  # xy loss fraction
+       'xy': 0.1367,  # xy loss fraction
-       'wh': 0.01524,  # wh loss fraction
+       'wh': 0.01057,  # wh loss fraction
-       'cls': 0.01509,  # cls loss fraction
+       'cls': 0.01181,  # cls loss fraction
-       'conf': 0.9013,  # conf loss fraction
+       'conf': 0.8409,  # conf loss fraction
-       'iou_t': 0.1003,  # iou target-anchor training threshold
+       'iou_t': 0.1287,  # iou target-anchor training threshold
-       'lr0': 0.001325,  # initial learning rate
+       'lr0': 0.001028,  # initial learning rate
-       'lrf': -3.853,  # final learning rate = lr0 * (10 ** lrf)
+       'lrf': -3.441,  # final learning rate = lr0 * (10 ** lrf)
-       'momentum': 0.8948,  # SGD momentum
+       'momentum': 0.9127,  # SGD momentum
-       'weight_decay': 0.0004053,  # optimizer weight decay
+       'weight_decay': 0.0004841,  # optimizer weight decay
        }
 
 
+# 0.856       0.95      0.935      0.887        1.3      8.488     0.1081    0.01351    0.01351     0.8649        0.1      0.001         -3        0.9     0.0005
+# hyp = {'k': 8.4875,  # loss multiple
+#        'xy': 0.108108,  # xy loss fraction
+#        'wh': 0.013514,  # wh loss fraction
+#        'cls': 0.013514,  # cls loss fraction
+#        'conf': 0.86486,  # conf loss fraction
+#        'iou_t': 0.1,  # iou target-anchor training threshold
+#        'lr0': 0.001,  # initial learning rate
+#        'lrf': -3.,  # final learning rate = lr0 * (10 ** lrf)
+#        'momentum': 0.9,  # SGD momentum
+#        'weight_decay': 0.0005,  # optimizer weight decay
+#        }
+
+
 def train(
         cfg,
         data_cfg,
@@ -89,12 +103,9 @@ def train(
     # Scheduler (reduce lr at epochs 218, 245, i.e. batches 400k, 450k)
     # lf = lambda x: 1 - x / epochs  # linear ramp to zero
     # lf = lambda x: 10 ** (-2 * x / epochs)  # exp ramp to lr0 * 1e-2
-    # lf = lambda x: 1 - 10 ** (hyp['lrf'] * (1 - x / epochs))  # inv exp ramp to lr0 * 1e-2
+    lf = lambda x: 1 - 10 ** (hyp['lrf'] * (1 - x / epochs))  # inv exp ramp to lr0 * 1e-2
-    # scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lf, last_epoch=start_epoch - 1)
+    scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lf, last_epoch=start_epoch - 1)
-    scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
+    # scheduler = optim.lr_scheduler.MultiStepLR(optimizer,milestones=[218, 245],gamma=0.1,last_epoch=start_epoch - 1)
-                                               milestones=[218, 245],
-                                               gamma=0.1,
-                                               last_epoch=start_epoch - 1)
 
     # Plot lr schedule
     # y = []
@@ -311,7 +322,7 @@ if __name__ == '__main__':
             # Mutate hyperparameters
             old_hyp = hyp.copy()
             init_seeds(seed=int(time.time()))
-            s = [.2, .2, .2, .2, .2, .2, .2, .2, .02, .2]
+            s = [.2, .2, .2, .2, .2, .3, .2, .2, .02, .3]
             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/datasets.py

@@ -74,7 +74,7 @@ class LoadImages:  # for inference
             print('image %g/%g %s: ' % (self.count, self.nF, path), end='')
 
         # Padded resize
-        img, _, _, _ = letterbox(img0, height=self.height)
+        img, _, _, _ = letterbox_rect(img0, height=self.height)
 
         # Normalize RGB
         img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB
@@ -239,9 +239,11 @@ def letterbox(img, height=416, color=(127.5, 127.5, 127.5)):
     # Resize a rectangular image to a padded square
     shape = img.shape[:2]  # shape = [height, width]
     ratio = float(height) / max(shape)  # ratio  = old / new
-    new_shape = (round(shape[1] * ratio), round(shape[0] * ratio))
+    new_shape = (round(shape[1] * ratio), round(shape[0] * ratio))  # new_shape = [width, height]
+
     dw = (height - new_shape[0]) / 2  # width padding
     dh = (height - new_shape[1]) / 2  # height padding
+
     top, bottom = round(dh - 0.1), round(dh + 0.1)
     left, right = round(dw - 0.1), round(dw + 0.1)
     img = cv2.resize(img, new_shape, interpolation=cv2.INTER_AREA)  # resized, no border
@@ -249,6 +251,22 @@ def letterbox(img, height=416, color=(127.5, 127.5, 127.5)):
     return img, ratio, dw, dh
 
 
+def letterbox_rect(img, height=416, color=(127.5, 127.5, 127.5)):
+    # Resize a rectangular image to a 32 pixel multiple rectangle
+    shape = img.shape[:2]  # shape = [height, width]
+    ratio = float(height) / max(shape)  # ratio  = old / new
+    new_shape = (round(shape[1] * ratio), round(shape[0] * ratio))  # new_shape = [width, height]
+
+    dw = np.mod(height - new_shape[0], 32) / 2  # width padding
+    dh = np.mod(height - new_shape[1], 32) / 2  # height padding
+
+    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
+    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
+    img = cv2.resize(img, new_shape, interpolation=cv2.INTER_AREA)  # resized, no border
+    img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # padded square
+    return img, ratio, dw, dh
+
+
 def random_affine(img, targets=(), degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-2, 2),
                   borderValue=(127.5, 127.5, 127.5)):
     # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))

utils/utils.py

@@ -100,11 +100,11 @@ def xywh2xyxy(x):
     return y
 
 
-def scale_coords(img_size, coords, img0_shape):
+def scale_coords(img1_shape, coords, img0_shape):
-    # Rescale x1, y1, x2, y2 from 416 to image size
+    # Rescale coords1 (xyxy) from img1_shape to img0_shape
-    gain = float(img_size) / max(img0_shape)  # gain  = old / new
+    gain = max(img1_shape[1:3]) / max(img0_shape[:2])  # gain  = old / new
-    pad_x = (img_size - img0_shape[1] * gain) / 2  # width padding
+    pad_x = (img1_shape[2] - img0_shape[1] * gain) / 2  # width padding
-    pad_y = (img_size - img0_shape[0] * gain) / 2  # height padding
+    pad_y = (img1_shape[1] - img0_shape[0] * gain) / 2  # height padding
     coords[:, [0, 2]] -= pad_x
     coords[:, [1, 3]] -= pad_y
     coords[:, :4] /= gain