assert best possible recall > 0.9 before training

train.py

@@ -191,7 +191,7 @@ def train(hyp):
     model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device)  # attach class weights
     model.names = data_dict['names']
 
-    # class frequency
+    # Class frequency
     labels = np.concatenate(dataset.labels, 0)
     c = torch.tensor(labels[:, 0])  # classes
     # cf = torch.bincount(c.long(), minlength=nc) + 1.
@@ -199,6 +199,9 @@ def train(hyp):
     plot_labels(labels)
     tb_writer.add_histogram('classes', c, 0)
 
+    # Check anchors
+    check_best_possible_recall(dataset, anchors=model.model[-1].anchor_grid, thr=hyp['anchor_t'])
+
     # Exponential moving average
     ema = torch_utils.ModelEMA(model)
 

utils/datasets.py

@@ -291,8 +291,6 @@ class LoadImagesAndLabels(Dataset):  # for training/testing
         self.label_files = [x.replace('images', 'labels').replace(os.path.splitext(x)[-1], '.txt')
                             for x in self.img_files]
 
-        # Rectangular Training  https://github.com/ultralytics/yolov3/issues/232
-        if self.rect:
         # Read image shapes (wh)
         sp = path.replace('.txt', '') + '.shapes'  # shapefile path
         try:
@@ -303,8 +301,12 @@ class LoadImagesAndLabels(Dataset):  # for training/testing
             s = [exif_size(Image.open(f)) for f in tqdm(self.img_files, desc='Reading image shapes')]
             np.savetxt(sp, s, fmt='%g')  # overwrites existing (if any)
 
+        self.shapes = np.array(s, dtype=np.float64)
+
+        # Rectangular Training  https://github.com/ultralytics/yolov3/issues/232
+        if self.rect:
             # Sort by aspect ratio
-            s = np.array(s, dtype=np.float64)
+            s = self.shapes  # wh
             ar = s[:, 1] / s[:, 0]  # aspect ratio
             irect = ar.argsort()
             self.img_files = [self.img_files[i] for i in irect]

utils/utils.py

@@ -51,6 +51,19 @@ def check_img_size(img_size, s=32):
     return make_divisible(img_size, s)  # nearest gs-multiple
 
 
+def check_best_possible_recall(dataset, anchors, thr):
+    # Check best possible recall of dataset with current anchors
+    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(dataset.shapes, dataset.labels)]))  # width-height
+    ratio = wh[:, None] / anchors.view(-1, 2)[None]  # ratio
+    m = torch.max(ratio, 1. / ratio).max(2)[0]  # max ratio
+    bpr = (m.min(1)[0] < thr).float().mean()  # best possible recall
+    mr = (m < thr).float().mean()  # match ratio
+    print(('Label width-height:' + '%10s' * 6) % ('n', 'mean', 'min', 'max', 'matching', 'recall'))
+    print(('                   ' + '%10.4g' * 6) % (wh.shape[0], wh.mean(), wh.min(), wh.max(), mr, bpr))
+    assert bpr > 0.9, 'Best possible recall %.3g (BPR) below 0.9 threshold. Training cancelled. ' \
+                      'Compute new anchors with utils.utils.kmeans_anchors() and update model before training.' % bpr
+
+
 def make_divisible(x, divisor):
     # Returns x evenly divisble by divisor
     return math.ceil(x / divisor) * divisor