--- a
+++ b/experimental/get_anchors.py
@@ -0,0 +1,23 @@
+from sklearn.cluster import KMeans
+import numpy as np, pandas as pd, brambox as bb
+import pickle, argparse
+
+p=argparse.ArgumentParser()
+p.add_argument('--patch_size',default=512,type=int)
+p.add_argument('--n_anchors',default=20,type=int)
+p.add_argument('--sample_p',default=1.,type=float)
+
+args=p.parse_args()
+np.random.seed(42)
+patch_size=args.patch_size
+n_anchors=args.n_anchors
+sample_p=args.sample_p
+annotation_file = 'annotations_bbox_{}.pkl'.format(patch_size)
+annotations=bb.io.load('pandas',annotation_file)
+if sample_p<1.:
+    annotations=annotations.sample(frac=sample_p)
+
+X=annotations[['x_top_left','y_top_left']].astype(float).values+(annotations['width']/2.).astype(float).values.reshape(-1,1)
+km=KMeans(n_clusters=n_anchors,n_jobs=-1).fit(X)
+anchors=km.cluster_centers_
+pickle.dump(anchors,open('anchors.pkl','wb'))