--- a
+++ b/ecg/examples/cinc17/build_datasets.py
@@ -0,0 +1,55 @@
+import json
+import numpy as np
+import os
+import random
+import scipy.io as sio
+import tqdm
+
+STEP = 256
+
+def load_ecg_mat(ecg_file):
+    return sio.loadmat(ecg_file)['val'].squeeze()
+
+def load_all(data_path):
+    label_file = os.path.join(data_path, "../REFERENCE-v3.csv")
+    with open(label_file, 'r') as fid:
+        records = [l.strip().split(",") for l in fid]
+
+    dataset = []
+    for record, label in tqdm.tqdm(records):
+        ecg_file = os.path.join(data_path, record + ".mat")
+        ecg_file = os.path.abspath(ecg_file)
+        ecg = load_ecg_mat(ecg_file)
+        num_labels = ecg.shape[0] // STEP
+        dataset.append((ecg_file, [label]*num_labels))
+    return dataset 
+
+def split(dataset, dev_frac):
+    dev_cut = int(dev_frac * len(dataset))
+    random.shuffle(dataset)
+    dev = dataset[:dev_cut]
+    train = dataset[dev_cut:]
+    return train, dev
+
+def make_json(save_path, dataset):
+    with open(save_path, 'w') as fid:
+        for d in dataset:
+            datum = {'ecg' : d[0],
+                     'labels' : d[1]}
+            json.dump(datum, fid)
+            fid.write('\n')
+
+from pathlib import Path
+
+
+if __name__ == "__main__":
+    random.seed(2018)
+
+    dev_frac = 0.1
+    home = str(Path.home())
+    data_path = home + "/ecg/examples/cinc17/data/training2017/"
+    dataset = load_all(data_path)
+    train, dev = split(dataset, dev_frac)
+    make_json("train.json", train)
+    make_json("dev.json", dev)
+