--- a
+++ b/examples/irhythm/build_datasets.py
@@ -0,0 +1,151 @@
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import fnmatch
+import glob
+import json
+import numpy as np
+import os
+import random
+import tqdm
+
+STEP = 256
+RELABEL = {"NSR": "SINUS", "SUDDEN_BRADY": "AVB",
+           "AVB_TYPE2": "AVB", "AFIB": "AF", "AFL": "AF"}
+
+def get_all_records(path, blacklist=set()):
+    records = []
+    for root, dirnames, filenames in os.walk(path):
+        for filename in fnmatch.filter(filenames, '*.ecg'):
+            if patient_id(filename) not in blacklist:
+                records.append(os.path.abspath(
+                    os.path.join(root, filename)))
+    return records
+
+def patient_id(record):
+    return os.path.basename(record).split("_")[0]
+
+def round_to_step(n, step):
+    diff = (n - 1) % step 
+    if diff < (step / 2):
+        return n - diff
+    else:
+        return n + (step - diff)
+
+def load_episodes(record, epi_ext):
+
+    base = os.path.splitext(record)[0]
+    ep_json = base + epi_ext
+    ep_json = glob.glob(ep_json)[0]
+
+    with open(ep_json, 'r') as fid:
+        episodes = json.load(fid)['episodes']
+    episodes = sorted(episodes, key=lambda x: x['onset'])
+
+    for episode in episodes:
+        episode['onset_round'] = round_to_step(episode['onset'], STEP)
+        rn = episode['rhythm_name']
+        episode['rhythm_name'] = RELABEL.get(rn, rn)
+
+    for e, episode in enumerate(episodes):
+        if e == len(episodes) - 1:
+            episode['offset_round'] = episode['offset']
+        else:
+            episode['offset_round'] = episodes[e+1]['onset_round'] - 1
+    return episodes
+
+def make_labels(episodes):
+    labels = []
+    for episode in episodes:
+        rhythm_len = episode['offset_round'] - episode['onset_round'] + 1
+        rhythm_labels = int(rhythm_len / STEP)
+        rhythm = [episode['rhythm_name']] * rhythm_labels
+        labels.extend(rhythm)
+    trunc_samp = int(episodes[-1]['offset'] / STEP)
+    labels = labels[:trunc_samp]
+    return labels
+
+def build_blacklist(blacklist_paths):
+    print('Building blacklist...')
+    blacklist = set()
+    for blacklist_path in blacklist_paths:
+        print(blacklist_path)
+        for record in get_all_records(blacklist_path):
+            blacklist.add(patient_id(record))
+    return blacklist
+
+def construct_dataset(records, epi_ext='.episodes.json'):
+    data = []
+    for record in tqdm.tqdm(records):
+        labels = make_labels(load_episodes(record, epi_ext))
+        assert len(labels) != 0, "Zero labels?"
+        data.append((record, labels))
+    return data
+
+def stratify(records, dev_frac):
+    pids = list(set(patient_id(record) for record in records))
+    random.shuffle(pids)
+    cut = int(len(pids) * dev_frac)
+    dev_pids = set(pids[:cut])
+    train = [r for r in records if patient_id(r) not in dev_pids] 
+    dev = [r for r in records if patient_id(r) in dev_pids] 
+    return train, dev 
+
+def load_train(data_path, dev_frac, blacklist_paths):
+    blacklist = build_blacklist(blacklist_paths)
+    records = get_all_records(data_path, blacklist)
+    train, dev = stratify(records, dev_frac)
+    print("Constructing train...")
+    train = construct_dataset(train)
+    print("Constructing dev...")
+    dev = construct_dataset(dev)
+    return train, dev
+
+def load_rev_id(record, epi_ext):
+
+    base = os.path.splitext(record)[0]
+    ep_json = base + epi_ext
+    ep_json = glob.glob(ep_json)[0]
+
+    with open(ep_json, 'r') as fid:
+        return json.load(fid)['reviewer_id']
+
+def load_test(data_path, epi_ext):
+    records = get_all_records(data_path)
+    print("Constructing test...")
+    test = construct_dataset(records, epi_ext)
+    # Get the reviewer id
+    reviewers = [load_rev_id(r, epi_ext) for r in records]
+    test = [(e, l, r)
+            for (e, l), r in zip(test, reviewers)]
+    return test
+
+def make_json(save_path, dataset):
+    with open(save_path, 'w') as fid:
+        for d in dataset:
+            datum = {'ecg' : d[0],
+                     'labels' : d[1]}
+            if len(d) == 3:
+                datum['reviewer'] = d[2]
+            json.dump(datum, fid)
+            fid.write('\n')
+
+
+if __name__ == "__main__":
+    data_dir = "/deep/group/med/irhythm/ecg/clean_30sec_recs/"
+    blacklist_paths = [
+            os.path.join(data_dir, "label_review/CARDIOL_MAY_2017/"),
+            os.path.join(data_dir, "batches/kids_blacklist"),
+            os.path.join(data_dir, "batches/vf_blacklist")]
+    data_path = os.path.join(data_dir, "batches")
+    dev_frac = 0.1
+    train, dev = load_train(data_path, dev_frac, blacklist_paths)
+    make_json("train.json", train)
+    make_json("dev.json", dev)
+    test_dir = os.path.join(data_dir, "label_review/CARDIOL_UNIQ_P/")
+    test = load_test(test_dir, '_grp*.episodes.json')
+    make_json("test.json", test)
+    for i in range(6):
+        test = load_test(test_dir, "_rev{}.episodes.json".format(i))
+        make_json("test_rev{}.json".format(i), test)