# Define a sampler to balance the classes
# training dataset
lbls = [dataset[idx][1] for idx in train_set.indices]
bc = np.bincount(lbls)
p_nOK = bc.sum()/bc[0]
p_OK = bc.sum()/bc[1]
lst_train = [p_nOK if lbl==0 else p_OK for lbl in lbls]
train_sampler = WeightedRandomSampler(weights=lst_train, num_samples=len(lbls))