--- a
+++ b/inpainting/data/data.py
@@ -0,0 +1,51 @@
+import torch
+import numpy as np
+import cv2
+import os
+from torch.utils.data import Dataset
+
+
+class ToTensor(object):
+    def __call__(self, sample):
+        entry = {}
+        for k in sample:
+            if k == 'rect':
+                entry[k] = torch.IntTensor(sample[k])
+            else:
+                entry[k] = torch.FloatTensor(sample[k])
+        return entry
+
+
+class InpaintingDataset(Dataset):
+    def __init__(self, info_list, root_dir='', im_size=(256, 256), transform=None):
+        self.filenames = open(info_list, 'rt').read().splitlines()
+        self.root_dir = root_dir
+        self.transform = transform
+        self.im_size = im_size
+        np.random.seed(2018)
+
+    def __len__(self):
+        return len(self.filenames)
+
+    def read_image(self, filepath):
+        image = cv2.imread(filepath)
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+        h, w, c = image.shape
+        if h != self.im_size[0] or w != self.im_size[1]:
+            ratio = max(1.0 * self.im_size[0] / h, 1.0 * self.im_size[1] / w)
+            im_scaled = cv2.resize(image, None, fx=ratio, fy=ratio)
+            h, w, _ = im_scaled.shape
+            h_idx = (h - self.im_size[0]) // 2
+            w_idx = (w - self.im_size[1]) // 2
+            im_scaled = im_scaled[h_idx:h_idx + self.im_size[0], w_idx:w_idx + self.im_size[1], :]
+            im_scaled = np.transpose(im_scaled, [2, 0, 1])
+        else:
+            im_scaled = np.transpose(image, [2, 0, 1])
+        return im_scaled
+
+    def __getitem__(self, idx):
+        image = self.read_image(os.path.join(self.root_dir, self.filenames[idx]))
+        sample = {'gt': image}
+        if self.transform:
+            sample = self.transform(sample)
+        return sample