--- a
+++ b/dsb2018_topcoders/selim/pred_oof.py
@@ -0,0 +1,145 @@
+import os
+from params import args
+
+os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
+
+from keras.applications.imagenet_utils import preprocess_input
+from keras.preprocessing.image import img_to_array, load_img
+
+from models.model_factory import make_model
+
+
+from os import path, mkdir, listdir
+import numpy as np
+
+np.random.seed(1)
+import random
+
+random.seed(1)
+import tensorflow as tf
+
+tf.set_random_seed(1)
+import timeit
+import cv2
+import pandas as pd
+from tqdm import tqdm
+
+images_folder = args.images_dir
+train_pred = os.path.join(args.out_root_dir, args.out_masks_folder)
+
+df = pd.read_csv(args.folds_csv)
+
+all_ids = []
+all_images = []
+all_masks = []
+
+weights = [os.path.join(args.models_dir, m) for m in args.models]
+
+OUT_CHANNELS = args.out_channels
+
+def preprocess_inputs(x):
+    return preprocess_input(x, mode=args.preprocessing_function)
+
+if __name__ == '__main__':
+    t0 = timeit.default_timer()
+
+    fold_nums = [0,1,2,3]  # , 2, 3
+
+    if not path.isdir(train_pred):
+        mkdir(train_pred)
+
+    df = df.reset_index().copy()
+    all_ids = df['img_id'].values
+    all_sources = df['source'].values
+
+    for i in tqdm(range(len(all_ids))):
+        img_id = all_ids[i]
+        img = img_to_array(load_img(os.path.join(args.images_dir, img_id + ".png")))
+
+        all_images.append(img)
+
+    model = make_model(args.network, (None, None, 3))
+
+    for it in range(4):
+        if it not in fold_nums:
+            continue
+
+        val_idx = df[(df['fold'] == it)].index.values  # (df['hold_out'] == 0) &
+
+        model.load_weights(weights[it])
+        print('Predicting fold', it)
+        #        if it < 2:
+        #            continue
+        for i in tqdm(val_idx):
+            final_mask = None
+            for scale in range(1):
+                fid = all_ids[i]
+                img = all_images[i]
+                if final_mask is None:
+                    final_mask = np.zeros((img.shape[0], img.shape[1], OUT_CHANNELS))
+                if scale == 1:
+                    img = cv2.resize(img, None, fx=0.75, fy=0.75)
+                elif scale == 2:
+                    img = cv2.resize(img, None, fx=1.25, fy=1.25)
+                x0 = 16
+                y0 = 16
+                x1 = 16
+                y1 = 16
+                if (img.shape[1] % 32) != 0:
+                    x0 = int((32 - img.shape[1] % 32) / 2)
+                    x1 = (32 - img.shape[1] % 32) - x0
+                    x0 += 16
+                    x1 += 16
+                if (img.shape[0] % 32) != 0:
+                    y0 = int((32 - img.shape[0] % 32) / 2)
+                    y1 = (32 - img.shape[0] % 32) - y0
+                    y0 += 16
+                    y1 += 16
+                img0 = np.pad(img, ((y0, y1), (x0, x1), (0, 0)), 'symmetric')
+
+
+                inp0 = []
+                inp1 = []
+                for flip in range(2):
+                    for rot in range(4):
+                        if flip > 0:
+                            img = img0[::-1, ...]
+                        else:
+                            img = img0
+                        if rot % 2 == 0:
+                            inp0.append(np.rot90(img, k=rot))
+                        else:
+                            inp1.append(np.rot90(img, k=rot))
+
+                inp0 = np.asarray(inp0)
+                inp0 = preprocess_inputs(inp0)
+                inp1 = np.asarray(inp1)
+                inp1 = preprocess_inputs(inp1)
+
+                mask = np.zeros((img0.shape[0], img0.shape[1], OUT_CHANNELS))
+
+                pred0 = model.predict(inp0, batch_size=1)
+                pred1 = model.predict(inp1, batch_size=1)
+                j = -1
+                for flip in range(2):
+                    for rot in range(4):
+                        j += 1
+                        if rot % 2 == 0:
+                            pr = np.rot90(pred0[int(j / 2)], k=(4 - rot))
+                        else:
+                            pr = np.rot90(pred1[int(j / 2)], k=(4 - rot))
+                        if flip > 0:
+                            pr = pr[::-1, ...]
+                        mask += pr  # [..., :2]
+
+                mask /= 8
+                mask = mask[y0:mask.shape[0] - y1, x0:mask.shape[1] - x1, ...]
+                if scale > 0:
+                    mask = cv2.resize(mask, (final_mask.shape[1], final_mask.shape[0]))
+                final_mask += mask
+            final_mask /= 1
+            final_mask = final_mask * 255
+            final_mask = final_mask.astype('uint8')
+            if OUT_CHANNELS == 2:
+                final_mask = np.concatenate([final_mask, np.zeros_like(final_mask)[..., 0:1]], axis=-1)
+            cv2.imwrite(path.join(train_pred, '{0}.png'.format(fid)), final_mask, [cv2.IMWRITE_PNG_COMPRESSION, 9])
\ No newline at end of file