make dataset split value more robust

Author: Clément Pinard

datasets/util.py

@@ -6,10 +6,15 @@ def split2list(images, split, default_split=0.9):
         with open(split) as f:
             split_values = [x.strip() == '1' for x in f.readlines()]
         assert(len(images) == len(split_values))
-    elif isinstance(split, float):
-        split_values = np.random.uniform(0,1,len(images)) < split
-    else:
+    elif split is None:
         split_values = np.random.uniform(0,1,len(images)) < default_split
+    else:
+        try:
+            split = float(split)
+        except TypeError:
+            print("Invalid Split value, it must be either a filepath or a float")
+            raise
+        split_values = np.random.uniform(0,1,len(images)) < split
     train_samples = [sample for sample, split in zip(images, split_values) if split]
     test_samples = [sample for sample, split in zip(images, split_values) if not split]
     return train_samples, test_samples

main.py

@@ -33,7 +33,8 @@
 group.add_argument('-s', '--split-file', default=None, type=str,
                    help='test-val split file')
 group.add_argument('--split-value', default=0.8, type=float,
-                   help='test-val split proportion (between 0 (only test) and 1 (only train))')
+                   help='test-val split proportion between 0 (only test) and 1 (only train), '
+                        'will be overwritten if a split file is set')
 parser.add_argument('--arch', '-a', metavar='ARCH', default='flownets',
                     choices=model_names,
                     help='model architecture, overwritten if pretrained is specified: ' +