FIX: decode_csv did not place features in a single tensor.

- this broke with tf 1.5, where we needed to use tf.stack on the list of
  features. Probably, this stacking before was done internally in dataset_ops.
  Now, they have changed it to treating it as a tuple which breaks things
  lateron.

Author: FrederikHeber

src/DataDrivenSampler/common.py

@@ -349,7 +349,7 @@ def decode_csv_line(line, defaults):
     items = tf.decode_csv(line, list(defaults.values()))
 
     # reshape into proper tensors
-    features = items[0:-1]
+    features = tf.stack(items[0:-1])
     label = tf.reshape(tf.convert_to_tensor(items[-1], dtype=tf.int32), [1])
 
     # return last element as label, rest as features