Merge pull request #880 from Sentient07/non-deterministic

Added a warning note in lasagne/random.py

lasagne/random.py

@@ -3,6 +3,13 @@
 used for weight initialization and seeding noise layers.
 This can be replaced by a :class:`numpy.random.RandomState` instance with a
 particular seed to facilitate reproducibility.
+
+Note: When using cuDNN, the backward passes of convolutional and max-pooling
+layers will introduce additional nondeterminism (for performance reasons).
+For 2D convolutions, you can enforce a deterministic backward pass
+implementation via the Theano flags ``dnn.conv.algo_bwd_filter=deterministic``
+and ``dnn.conv.algo_bwd_data=deterministic``. Alternatively, you can disable
+cuDNN completely with ``dnn.enabled=False``.
 """
 
 import numpy as np