Segmentation fault with generators

[tristandeleu]

Ever since I switched to general generators for training data, I get Segmentation Fault errors every so often when I am training a model. It can equally happen either when the training stops because it reached the max_iter number of iterations or if I interrupt the training with Ctrl+C.

...
Batch #50000: 0.000007
Batch #50500: 0.000006
^C
In [1]: [1]    32012 segmentation fault  THEANO_FLAGS='mode=FAST_RUN' ipython examples/repeat-copy.py -i