I'm trying to extend the deeplearning1 notebook
mnist.ipynb and am running into some bugs. On the line
lm.fit_generator(batches, batches.N, nb_epoch=1, validation_data=test_batches, nb_val_samples=test_batches.N)
I first had the error message
AttributeError: 'NumpyArrayIterator' object has no attribute 'N'
which is raised on this thread, with a solution but not an explanation. I changed
.n as suggested there and then had an exception on the same line:
error: identifier "cudnnSetFilterNdDescriptor_v4" is undefined
which seems to be a problem with incompatibility between different versions of the libraries, discussed on github.
I've already been warned:
Using gpu device 0: Tesla K80 (CNMeM is enabled with initial size: 95.0% of memory, cuDNN 6021)
Your cuDNN version is more recent than the one Theano officially supports.
If you see any problems, try updating Theano or downgrading cuDNN to version 5.1.
I'm wary of fooling around trying to change library versions when I don't fully understand the crestle environment, and I don't know the command to issue to downgrade cuDNN or to upgrade theano. @anurag, can you suggest a good way to work around this? Thanks!