Alex Park

For L1 the IFM->OFM is 3->96, but in L2, it goes from 64->128, were these numbers selected to intentionally be mismatched for benchmarking purposes? Or should L2 be 96->128?

Could you give me some additional details of what you are trying to do? By input layers, do you mean the input to the network? Or do you just mean...

I'm not sure why this would be the case, but I can no longer replicate this behavior after upgrading to driver version 375.

did you create image batches for the dataset first? if not, then you will need to create them first. If you did, then it might be helpful to know the...

hmm... that is a strange one. could you try changing line 104 on `/usr/local/lib/python2.7/dist-packages/neon/data/dataloader.py` to instead return ``` return [self.be.iobuf(dim0=dim0, dtype=dtype, persist_values=False) for _ in range(2)] ```

hmm... have you been able to run any other neon examples (e.g. cifar_conv.py in the examples directory)? which gpu do you have and which version of pycuda? thanks, On Fri,...

ah ok -- it's a non-maxwell card. i guess there are still some issues for running dataloader dependent examples (cifar_msra) on kepler cards. Seems like the device buffer for storing...

as another data point, we have observed that driver versions up to 361.42 seem to allow this, but fail for newer versions

this thread seems to document the behavior https://devtalk.nvidia.com/default/topic/973477/cuda-programming-and-performance/-cuda8-0-bug-child-process-forked-after-cuinit-get-cuda_error_not_initialized-on-cuinit-/