The current version does not support Windows? about sonnet HOT 8 CLOSED

@shushan2017 We don't officially support Windows, although now that the PyPI packages don't contain any C++ (and neither does DNC) this should work. Could you post the error message you receive along with a full backtrace?

from sonnet.

@malcolmreynolds thanks!!!!!
My environment is： win7(64bit) Anaconda3 (64-bit) python3.6(64bit) 1080ti-card(CUDA8+CUDNN6) tf-gpu1.3
then i install dm-sonnet-gpu ,and run \sonnet\examples\rnn_shakespeare.py , It works on GPU
but run dm-DNC(https://github.com/deepmind/dnc----train.py) it is error：
An error ocurred while starting the kernel
2017???? 08:37:34.935876: W C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017???? 08:37:35.741301: I C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\common_runtime\gpu\gpu_device.cc:955] Found device 0 with properties:
name: GeForce GTX 1080 Ti
major: 6 minor: 1 memoryClockRate (GHz) 1.721
pciBusID 0000:04:00.0
Total memory: 11.00GiB
Free memory: 10.72GiB
2017???? 08:37:35.741301: I C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\common_runtime\gpu\gpu_device.cc:976] DMA: 0
2017???? 08:37:35.741301: I C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\common_runtime\gpu\gpu_device.cc:986] 0: Y
2017???? 08:37:35.741301: I C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\common_runtime\gpu\gpu_device.cc:1045] Creating TensorFlow device (/gpu:0) ?> (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:04:00.0)
2017???? 08:37:42.907917: E C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\stream_executor\cuda\cuda_event.cc:49] Error polling for event status: failed to query event: CUDA_ERROR_ILLEGAL_ADDRESS
2017???? 08:37:42.907917: F C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\common_runtime\gpu\gpu_event_mgr.cc:203] Unexpected Event status: 1
2017???? 08:39:55.881633: W C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017???? 08:39:56.474434: I C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\common_runtime\gpu\gpu_device.cc:955] Found device 0 with properties:
name: GeForce GTX 1080 Ti
major: 6 minor: 1 memoryClockRate (GHz) 1.721
pciBusID 0000:04:00.0
Total memory: 11.00GiB
Free memory: 10.72GiB
2017???? 08:39:56.474434: I C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\common_runtime\gpu\gpu_device.cc:976] DMA: 0
2017???? 08:39:56.474434: I C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\common_runtime\gpu\gpu_device.cc:986] 0: Y
2017???? 08:39:56.474434: I C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\common_runtime\gpu\gpu_device.cc:1045] Creating TensorFlow device (/gpu:0) ?> (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:04:00.0)
2017???? 08:40:02.511645: E C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\stream_executor\cuda\cuda_event.cc:49] Error polling for event status: failed to query event: CUDA_ERROR_ILLEGAL_ADDRESS
2017???? 08:40:02.511645: F C:\tf_jenkins\home\workspace\nightly?win\M\windows?gpu\PY\36\tensorflow\core\common_runtime\gpu\gpu_event_mgr.cc:203] Unexpected Event status: 1

from sonnet.

@shushan2017 Sorry for the delay - I think this is a TF issue rather than anything Sonnet-specific. This issue looks related: tensorflow/tensorflow#3224 - can you see if the various suggestions there work for you?

from sonnet.

@malcolmreynolds Oh, it doesn't matter. Thank you very much for your attention and answer. I'll go and see.

from sonnet.

After the installation of tf1.5gpu, the operation is normal, it seems to be the problem of TF, but sonnet seems to only need tf1.3 version, I forced the installation of version 1.5.
But the problem seems to come out, the code runs very slowly, 10 times slower than CPU, and the calculations don't seem quite right.

from sonnet.

@shushan2017 as far as I'm aware the most up to date version of TF is 1.4.0rc0 - I'm not aware of any 1.5 branch. You're correct that Sonnet requests the exact version 1.3.0 - we aim to track whatever version of TF is installed by pip install tensorflow{-gpu}. When TF 1.4.0 hits stable (ie, there is a non-release-candidate tagged version) we will update Sonnet to make that the supported version. As you've found, if you want to use the current version of Sonnet with a different version of TF you will have to hack things, and the result is unsupported.

It doesn't surprise me too much that the GPU version runs slowly - the DNC architecture is not particularly well suited to GPU, having lots of smaller Tensors that need to be operated on as opposed to larger matrix multiples / convolutions. During development of NTM / DNC we never used GPUs as there seemed to be no benefit to them. A 10x slowdown wouldn't particularly surprise me.

That said, I would expect the answers produced to be the same. When you say calculations "don't quite seem right", can you be more specific?

from sonnet.

@malcolmreynolds
Thank you for your answer. I'm concerned about GPU. I want to get higher speed. If DNC isn't suitable for GPU, then I don't bother with it.
About the calculation errors, I just feel sorry, 0.4 is probably remember before the calculation of ACC, and GPU, into 0.003, because the calculation is too slow, I run only once, this information can not explain what. The conclusion is that DNC is really not suitable for running on gpu.
Again, thank you very much！！！！！

from sonnet.

@shushan2017 all the results for the Nature paper were done with a Hogwild SGD setup across multiple CPUs (32 cores in some cases) - I'd recommend looking into that for the best performance.

from sonnet.