CUDA out of memory issue

[danieleghisi]

Hi, I'm having an "out of memory" issue while running the demo.

Snippet: I tensorflow/core/common_runtime/gpu/gpu_device.cc:885] Found device 0 with properties: name: GeForce GTX TITAN X major: 5 minor: 2 memoryClockRate (GHz) 1.076 pciBusID 0000:04:00.0 Total memory: 11.91GiB Free memory: 11.67GiB (full log below)

I have tried to lower the model parameters, but nothing seems to work. Do you have any advice? Why does the demo take so much GPU memory? Thanks a lot, Daniele

Full log: python demo.py I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcublas.so locally I tensorflow/stream_executor/dso_loader.cc:119] Couldn't open CUDA library libcudnn.so. LD_LIBRARY_PATH: I tensorflow/stream_executor/cuda/cuda_dnn.cc:3459] Unable to load cuDNN DSO I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcufft.so locally I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcuda.so.1 locally I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcurand.so locally I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero I tensorflow/core/common_runtime/gpu/gpu_device.cc:885] Found device 0 with properties: name: GeForce GTX TITAN X major: 5 minor: 2 memoryClockRate (GHz) 1.076 pciBusID 0000:04:00.0 Total memory: 11.91GiB Free memory: 11.67GiB W tensorflow/stream_executor/cuda/cuda_driver.cc:590] creating context when one is currently active; existing: 0x48c4140 I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero I tensorflow/core/common_runtime/gpu/gpu_device.cc:885] Found device 1 with properties: name: GeForce GTX 980 major: 5 minor: 2 memoryClockRate (GHz) 1.342 pciBusID 0000:0a:00.0 Total memory: 3.94GiB Free memory: 487.88MiB W tensorflow/stream_executor/cuda/cuda_driver.cc:590] creating context when one is currently active; existing: 0x48c0320 E tensorflow/core/common_runtime/direct_session.cc:135] Internal: failed initializing StreamExecutor for CUDA device ordinal 2: Internal: failed call to cuDevicePrimaryCtxRetain: CUDA_ERROR_OUT_OF_MEMORY; total memory reported: 18446744073648275456 Traceback (most recent call last): File "demo.py", line 16, in gpu_fraction=gpu_fraction) File "/home/daniele/fast-wavenet-master/wavenet/models.py", line 54, in init sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) File "/home/daniele/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1186, in init super(Session, self).init(target, graph, config=config) File "/home/daniele/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 551, in init self._session = tf_session.TF_NewDeprecatedSession(opts, status) File "/usr/lib/python2.7/contextlib.py", line 24, in exit self.gen.next() File "/home/daniele/.local/lib/python2.7/site-packages/tensorflow/python/framework/errors_impl.py", line 469, in raise_exception_on_not_ok_status pywrap_tensorflow.TF_GetCode(status)) tensorflow.python.framework.errors_impl.InternalError: Failed to create session.

[ianni67]

I'm facing the same error. Titan X, 11.92GB memory available. Looks like TF is trying to get as much memory as the card exposes, thus being never satisfied. There must be some wrong configuration somewhere.

[tomlepaine]

Hi @danieleghisi @ianni67 I have definitely run the demo with less memory. 6 GB I think. Did you have any luck?

[danieleghisi]

Not really, still same issue for me...

----- Messaggio originale ----- Da: "Tom Le Paine" [email protected] Inviato: ‎01/‎02/‎2017 01:53 A: "tomlepaine/fast-wavenet" [email protected] Cc: "danieleghisi" [email protected]; "Mention" [email protected] Oggetto: Re: [tomlepaine/fast-wavenet] CUDA out of memory issue (#10)

Hi @danieleghisi @ianni67 I have definitely run the demo with less memory. 6 GB I think. Did you have any luck? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[ianni67]

I'm very sorry for my late reply. I eventually solved the issue, and, alas, (or luckily?) it was my own fault. I did not manage the GPUs correctly, so two processes where contending the same GPU, both requesting all its memory. Confining the processes on separate GPUs (by using CUDA_VISIBLE_DEVICES appropriately) solved the problem. Sorry for the wrong issue and the late reply!

[tomlepaine]

@ianni67, I'm glad you resolved your issue. @danieleghisi do you think you have the same problem?

[danieleghisi]

Sorry for the late reply. CUDA_VISIBLE_DEVICES solves the memory issues, but now I get a CuDNN version error:

E tensorflow/stream_executor/cuda/cuda_dnn.cc:378] Loaded runtime CuDNN library: 5005 (compatibility version 5000) but source was compiled with 5105 (compatibility version 5100). If using a binary install, upgrade your CuDNN library to match. If building from sources, make sure the library loaded at runtime matches a compatible version specified during compile configuration.

Looks like an issue in tensorflow installation (I can't update CuDNN, due to other dependencies.) I'll try to google how to solve this...

Daniele

[danieleghisi]

@tomlepaine Thanks Tom, I've managed to update CuDNN and dependencies, and the training works fine! On the other hand, the generation is always a constant value, the seed (perfectly flat waveform)...

May I be doing something wrong? I've just added a scipy.io.wavfile.write("out.wav", 44100, numpy.array(predictions[0])) as very last line, to save the output file...

Thanks again for your support, Daniele

[tomlepaine]

@danieleghisi why are you saving predictions[0]? What is the shape of predictions?

Glad you are closer to getting the code working!

[danieleghisi]

Hi Tom, the shape of predictions is (1, 32000) However don't worry, I've just noticed that the tensorflow-wavenet implementation uses this fast generation; I'll stick to that model for now! Thanks for your support, d

[ishandutta2007]

@danieleghisi if you have succeeded in training it now, can you please share the model.

[francois-baptiste]

Setting "gpu_fraction = .95" instead of 1 in the demo.py solve the problem for me!

[ishandutta2007]

@francois-baptiste can you share the model.