每次尝试使用gaussian_noise_injection_std_dev> 0的学习者时,我都会遇到此问题。 根据GitHub网站上第444期提供的答案,解决方案是在CPU上进行初始化,但我无法在Python API中找到此选项。
CURAND失败105 :(参见curand.h&寻找curandStatus或CURAND_STATUS_xxx); GPU = 0; hostname = ROSIE; expr = curandGenerateNormal(((curandGenerator_t *)s_curandGenerator)[0],reinterpret_cast(Data()),GetNumElements(),(float)mean,(float)sigma) Traceback(最近一次调用最后一次): 文件" C:\ local \ Anaconda3-4.1.1-Windows-x86_64 \ envs \ cntk-py35 \ lib \ site-packages \ IPython \ core \ interactiveshell.py",第2881行,在run_code中 exec(code_obj,self.user_global_ns,self.user_ns) 文件"",第5行,in trainer.train_minibatch(data)#news model with it 文件" C:\ local \ Anaconda3-4.1.1-Windows-x86_64 \ envs \ cntk-py35 \ lib \ site-packages \ cntk \ trainer.py",第120行,在train_minibatch中 参数,设备) 文件" C:\ local \ Anaconda3-4.1.1-Windows-x86_64 \ envs \ cntk-py35 \ lib \ site-packages \ cntk \ cntk_py.py",第2022行,在train_minibatch_overload_for_minibatchdata中 return _cntk_py.Trainer_train_minibatch_overload_for_minibatchdata(self,args) RuntimeError:CURAND失败105 :(参见curand.h&查找curandStatus或CURAND_STATUS_xxx); GPU = 0; hostname = ROSIE; expr = curandGenerateNormal(((curandGenerator_t)s_curandGenerator)[0],reinterpret_cast(Data()),GetNumElements(),(float)mean,(float)sigma) [CALL STACK]
Microsoft :: MSR :: CNTK :: CudaTimer :: Stop - Microsoft :: MSR :: CNTK :: GPUMatrix :: SetGaussianRandomValue - Microsoft :: MSR :: CNTK :: Matrix :: SetGaussianRandomValue - CNTK :: Internal :: ZeroesWithDynamicAxesLike - CNTK :: Internal :: ZeroesWithDynamicAxesLike - CNTK :: SGDLearner - CNTK :: TrainingParameterSchedule :: Serialize - CNTK :: Trainer :: TotalNumberOfSamplesSeen - CNTK :: Trainer :: TrainMinibatch - CNTK :: Trainer :: TrainMinibatch - PyInit__cntk_py - PyInit__cntk_py - PyCFunction_Call - PyEval_GetFuncDesc - PyEval_EvalFrameEx - PyEval_EvalFrameEx 我有GTX 970,4GB,在Win10上运行beta 12。
此致 Slawek
答案 0 :(得分:0)
对此的修复现在是主人。 CNTK 2.0rc2将正常工作。