Shakerato

1. Install Anaconda3, GPU driver, CUDA, cudnn2. Run 'cmd' as administrator3. conda create -n [myenv] python=3.6 anaconda4. activate [myenv]5. conda install theano pygpu(if you got error, try this command: "conda install -c conda-forge pygpu") 6. conda install numpy scipy mkl-service libpython 7. Create a file "C:\Users\[Username]\.theanorc"[global]floatX = float32device = cuda0 8. Test code on G..

Limited GPU Memory GPU usually has lesser device memory than host memoryThe latest high-end GPU (such as NVIDIA GPU P100)12–16 GB device memoryHost system memory256GBTrend for deep learning models is to have a “deeper and wider” architectureEspecially, RNN needs a lot of memory 1. First Solution: distributed Deep LearningSource: M. Cho et al., "PowerAI DDL", 2017PowerAI DDL provides a unified in..