Allocation of X exceeds 10% of system memory.

tensorflow/core/framework/allocator.cc:101] Allocation of X exceeds 10% of system memory. 

==> 將batch size 縮小

CUDA driver version is insufficient for CUDA runtime version

問題1: 程式只要 import keras 就會掛掉!!

解法==> conda update keras
                conda update tensorflow

問題2: UDA driver version is insufficient for CUDA runtime version

\Anaconda3\lib\site-packages\tensorflow\python\client\session.py in __init__(self, target, graph, config)
    674     try:
    675       # pylint: disable=protected-access
--> 676       self._session = tf_session.TF_NewSessionRef(self._graph._c_graph, opts)
    677       # pylint: enable=protected-access
    678     finally:

InternalError: cudaGetDevice() failed. Status: CUDA driver version is insufficient for CUDA runtime version

解法==> 就是要讓 cuda driver和 cudnn 匹配. 若是在Anaconda , 因為Anaconda 會去做軟體版本匹配的整合, 所以最簡單的作法就是直接執行  conda update --all


若tensorflow 要使用底層GPU架構,要安裝對應的軟體.

Keras --> Tensorflow --> cuDNN --> CUDA Driver  --> GPU graphic card

cuDNN [1]: The NVIDIA CUDA® Deep Neural Network library (cuDNN) is a GPU-accelerated library of primitives for deep neural networks. cuDNN provides highly tuned implementations for standard routines such as forward and backward convolution, pooling, normalization, and activation layers. cuDNN is part of the NVIDIA Deep Learning SDK.

CUDA Driver : Driver for NVIDIA graphic card

References:

[1] https://developer.nvidia.com/cudnn