Recently I started playing Tensorflow on my mac, it very fun. But I found there are some traps about Tensorflow on Mac and AWS, so tell you know.
- Tensorflow is very slow.
- Now Tensorflow doesn’t support CUDA on Mac officially, but you can easy find a way to compile it to support CUDA. But almost all new Mac don’t contain a NVIDIA display card, so you still can’t use CUDA. OpenCL support is on the way, open to contributions, but how knows.
- The HPC of AWS ( g2.2xlarge
and g2.8xlarge ) contain NVIDIA display card, but the card is blow the requirement of default CUDA setting of Tensorflow, so you can’t user precompiled Tensorflow, you need compile it by yourself, and set compute capability is 3.0. And there is a AMI (Amazon Linux x64 AMI with TensorFlow (GPU)), already include Tensorflow on AWA’ AMI market, but this AMI is using precompiled Tensorflow, so it is useless, you can’t use the GPU. And even after your compile and install you own Tensorflow, there is a bug of NVIDIA display driver card, you must update NVIDIA display card driver, and finally you can use GPU from the HPC.