My personal blog http://blog.omeryavuz.gen.tr
Incredible, last attack from Tensorflow. They can not be successful on symbolic conversation. They try to accomplish physical limits. It seems that we won't see Tensorflow in a few year later. Their last attack mitigate fundamental problems?
Yes, intuitive user interface and easy coding with Keras is possible.
But,
- there is no efficient symbolic structure it must be over F#, Lisp, but python is questionable to accomplish expected quality.
- there is no very good background and culture on Tensorflow block, they tested on some systems and shared restricted content for commercial purposes.
- their GPU resource usage are too limited, if you need multi GPU over 4 GPU you must consider patches from different vendors because of performance bottleneck.
- performance are not acceptable for heavy users.
- framework designed for commercial purposes, lack of freedom does not acceptable for serious investors.
- Tensorflow backend is not feasible on low and mid level Nvidia GPU machines i.e. 1070, 1070TI, 1060, 1080, their testing area is TPU, V100 etc. and they support only big cloud system hardware stack.
