At the end of the installation KERAS,TENSORFLOW,PYTORCH,OPENCV

This article is void

My next installment is the TensorFlow and Keras truth.

Environment:

Anaconda4.2;python3.5;windows10,64,cuda

Previous hard cuda9.1 useless, we want to use the GPU must choose cuda8.0, I thought the official will be corresponding update, naive. First TensorFlow don't recognize, moreover cudnn own all do not recognize, only 8.0.

Keras and TensorFlow are both Pip,pytorch and OpenCV are going to find WHL. About Keras backend and installation tensorflow, I do not know which link out of the problem, the other import in Pycharm are no problem, only import TensorFlow will appear: Process finished with exit Code " -1073741819 (0xc0000005)", in Cmd Python, is also the appearance of Python can not run, direct crash exit.

Read a lot of blog, some say is pycharm problem, some say is memory leak, some say is the problem of windows, some say is the Python version problem, and so on.

I may have made a mistake at the time:

Go directly to the environment variable new keras_backend, the value is TensorFlow. But I hesitate, this is not a path, and whether to add English quotes. have tried, or error, but in the search for Keras.josn file, found that backend is already tensorflow (that the environment variable Keras deleted), is the previous link is right (or Keras default is TensorFlow). Because in the repeated unloading of a reload tensorflow when Keras, no TensorFlow, Keras will only report no backend error, and TensorFlow, he will be like TensorFlow Import panic '. Later, the import of other modules can be used normally, obviously with the external hardware does not matter, then the TensorFlow uninstall clean, not only pip uninstall, there are tensorflow-board and other documents, deleted (emphasis one).

Because remember that before installing Keras, you need Conda to install Libpython and MinGW (a book proposal), and then I will upgrade my python3.5.2 and some related packages to 3.5.4, and I'll take it after I'm done. Sure enough to install TensorFlow played a bad role. Well, just update all the packages and upgrade to 3.5.4 (emphasis II).

Combine emphasis one and focus two, perfect.