Local install (Windows only:cpu)
- Git (preferably from cygwin installer)
Step 1 - Install Theano, Anaconda, Keras
- Follow GPU-accelerated Keras with Tensorflow or Theano on Windows 10 native which work well as of May 2017 (commit 3d0acee).
- Use Anaconda2, Python 2.7 (not v3)
- The above instructions work for both CPU and GPU configuration.
- If planning to use CPU only, ignore the CUDA installs and set the THEANO_FLAGS environment variable to that of THEANO_FLAGS_CPU
- NOTE Do not install Keras 2.x.x as per the instructions but instead use the same version as the course AWS install:
pip install keras==1.2.2
Step 2 - Fast.ai refinements
- From Anaconda Prompt (bcolz is required for saving & loading arrays, ffmpeg for sgd-intro.ipynb):
conda install bcolz conda install -c menpo ffmpeg
Execute a simple test to make sure theano and keras have been installed, from a python cli:
import theano, keras
The Keras config (%USERPROFILE%\.keras\keras.json) will default to using Tensor Flow. These lines:
"image_dim_ordering": "tf" "backend": "tensorflow"
should be changed to use Theano:
"image_dim_ordering": "th" "backend": "theano"
Step 3 - Execute Lesson 1 notebook
- You should now be able to execute lesson 1 after starting up Jupyter Notebook from the start menu.
- The Jupyter Notebook shortcut may be modified to change the starting location for Jupyter
Other historical sources: