Using CNTK with Keras

This page has migrated to our new site. Please update any bookmarks.

Using CNTK with Keras (Beta)

This page is under development

Steps to follow

We assume you have followed the Anaconda installed on your Windows or Linux machines.

1. We highly recommend creating a new anaconda environment.

conda create --name cntkkeraspy35 python=3.5 numpy scipy h5py jupyter

2. Activate the new environment

activate cntkkeraspy35

3. Install Keras

4. Install a GPU build of CNTK

pip install <URL to CNTK GPU wheel>

4. Update Keras to use CNTK as backend

On Windows:

SET KERAS_BACKEND=cntk

On Linux:

export KERAS_BACKEND=CNTK

change “backend” value to “cntk” ensure “image_dim_ordering” is set to “tf”

If you don’t have a keras.json, it means you have not run keras on the machine. Run Python and import keras and it will create the file. 7. Unzip the Keras examples to your machine unzip \stcvm-ls426\share\cntk\keras_examples.zip -d c:\local\keras_examples (or whatever unzip mechanism you’d like to use) 8. Try out the Keras examples python c:\local\keras_examples\examples\addition_rnn.py

9. Test your own scripts!

10. Send your feedback to cntkkeras

Known issues: • The following Keras APIs are only supported on GPU: normalize_batch_in_training, batch_normalization • The following Keras APIs are not yet supported: clear_session, cast, resize_images, resize_volumes, print_tensor, softplus, hard_sigmod, ctc_batch_cost, ctc_decode, map_fn, foldl, foldr • The following examples do not work yet: mnist_swwae.py, neural_doodle.py, neural_style_transfer.py, pretrained_word_embeddings.py • The following Keras layers are not supported yet: convolution with dilation rate, cropping, noise, merge dot • CPU performance needs to be improved • Models cannot be saved to HDFS