This is cool! RapidMiner releases a Keras Deep Learning extension!
Anyone download and play with it yet? I got to play with the earlier version and it's freaking cool!
Answers
-
I just tried it and got the following error: the script could not be parsed. No module named keras.callbacks
I was trying to run the s&p 500 regression example. When I test the keras extension in preferences, I get:
Keras is not installed.
None of the backends CNTK, TensorFlow or Theano are installed
Graphviz is not installed
Pydot is not installed
1 -
So far, so good, with a few hiccups! Some reflections.
Running RapidMiner 7.6 / Ubuntu 16.04, Keras with Tensorflow as the back end
General comments:
- Great job so far
- All Python processes always properly close regardless of the RM process end status
- GPU memory has no leaks and is always freed when the RM process stops
- There is still a large part of Keras not implemented
- It would be great to access history.history returned from model.fit, e.g. for plotting
- Is there any ability to backup the model?
- Is there any possibility to load the model?
I have started with the most complex example and then went down to the simplest. In all cases, I have added the Performance operator at the end and in s&p 500 I have also added Filter Example to remove missing values generated in the prediction.
Example: s&p 500 (regression)
Comments:
- Some predictions at the end are generated with missing values (31 out of 249), possibly due to batch size?
- Predicted value sometimes is returned as “Nominal” rather than of type of the label.
- Apply Keras Model sometimes produces an example data set (data frame) but it does not get passed to the next operator, which complains that “Apply Keras Model did not produce any output”.
- It seems that placing a break on Keras Model, Apply Keras Model or even the operator receiving the lebelled examples from Apply Keras Model ALWAYS results in generating a “Nominal” prediction and the process failing half way through.
Example: Iris (classification)
Comments:
- Worked well (with and without break points)
Example: Boston housing prices (regression)
Comments:
- Worked well (with and without break points)
You have beatten me to this extension as I've been working through a similar interface to Keras but stopped at Python interfaces Note that my earlier comments about cleaning Python processes and GPU memory is something that I really struggled with, so this is a really clean job!
Jacob
2 -
@lbookman You need to first install Python / Anaconda and then Tensorflow (or Theano or CNTK) and Keras on top, in the process you may need to install quite a few different libraries as well, depending on if you are running it on Windows or Linux. Graphviz and Pydot are needed to plot the Keras model (the Graph option for the Keras model). RapidMiner will not install any of these for you, I am afraid.
Jacob
0 -
I tried to install it using Anaconda and Windows and didn't have success. When I was testing it internally a month ago I had to install it on Windows without using Anaconda and got it to work. I think I'm going to try it on a Ubuntu VM and see how it goes.
0 -
Assuming you've followed the Anaconda install instructions there's a piece in the instructions that's a little unclear.
When you put the path to your new Keras enabled executable you must ALSO use the same path in your python script settings.0 -
Well, this is a great news!!
(To RapidMiner team) Hopefully in the future they will implement extensions for TensorFlow, Theano, Statsmodel, and Scikit-Learn. It would also be nice to build extensions for R statistical packages.
0