Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
Pramit Choudhary
@pramitchoudhary
Welcome Brett
Brett Olmstead
@Brett777
Hey guys
Am interested in including some Skater in a demo
Some of the demos I show use H2O models
What are the currently supported libraries?
Just SKLearn?
Pramit Choudhary
@pramitchoudhary
We had started work on supporting h20 but couldn't complete it. In-terms of supporting libraries, we have examples for python/R based models/third party vendors(indico.ai, algorithmia), deep-learning framework like keras
It would be great, if you could take a shot at H20 models and see how we can change improve the library to enable a more native support. Looking fwd to collaborating
Pramit Choudhary
@pramitchoudhary
This year NIPS had a lot of traction in-regards to Model Interpretability
Here is the keynote from Kate Crawford: https://tinyurl.com/yax2v2vr
Pramit Choudhary
@pramitchoudhary
Welcome @dipanjanS
Pramit Choudhary
@pramitchoudhary
@aikramer2 will it be possible to check the permission on travis once? I am not able to add a branch for build
Pramit Choudhary
@pramitchoudhary
@aikramer2 was able to get rid of the exception and get the builds back up. Thanks for pointing out the fix
Ben Van Dyke
@benvandyke
Hey @pramitchoudhary @aikramer2 Partial dependence plots helped me out this morning. Thanks!
Ben Van Dyke
@benvandyke
Any thoughts on getting spark compatibility? Maybe a similar treatment to the deployed api?
Pramit Choudhary
@pramitchoudhary
Hey @benvandyke, glad it was helpful. We had thought about it, we definitely need to in-corporate Spark compatibility. Just added a new ticket: datascienceinc/Skater#203. Possible to add some more details, you think we should consider.
Pramit Choudhary
@pramitchoudhary
@benvandyke will you have the bandwidth to explore the possibility of taking up this feature ?
Ben Van Dyke
@benvandyke
@pramitchoudhary I'll think about it some more, going to write the spark partial dependence manually and perhaps that can become the base for an actual feature
Pramit Choudhary
@pramitchoudhary
@benvandyke thats a great idea :), that will enable us to make it more performant with robust scalability. Lets capture our thoughts in the ticket datascienceinc/Skater#203 as we go along. Thanks for look into it
datadave
@datadave
FYI Pramit and I recently watched https://www.ted.com/talks/cathy_o_neil_the_era_of_blind_faith_in_big_data_must_end/transcript and Pramit and Cathy have a skype at some point in the near future to get her input on model interpretation and skater.
Pramit Choudhary
@pramitchoudhary
Thanks @datadave. Her book titled “Weapons of Math Destruction” is pretty nice. I have been reading that. Would recommend to others as well. https://weaponsofmathdestructionbook.com/author/mathbabe/
Pramit Choudhary
@pramitchoudhary
Hi Everyone,
hopefully, soon we will be adding the support for building interpretable models as well within Skater. Will keep you guys updated on the progress
Pramit Choudhary
@pramitchoudhary
@Brett777 where you able to make any progress with the h20 models?
Pramit Choudhary
@pramitchoudhary
@/all
Welcoming @christophM to the group. If i have to describe him in a word, I would say he is "awesome"
Checkout his R implementation on the idea of MI here https://github.com/christophM/iml
Pramit Choudhary
@pramitchoudhary
Pramit Choudhary
@pramitchoudhary
@/all Hey guys, we have a new beta release of Skater with the ability to build interpretable models using rules extraction(Bayesian Rule List).
Steps to get started:
Quick summary to install conda and setup the python environment(recommended steps for using python3.x)

1. wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh
2. bash miniconda.sh -b -p $HOME/miniconda
3. export PATH="$HOME/miniconda/bin:$PATH"
4. conda config --set always_yes yes --set changeps1 no
5. conda info -a

Installation:
Option1:
1. conda install gxx_linux-64
2. git clone the repo
3. sudo python setup.py install --ostype=linux-ubuntu --rl=True

Option2:
1. conda install gxx_linux-64
2. sudo pip install -U --no-deps --force-reinstall --install-option="--rl=True" skater==1.1.0b1
The docs are pending an update. Should be able to get to it over the weekend.
Pramit Choudhary
@pramitchoudhary
More exciting stuff to follow
Pramit Choudhary
@pramitchoudhary
Image and text interpretability coming soon to Skater
Aaron Kramer
@aikramer2
Hey @pramitchoudhary , we ought to push 1.0.3 to pypi (git checkout tags/v1.0.3 ought to take you there). There are some key fixes in that version that would be good to publish as the current release.
Pramit Choudhary
@pramitchoudhary
@aikramer2 all the fixes from v1.0.3 should be in the new release right? unless I am missing something
checking versions on pypi
Pramit Choudhary
@pramitchoudhary
@aikramer2; just checked v1.0.3 is still there on pypi. Is pip install skater==1.0.3 not working?
Pramit Choudhary
@pramitchoudhary
I think the above issue with v1.0.3 is resolved with the new uploaded version v1.0.4. Currently work on improving the docs and moving v1.1.0-b1 to v1.1.0
Pramit Choudhary
@pramitchoudhary

Hi Guys, move v1.1.0b1 to v1.1.0. This release also has some document improvement with an explicit section on Jupyter Notebooks(https://datascienceinc.github.io/Skater/gallery.html).
This might be helpful in finding examples describing the use of different algorithms. The concept of rule list is still experimental as more work needs to be done there and hopefully we will get there.

This release will be followed up very soon, with another release supporting interpretability in-regards to DNNs and further documentation improvements. Feel free to reach out for any suggestions or improvements

Pramit Choudhary
@pramitchoudhary
Hi Guys, excited to announce native support for inferring DNNs with the beta release of skater-1.1.1b2. Have also updated the gallery with new examples
https://datascienceinc.github.io/Skater/gallery.html
sudo pip install -U --force-reinstall skater==1.1.1b2
Romit Singhai
@Romits
I am getting following error for build_visual_explainer
I am getting following error for build_visual_explainer, encoding' is an invalid keyword argument for this function. I am using skater 1.1.1b2
Pramit Choudhary
@pramitchoudhary
Romit Singhai
@Romits
Hello @pramitchoudhary I just looked at the plotting function.
Pramit Choudhary
@pramitchoudhary
oh ok, so is it working now @Romits
or still throwing an error. I had checked the example just a while back and it had worked fine
I will update you on the other request you had mentioned about
Romit Singhai
@Romits
@pramitchoudhary it is working for me too. Fixed the issue.
Pramit Choudhary
@pramitchoudhary
thats great to hear @Romits
Pramit Choudhary
@pramitchoudhary
@here starting a new channel related to ML experiments(MLLab). Here is video of Skater applied on a simple autonmous driving example. Hope you guys like it
https://youtu.be/dmXF7WxT63E
Pramit Choudhary
@pramitchoudhary
Hi Guys, just started working on adding support for interpreting tree. More updates coming soon