TPOT one-line pipeline representation

[jni]

Just a quick (I hope) question. When I interrupt TPOT training, I get the following line output:

Best pipeline: _random_forest(ARG0, mul(41, mul(38, 5)), sub(3, 94))

But when I export the pipeline, I just get a simple RF with 500 estimators. Is this expected? What's the format of these one-line summaries?

[rhiever]

That's simply because the most trees allowed in a TPOT RF is 500. So when
the mathematical expression evaluates to >500, we clamp it to 500.

On Wednesday, March 2, 2016, Juan Nunez-Iglesias notifications@github.com
wrote:

Just a quick (I hope) question. When I interrupt TPOT training, I get the
following line output:

Best pipeline: _random_forest(ARG0, mul(41, mul(38, 5)), sub(3, 94))

But when I export the pipeline, I just get a simple RF with 500
estimators. Is this expected? What's the format of these one-line summaries?

—
Reply to this email directly or view it on GitHub
#101.

Randal S. Olson, Ph.D.
Postdoctoral Researcher, Institute for Biomedical Informatics
University of Pennsylvania

E-mail: rso@randalolson.com | Twitter: @randal_olson
https://twitter.com/randal_olson
http://www.randalolson.com

[jni]

Ah sure, so those expressions ((41 * (38 * 5), (3 - 94)?) evaluate to parameters to the sklearn RandomForestClassifier? Which parameters?

[rhiever]

That's right. The first one is the # of trees, the second one is
max_features: https://github.com/rhiever/tpot/blob/master/tpot/tpot.py#L439

On Wed, Mar 2, 2016 at 6:18 PM, Juan Nunez-Iglesias <
notifications@github.com> wrote:

Ah sure, so those expressions ((41 * (38 * 5), (3 - 94)?) evaluate to
parameters to the sklearn RandomForestClassifier? Which parameters?

—
Reply to this email directly or view it on GitHub
#101 (comment).

Randal S. Olson, Ph.D.
Postdoctoral Researcher, Institute for Biomedical Informatics
University of Pennsylvania

E-mail: rso@randalolson.com | Twitter: @randal_olson
https://twitter.com/randal_olson
http://www.randalolson.com

[jni]

Ah, thanks for the source pointer, that clarifies things a lot!

But LOL it's a little bit depressing that TPOT came back to me with "use a random forest with as many trees as you can afford and automatic feature selection". =D

[rhiever]

How long did you run it for? Throwing a random forest at it usually comes out on top early on, depending on the data. Random forests became an industry standard for a long time for a reason. :-)

[jni]

@rhiever in this case, I think it was like 30 generations (and the best hadn't changed for a big chunk of those).

My hope using TPOT was that it would find something that could match RF's accuracy but have better test-time speed. This result was with a toy dataset though, I'll keep fiddling. =)

[rhiever]

Gotcha. I'm currently doing a comprehensive analysis of TPOT on about 180 data sets. It'll be interesting to see what comes out of that.