boosted_trees_classifier.create early_stopping_rounds stops on first iteration in 2.0.1

User 2568 | 7/17/2016, 8:05:15 AM

If I use the example in from boostedtreesclassifier.create, and add metric="logloss", maxiterations=50, earlystoppingrounds=5

url = 'https://static.turi.com/datasets/xgboost/mushroom.csv'
data = graphlab.SFrame.read_csv(url)
train, test = data.random_split(0.8)
model = graphlab.boosted_trees_classifier.create(train, target='label', metric="log_loss", 
							max_iterations=50, early_stopping_rounds=5)

the algorithm stops after the first iteration. If I remove the earlystoppingrounds, then it goes to 11, with the Validation-log_loss continuiously dropping. I found the same behaviour with a model I'm working on.

This looks like earlystoppingrounds no longer works. When earlystoppingrounds, the prior version would have gone to 50 iterations, so I'm not sure why it now stops at 11

Comments

User 91 | 7/18/2016, 6:21:34 PM

We are able to reproduce this issue. We will look into it and give you an update soon. Thanks for pointing it out!


User 91 | 7/20/2016, 4:56:06 AM

We were able to identify and fix the issue. The next release (coming soon) will have the fix.


User 2568 | 7/20/2016, 5:20:07 AM

Great. In the meantime I'd like to install 1.10.1 in another anaconda environment. I tired this but seemed to get 2.01 instead

conda create -n GraphLab1.10.1 python=2.7 anaconda
source activate GraphLab1.10.1

pip install --no-cache-dir https://get.graphlab.com/GraphLab-Create/1.10.1/xxxx/xxxx/GraphLab-Create-License.tar.gz

The output shows thsi is is installing 2.0.1

 Downloading https://get.graphlab.com/GraphLab-Create/1.10.1/kevin.mcisaac@gmail.com/7216-F60A-6BD4-96B9-87F5-1D10-B01A-9AAF/GraphLab-Create-License.tar.gz
Collecting graphlab-create>=1.10.1 (from GraphLab-Create-License==1.10.1)
  Downloading GraphLab_Create-2.0.1-py2.7.tar.gz (51.2MB)
100% |████████████████████████████████| 51.2MB 82.4MB/s