To @ChenglongChen , can you try to add seed to the python param to see if it fixed the problem? Seems the python module did not call seed during initialization as main console version did.

from xgboost.

Ok. Will try it soon. Thanks.

ÔÚ 2014Äê5ÔÂ21ÈÕ£¬0:14£¬Tianqi Chen [email protected] дµÀ£º

To @ChenglongChen , can you try to add seed to the python param to see if it fixed the problem? Seems the python module did not call seed during initialization as main console version did.

¡ª
Reply to this email directly or view it on GitHub.

from xgboost.

It seems there is actually a "bug" in my code, as I random shuffle the data each run. I remove that part and test with subsample<1, the results can be reproducible without explicitly setting the seed param. Very sorry :(

from xgboost.

No worries, glad that you find the problem :) Thanks for using XGboost

from xgboost.

Thanks for reporting the problem. Indeed the python module's seed via
set_param was not supported in previous code. I have made a fix to the
code, the most recent version of the code in repo should be able to
reproduce the result.

Tianqi

On Sat, May 24, 2014 at 11:25 AM, André Panisson
[email protected]:

I'm trying to reproduce a test with no success.
The following code, using the Kaggle Higgs dataset, should give the same
result for all steps, but I have different values in each step:

import numpy as npimport xgboost as xgbfrom sklearn import metrics
dtrain = np.loadtxt( open('../data/training.csv'), delimiter=',', skiprows=1,
converters={32: lambda x:int(x=='s'.encode('utf-8')) } )
X = dtrain[:1000,1:31]label = dtrain[:1000,32]weight = dtrain[:1000,31]

Create parametersparam = {}param['objective'] = 'binary:logitraw'param['eval_metric'] = 'auc'param['nthread'] = 1param['seed'] = 42num_round = 120

Create train and test setstrain_indices = np.arange(len(label)_0.8).astype(int)test_indices = np.arange(len(label)_0.8, len(label)).astype(int)

X_train, X_test = X[train_indices], X[test_indices]y_train, y_test = label[train_indices], label[test_indices]w_train, w_test = weight[train_indices], weight[test_indices]
xgmat = xgb.DMatrix(X_train, label=y_train, missing=-999.0, weight=w_train)xgmat_test = xgb.DMatrix(X_test, missing=-999.0)
for i in range(10):
bst = xgb.train(param.items(), xgmat, num_round)
y_out = bst.predict(xgmat_test)
print metrics.roc_auc_score(y_test, y_out)

I've set the seed parameter, but it does not help. Setting the python
random seed also does not help.

—
Reply to this email directly or view it on GitHubhttps://github.com/tqchen/xgboost/issues/9#issuecomment-44095344
.

Sincerely,

Tianqi Chen
Computer Science & Engineering, University of Washington

from xgboost.

@panisson it seems I mistakenly delete your previous comment. Really sorry about that.

from xgboost.

Thanks to @panisson , the python module should now support seeding and thus reproducible boosting results

from xgboost.