These are chat archives for beniz/deepdetect

2nd
Feb 2017
Ardalan
@ArdalanM
Feb 02 2017 11:08

Hi guys, @beniz :)
I am playing around xgboost on deepdetect, I was wondering if scale_pos_weight (deals with unbalanced dataset) is available ?
I am able to train xgboost for binary classification with this call:

{"async": true, "data": ["/data/ardalan.mehrani/projects/pyDD/examples/x_train.svm", "/data/ardalan.mehrani/projects/pyDD/examples/x_test.svm"], "parameters": {"input": [{}], "mllib": {"booster_params": {"grow_policy": "depthwise", "colsample": 1.0, "max_depth": 10, "lambda_bias": 0.0, "gamma": 0.0, "subsample": 1.0, "tree_method": "auto", "scale_pos_weight": 1000, "max_delta_step": 0, "min_child_weight": 1, "lambda": 1.0, "eta": 0.3, "alpha": 0.0}, "save_period": 0, "booster": "gbtree", "base_score": 0.5, "seed": 0, "objective": "multi:softprob", "eval_metric": "auc", "test_interval": 1, "nthread": -1, "iterations": 100}, "output": [{"measure": ["auc", "acc"]}]}, "service": "pyDD_MLP_2017-02-02-12-06-17-498271"}

However, scale_pos_weight does not seem to have any effect into the predictions

Emmanuel Benazera
@beniz
Feb 02 2017 11:16
hey @ArdalanM I believe xgboost has now support for many more parameters than when we started supporting it, e.g. https://github.com/dmlc/xgboost/blob/master/doc/parameter.md
we could have a blind support for parameters but at the time I had opted for a more controlled view in order to type any input parameter at the API level
you may want to open a larger issue for support of new XGBoost parameter, it'll then be executed in a short time, thanks.
Ardalan
@ArdalanM
Feb 02 2017 11:19
will do, thanks !
Emmanuel Benazera
@beniz
Feb 02 2017 11:19
typically, I think the DART booster is rather new.
Ardalan
@ArdalanM
Feb 02 2017 11:20
gpu training is fancy too :)
Emmanuel Benazera
@beniz
Feb 02 2017 11:20
OK, this should go into another issue
as a feature request
I will open the latter one
Ardalan
@ArdalanM
Feb 02 2017 11:22
great, I will just open an issue about new "regular" parameter
Emmanuel Benazera
@beniz
Feb 02 2017 11:22
All parameters can be supported, that's fine
tks