Hello, ytsaig!

I have no idea about how to contact you, so I opened issues here !

Since I am doing implementation of multiclass custom objective, so I carefully read you comments on the issue #2113 Missing example of multiclass custom objective in XGBoost, and also viewed the code you published on gist.

I have question that:
Do we need to do softmax transform on preds before computing grad and hess in your code of the function softmaxobj(preds, dtrain) ? and does the corresponding code need to be updated ?

The Concern is from the reference custom_objective.py updated by the author of Tianqi Chen, details of which are as follows:

# user defined evaluation function, return a pair metric_name, result
# NOTE: when you do customized loss function, the default prediction value is margin
# this may make builtin evaluation metric not function properly
# for example, we are doing logistic loss, the prediction is score before logistic transformation
# the builtin evaluation error assumes input is after logistic transformation
# Take this in mind when you use the customization, and maybe you need write customized evaluation function

Look forward to your reply ! Sorry for disturbing!
Thx !