ngroup=0 when using xgb_model.pred(pred_interactions=True)

[kyoungrok0517]

Related: shap/shap#464

---

Problem
I get the following error when I use pred_interactions=True to get the feature interaction. The error occurs when reshaping the result array, and I suspect this is because the ngroup (2nd axis of (35255,0,601,601)) becomes 0, which seems to be an error.

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-14-6baa08e16bb5> in <module>()
----> 1 preds = model.predict(dtest, pred_contribs=True, approx_contribs=True, pred_interactions=True)

~/anaconda/lib/python3.6/site-packages/xgboost/core.py in predict(self, data, output_margin, ntree_limit, pred_leaf, pred_contribs, approx_contribs, pred_interactions, validate_features)
   1237                     preds = preds.reshape(nrow, data.num_col() + 1, data.num_col() + 1)
   1238                 else:
-> 1239                     preds = preds.reshape(nrow, ngroup, data.num_col() + 1, data.num_col() + 1)
   1240             elif pred_contribs:
   1241                 ngroup = int(chunk_size / (data.num_col() + 1))

ValueError: cannot reshape array of size 148317785 into shape (35255,0,601,601)

[kyoungrok0517]

Here's necessary files for testing.

• Model
• Data
• Script

How to run
It won't take long. The dataset is small.

python SHAP_Analysis_For_Report.py ./model_dense.pkl ./report.dmatrix

Expected error
ValueError: cannot reshape array of size 63105 into shape (15,0,601,601)

[00:02:25] 140190x600 matrix with 76352399 entries loaded from ./data/dense/test.dmatrix
Using 15 samples
Traceback (most recent call last):
  File "SHAP_Analysis_For_Report.py", line 91, in <module>
    preds = model.predict(dtest, **shap_params)
  File "C:\Users\kyoun\Anaconda3\lib\site-packages\xgboost\core.py", line 1306, in predict
    preds = preds.reshape(nrow, ngroup, data.num_col() + 1, data.num_col() + 1)
ValueError: cannot reshape array of size 63105 into shape (15,0,601,601)

[rnarukulla-deloitte]

I have the same error while I am using the xgboost model which is built using CLI interface. Can some one help

[hcho3]

Related: https://discuss.xgboost.ai/t/scala-spark-xgboost-v0-81-shap-problem/817/3

[kyoungrok0517]

So the devs are not interested in this problem?

[hcho3]

@kyoungrok0517 Sorry I didn't have a chance to look at this. We have a backlog of issues currently, and are trying to address them when we are able to. A little patience would be appreciated.

[hcho3]

@kyoungrok0517 And I apologize for not giving you any update for more than 2 months. I will look at this issue over this weekend.

[hcho3]

@kyoungrok0517 Quick update: One thing I notice is that your script enables pred_contribs=True, approx_contribs=True, pred_interactions=True. I think only one of pred_contribs and pred_interactions can be used at a time, but somehow XGBoost does not produce error for this. I will see if the problem stays when using pred_interactions=True exclusively.

[kyoungrok0517]

Yes please try pred_interactions=True. I’ having a problem with that setting. 2019년 5월 31일 (금) 오후 6:14, Philip Hyunsu Cho <notifications@github.com>님이 작성:

…

@kyoungrok0517 <https://github.com/kyoungrok0517> Quick update: One thing I notice is that your script enables pred_contribs=True, approx_contribs=True, pred_interactions=True. Only one of them should have been enabled, but somehow XGBoost does not produce error for this. I will see if the problem stays when using pred_interactions=True exclusively. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#4276?email_source=notifications&email_token=AAIAZ7BQ7IUUZA4N4Q5YWSLPYDT75A5CNFSM4G7RMK42YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODWUVYMY#issuecomment-497638451>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAIAZ7DNW2LZ3NVZV3QZLRTPYDT75ANCNFSM4G7RMK4Q> .

[hcho3]

@kyoungrok0517 Interesting, so I changed your script from

# interactions
shap_params = {
    'pred_contribs': True,
    'approx_contribs': True,
    'pred_interactions': True
}
preds = model.predict(dtest, **shap_params)

to

# interactions
shap_params = {
    'pred_interactions': True,
    'approx_contribs': True
}
preds = model.predict(dtest, **shap_params)

and I don't get ngroup = 0 error any more. In about 20 minutes on C5.9xlarge machine, the modified script completes running.

I filed #4522 to fix the issue. Specifically, you will get an error when both pred_contribs and pred_interactions are set.

[kyoungrok0517]

I see. Thanks for the response! 2019년 5월 31일 (금) 오후 7:26, Philip Hyunsu Cho <notifications@github.com>님이 작성:

…

@kyoungrok0517 <https://github.com/kyoungrok0517> Interesting, so I changed your script from # interactions shap_params = { 'pred_contribs': True, 'approx_contribs': True, 'pred_interactions': True } preds = model.predict(dtest, **shap_params) to # interactions shap_params = { 'pred_interactions': True, 'approx_contribs': True } preds = model.predict(dtest, **shap_params) and I don't get ngroup = 0 error any more. In about 20 minutes on C5.9xlarge machine, the modified script completes running. I filed #4522 <#4522> to fix the issue. Specifically, you will get an error when both pred_contribs and pred_interactions are set. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#4276?email_source=notifications&email_token=AAIAZ7G7NVXE5PBN6HH4DUDPYD4OXA5CNFSM4G7RMK42YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODWU3EWA#issuecomment-497660504>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAIAZ7EMEZYU6O255FWYFKDPYD4OXANCNFSM4G7RMK4Q> .