Hi,

I'm trying to apply the TreeExplainer to get shap_values on XGBoost model on regression problem with large dataset. During the hyperparameter tuning, it failed due to segmentation fault at the explainer.shap_values() step in certain hyperparameter sets. I used fasttreeshap=0.1.1, xgboost=1.4.1 (also tested 1.6.0) and the machine came with CPU:"Intel Xeon E5-2640 v4 (20) @ 3.400GHz" and Memory:"128GB". The sample code below is a toy script to reproduce the issue using the Superconductor dataset from example notebook:

# for debugging
import faulthandler
faulthandler.enable()

import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
import xgboost as xgb
import fasttreeshap

print(f"XGBoost version: {xgb.__version__}")
print(f"fasttreeshap version: {fasttreeshap.__version__}")

# source of data: https://archive.ics.uci.edu/ml/datasets/superconductivty+data
data = pd.read_csv("FastTreeSHAP/data/superconductor_train.csv", engine = "python")
train, test = train_test_split(data, test_size = 0.5, random_state = 0)
label_train = train["critical_temp"]
label_test = test["critical_temp"]
train = train.iloc[:, :-1]
test = test.iloc[:, :-1]

print("train XGBoost model")
xgb_model = xgb.XGBRegressor(
    max_depth = 100, n_estimators = 200, learning_rate = 0.1, n_jobs = -1, alpha = 0.12, random_state = 0)
xgb_model.fit(train, label_train)

print("run TreeExplainer()")
shap_explainer = fasttreeshap.TreeExplainer(xgb_model)

print("run shap_values()")
shap_values = shap_explainer.shap_values(train)

The time report of program execution also showed that the "Maximum resident set size" was only about 32GB.

~$ /usr/bin/time -v python segfault.py 
XGBoost version: 1.4.1
fasttreeshap version: 0.1.1
train XGBoost model
run TreeExplainer()
run shap_values()
Fatal Python error: Segmentation fault

Thread 0x00007ff2c2793740 (most recent call first):
  File "~/.local/lib/python3.8/site-packages/fasttreeshap/explainers/_tree.py", line 459 in shap_values
  File "segfault.py", line 27 in <module>
Segmentation fault (core dumped)

Command terminated by signal 11
        Command being timed: "python segfault.py"
        User time (seconds): 333.65
        System time (seconds): 27.79
        Percent of CPU this job got: 797%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:45.30
        Average shared text size (kbytes): 0
        Average unshared data size (kbytes): 0
        Average stack size (kbytes): 0
        Average total size (kbytes): 0
        Maximum resident set size (kbytes): 33753096
        Average resident set size (kbytes): 0
        Major (requiring I/O) page faults: 0
        Minor (reclaiming a frame) page faults: 8188488
        Voluntary context switches: 3048
        Involuntary context switches: 3089
        Swaps: 0
        File system inputs: 0
        File system outputs: 0
        Socket messages sent: 0
        Socket messages received: 0
        Signals delivered: 0
        Page size (bytes): 4096
        Exit status: 0

In some case (and the example above), forcing TreeExplainer(algorithm="v1") did help, which means the issue could only happen to "v2" (or "auto" passing the _memory_check()). However, by chance the v1 would raise another check_additivity issue which remained unsolved in the original algorithm.

Alternatively, passing approximate=True to explainer.shap_values() would work but have the inconsistency concerns for the reproducibility of our studies...

In this case, could you help me to debug with this issue?

Thanks you so much!

As highlighted in https://github.com/slundberg/shap/issues/2345, it seems that TreeSHAP does not compute the expected SHAP values.

The following notebook reproduces the problem with the FastTreeSHAP implementation: https://nbviewer.org/gist/glemaitre/9a30dd3a704675164b84d9bf7128882e

Note that it is not only a problem in the implementation but rather a problem in algorithm 1 of the original TreeSHAP paper (i.e. tree traversal)

First of all, thank you for this amazing project, it has really sped up my team's shap execution.

My problem arises because the numpy<1.22 restriction causes conflicts with other packages we use.

In your setup.py code it says that this restriction is due to using numba, but in their own setup.py the minimum numpy version is 1.18 and there is no maximum version requirement. In fact, I've been able to upgrade numpy's version after installing fasttreeshap in a clean environment and have had no problem at all during my test executions (using numpy==1.23.5). This is how I setup the environment on Windows 10:

$ python -m venv venv_fasttreeshap
$ venv_fasttreeshap\Scripts\activate
$ pip install fasttreeshap
$ pip install --upgrade numpy==1.23.5

It throws the following error, but pip is finaly able to install it anyway:

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
fasttreeshap 0.1.2 requires numpy<1.22, but you have numpy 1.23.5 which is incompatible.

Do you think you could upgrade the numpy version in the install_requires section of setup.py, so that it's easier to combine fasttreeshap with other packages?

Thank you!

Catboost produces a TreeEnsemble has no "num_nodes" error with this code. Btw do you support a background dataset parameter, like in shap for "interventional" vs "tree_path_dependent"? Because if your underlying code uses the "interventional" method this might be related to this bug: https://github.com/slundberg/shap/issues/2557

from catboost import CatBoostRegressor
import  fasttreeshap

X, y = shap.datasets.boston()

model = CatBoostRegressor(task_type="CPU",logging_level="Silent").fit(X, y)
explainer = fasttreeshap.TreeExplainer(model, algorithm="v2", n_jobs=-1)
shap_values = explainer(X)

# visualize the first prediction's explanation
shap.plots.waterfall(shap_values[25])

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Input In [131], in <cell line: 13>()
     10 # explain the model's predictions using SHAP
     11 # (same syntax works for LightGBM, CatBoost, scikit-learn, transformers, Spark, etc.)
     12 explainer = fasttreeshap.TreeExplainer(model, algorithm="v2", n_jobs=-1)
---> 13 shap_values = explainer(X)
     15 # visualize the first prediction's explanation
     16 shap.plots.waterfall(shap_values[25])

File ~\miniconda3\envs\Master\lib\site-packages\fasttreeshap\explainers\_tree.py:256, in Tree.__call__(self, X, y, interactions, check_additivity)
    253     feature_names = getattr(self, "data_feature_names", None)
    255 if not interactions:
--> 256     v = self.shap_values(X, y=y, from_call=True, check_additivity=check_additivity, approximate=self.approximate)
    257 else:
    258     assert not self.approximate, "Approximate computation not yet supported for interaction effects!"

File ~\miniconda3\envs\Master\lib\site-packages\fasttreeshap\explainers\_tree.py:379, in Tree.shap_values(self, X, y, tree_limit, approximate, check_additivity, from_call)
    376 algorithm = self.algorithm
    377 if algorithm == "v2":
    378     # check if memory constraint is satisfied (check Section Notes in README.md for justifications of memory check conditions in function _memory_check)
--> 379     memory_check_1, memory_check_2 = self._memory_check(X)
    380     if memory_check_1:
    381         algorithm = "v2_1"

File ~\miniconda3\envs\Master\lib\site-packages\fasttreeshap\explainers\_tree.py:483, in Tree._memory_check(self, X)
    482 def _memory_check(self, X):
--> 483     max_leaves = (max(self.model.num_nodes) + 1) / 2
    484     max_combinations = 2**self.model.max_depth
    485     phi_dim = X.shape[0] * (X.shape[1] + 1) * self.model.num_outputs

AttributeError: 'TreeEnsemble' object has no attribute 'num_nodes'

It would be nice to add a FastTreeSHAP package to conda-forge https://conda-forge.org/ (in addition to pypi)

You can use grayskull https://github.com/conda-incubator/grayskull to generate a boilerplate for the conda recipe.

Using a CatBoostRegressor() with your provided notebooks and the semiconductor dataset, I did not see any improvements regarding computing the shap files. Do you have any data on this regarding speed up or did I make a mistake? I did not find any information on this in the paper or blogpost.

Also how does fasttreeshap handle GPU support?

Code used:

cat =CatBoostRegressor(iterations=2000, task_type="CPU")#, devices="0-1")
cat.fit(train, label_train)

run_fasttreeshap(
    model = cat sample = test, interactions = False, algorithm_version = "v0", n_jobs = n_jobs,
    num_round = num_round, num_sample = num_sample, shortcut = False)

run_fasttreeshap(
    model =cat, sample = test, interactions = False, algorithm_version = "v1", n_jobs = n_jobs,
    num_round = num_round, num_sample = num_sample, shortcut = False)

Note This corresponds to an issue on slundberg/shap.

Description A LightGBM Booster does not necessarily have 'objective' in its param attribute dictionary. The documented default behavior in LightGBM is to assume regression without auto-setting the 'objective' key in the Booster.params dictionary if 'objective' is missing from param (a dictionary passed into the Training API function train).

Reproduction The following is a link to a notebook containing details for error reproduction and a view of the stack trace: Notebook for Reproduction

Thanks for your responsiveness!

Nandish Gupta Data Science Engineer, SolasAI

Description

Following the 0.1.3 release, the additivity of my logit-linked xgboost local tree shaps broke. I found that pinning my numpy version back to either 1.21.6 or 1.22.4 temporarily solved the issue. I also tested every numpy version from 1.23.1 to 1.23.5 and found that all numpy versions 1.23.* were associated with the value changes.

Brainstorming

It makes me wonder if there is any numpy or numpy-adjacent usage in FastTreeSHAP that could be reimplemented to be compatible with new versions. I would be glad to contribute if that's the case and someone could help restrict the scope of the problem (if I can't restrict it myself quickly enough). As we all know from #14, that would be highly preferable to pinning numpy versions. My instinct tells me this issue may link to an existing NumPy update note or issue... haven't looked too deeply yet.

Good Release

In any case, I still think the decision to relax the numpy constraint is the right one and anyone who runs into the same behavior can just pin the version themselves in the short term.

Environment Info

Partial result of pip list

Package                       Version
----------------------------- -----------
fasttreeshap                  0.1.3
pandas                        1.5.2
numba                         0.56.4
numpy                         1.23.5
setuptools                    65.6.3
shap                          0.41.0
wheel                         0.38.4
xgboost                       1.7.1

uname --all

Linux ip-**-**-*-*** 5.4.0-1088-aws #96~18.04.1-Ubuntu SMP Mon Oct 17 02:57:48 UTC 2022 x86_64 x86_64 x86_64 GNU/Linux

lscpu

Architecture:        x86_64
CPU op-mode(s):      32-bit, 64-bit
Byte Order:          Little Endian
CPU(s):              8
On-line CPU(s) list: 0-7
Thread(s) per core:  2
Core(s) per socket:  4
Socket(s):           1
NUMA node(s):        1
Vendor ID:           GenuineIntel
CPU family:          6
Model:               85
Model name:          Intel(R) Xeon(R) Platinum 8259CL CPU @ 2.50GHz
Stepping:            7
CPU MHz:             3102.640
BogoMIPS:            5000.00
Hypervisor vendor:   KVM
Virtualization type: full
L1d cache:           32K
L1i cache:           32K
L2 cache:            1024K
L3 cache:            36608K
NUMA node0 CPU(s):   0-7
Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single pti fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid mpx avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves ida arat pku ospke avx512_vnni

Thanks for your responsiveness to the community!

Nandish Gupta Senior AI Engineer, SolasAI

We would like to use fasttreeshap to calculate explainability values for our machine learning model. To run our model we create a linux container with a venv for the model. All the requirements are specified in a file and installed in the venv with pip -r requirements.txt. Our model depends on numpy==1.21.4 and when we try to install fasttreeshap we incur into this issue:

ERROR: Cannot install oldest-supported-numpy==0.12, oldest-supported-numpy==0.14, oldest-supported-numpy==0.15, oldest-supported-numpy==2022.1.30, oldest-supported-numpy==2022.3.27, oldest-supported-numpy==2022.4.10, oldest-supported-numpy==2022.4.18, oldest-supported-numpy==2022.4.8, oldest-supported-numpy==2022.5.27, oldest-supported-numpy==2022.5.28 and oldest-supported-numpy==2022.8.16 because these package versions have conflicting dependencies.
The conflict is caused by:
oldest-supported-numpy 2022.8.16 depends on numpy==1.17.3; python_version == "3.8" and platform_machine not in "arm64|aarch64|s390x|loongarch64" and platform_python_implementation != "PyPy"
oldest-supported-numpy 2022.5.28 depends on numpy==1.17.3; python_version == "3.8" and (platform_machine != "arm64" or platform_system != "Darwin") and platform_machine != "aarch64" and platform_machine != "s390x" and platform_machine != "loongarch64" and platform_python_implementation != "PyPy"
oldest-supported-numpy 2022.5.27 depends on numpy==1.17.3; python_version == "3.8" and (platform_machine != "arm64" or platform_system != "Darwin") and platform_machine != "aarch64" and platform_machine != "s390x" and platform_machine != "loongarch64" and platform_python_implementation != "PyPy"
oldest-supported-numpy 2022.4.18 depends on numpy==1.17.3; python_version == "3.8" and (platform_machine != "arm64" or platform_system != "Darwin") and platform_machine != "aarch64" and platform_machine != "s390x" and platform_machine != "loongarch64" and platform_python_implementation != "PyPy"
oldest-supported-numpy 2022.4.10 depends on numpy==1.17.3; python_version == "3.8" and (platform_machine != "arm64" or platform_system != "Darwin") and platform_machine != "aarch64" and platform_machine != "s390x" and platform_machine != "loongarch64" and platform_python_implementation != "PyPy"
oldest-supported-numpy 2022.4.8 depends on numpy==1.17.3; python_version == "3.8" and (platform_machine != "arm64" or platform_system != "Darwin") and platform_machine != "aarch64" and platform_machine != "s390x" and platform_machine != "loongarch64" and platform_python_implementation != "PyPy"
oldest-supported-numpy 2022.3.27 depends on numpy==1.17.3; python_version == "3.8" and (platform_machine != "arm64" or platform_system != "Darwin") and platform_machine != "aarch64" and platform_machine != "s390x" and platform_machine != "loongarch64" and platform_python_implementation != "PyPy"
oldest-supported-numpy 2022.1.30 depends on numpy==1.17.3; python_version == "3.8" and (platform_machine != "arm64" or platform_system != "Darwin") and platform_machine != "aarch64" and platform_machine != "s390x" and platform_python_implementation != "PyPy"
oldest-supported-numpy 0.15 depends on numpy==1.17.3; python_version == "3.8" and (platform_machine != "arm64" or platform_system != "Darwin") and platform_machine != "aarch64" and platform_machine != "s390x" and platform_python_implementation != "PyPy"
oldest-supported-numpy 0.14 depends on numpy==1.17.3; python_version == "3.8" and (platform_machine != "arm64" or platform_system != "Darwin") and platform_machine != "aarch64" and platform_machine != "s390x" and platform_python_implementation != "PyPy"
oldest-supported-numpy 0.12 depends on numpy==1.17.3; python_version == "3.8" and (platform_machine != "arm64" or platform_system != "Darwin") and platform_machine != "aarch64" and platform_python_implementation != "PyPy"

I can volunteer time and resources to add a wheel file for linux. Would you be interested in distributing a wheel file created by us ?

Thanks for making FastTreeShap! I'm excited to use it. I want to generate some classic SHAP plots. Can you provide an example, maybe in a jupyter notebook, of how to use your plotting library to visualize the SHAP values.

Hi team, I have both windows and linux laptop. I was able to install FastTreeSHAP package on window but not on linux. this is the error message that I got on linux: ERROR: Could not build wheels for fasttreeshap, whichh is required tot install pyproject.toml-based projects