SHAP takes very very long

[blanpa]

Hi,
first of all, I would like to say that Dalex is a very good package and there are many models available for the explanation that would otherwise be very difficult to include on their own.

Is it normal that the SHAP plot takes a long time? I myself did not get it to show the original SHAP plot.

Using 2000 background data samples could cause slower run times. Consider using shap.sample(data, K) or shap.kmeans(data, K) to summarize the background as K samples.
0%| | 4/1000 [08:41<35:50:59, 129.58s/it]

My Code:
#%%
from pycaret.datasets import get_data
df = get_data("bank")

#%%
import pycaret.classification as pycc

clf1 = pycc.setup(df, target = 'deposit', session_id=123, verbose= False, silent = True, use_gpu=False)

#%%
X = pycc.get_config('X').reset_index(drop=True)
y = pycc.get_config('y').reset_index(drop=True)
X_train = pycc.get_config('X_train').reset_index(drop=True)
X_test = pycc.get_config('X_test').reset_index(drop=True)
y_train = pycc.get_config('y_train').reset_index(drop=True)
y_test = pycc.get_config('y_test').reset_index(drop=True)

#%%%
import dalex as dx
dx.version

#%%
n, p = X_train.shape
#%%
METRICS = [
# keras.metrics.TruePositives(name='tp'),
# keras.metrics.FalsePositives(name='fp'),
# keras.metrics.TrueNegatives(name='tn'),
# keras.metrics.FalseNegatives(name='fn'),
tf.keras.metrics.BinaryAccuracy(name='accuracy'),
# keras.metrics.Precision(name='precision'),
tf.keras.metrics.Recall(name='recall'),
# keras.metrics.AUC(name='auc'),
]

tf.random.set_seed(11)

normalizer = tf.keras.layers.experimental.preprocessing.Normalization(input_shape=[p,])
normalizer.adapt(X.to_numpy())

model = tf.keras.Sequential([
normalizer,
tf.keras.Input(shape=(p,)),
tf.keras.layers.Dense(p2, activation='relu'),
# tf.keras.layers.Dense(p3, activation='relu'),
# tf.keras.layers.Dense(p3, activation='relu'),
# tf.keras.layers.Dense(p2, activation='relu'),
tf.keras.layers.Dense(p, activation='relu'),
tf.keras.layers.Dense(1, activation='sigmoid')
])

model.compile(
optimizer=tf.keras.optimizers.Adam(0.001),
loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
metrics=METRICS
)

model.summary()

#%%

import numpy as np
from sklearn.utils import class_weight
class_weight = class_weight.compute_class_weight('balanced',np.unique(y),y)
class_weight = dict(enumerate(class_weight))
class_weight

#%%

model.fit(X_train, y_train,
batch_size=int(n/10),
epochs=100,
class_weight = class_weight,
verbose=True)

#%%
import dalex as dx
dx.version

#%%

explainer = dx.Explainer(
model = model,
data = X_test[:2000],
y = y_test[:2000],
model_type= "classification")

#%%
explainer.model_performance()

#%%
explainer.model_parts().plot()

#%%

explainer.model_parts(type='shap_wrapper').plot()

%%

[hbaniecki]

Hi @pa-blan,

Yes, SHAP calculations take very very long. The only hint is in the warning (basically lower data dimension):
Using 2000 background data samples could cause slower run times. Consider using shap.sample(data, K) or shap.kmeans(data, K) to summarize the background as K samples.

This is why there is a "faster" method for tree ensembles (e.g. xgboost; see https://github.com/slundberg/shap#tree-ensemble-example-with-treeexplainer-xgboostlightgbmcatboostscikit-learnpyspark-models and relevant papers) and newer methods are developed.

model_parts will take especially long, because the values are calculated for many observations. For more information on SHAP I can reference https://pbiecek.github.io/ema/shapley.html.

[blanpa]

@hbaniecki Thanks for the quick response

[hbaniecki]

@pa-blan I believe mistakes were made. tf.keras.Sequential works faster with shap_explainer_type='GradientExplainer', which now will be the default. Also, added a warning when shap_explainer_type=None so that the user knows what happens.