def compute_explanation(model: Model, x, z):
""" generates an explanation how the model came to its result
Parameters
----------
model : Model
a hassbrain model
x : np.ndarray
the raw array conatining the sensor values
Returns
-------
"""
from matplotlib.pyplot import Figure
wrapped_model = ModelWrapper(model)
class_names = model.get_state_lbl_lst()
feature_names = model.get_obs_lbl_lst()
cat_idxs = [i for i in range(len(class_names))]
categorical_names = {}
for i in cat_idxs:
categorical_names[i] = {}
categorical_names[i][0] = "off"
categorical_names[i][1] = "on"
from skater.core.local_interpretation.lime.lime_tabular import LimeTabularExplainer
exp = LimeTabularExplainer(x,
mode='classification',
training_labels=z,
feature_names=feature_names,
categorical_features=cat_idxs,
categorical_names=categorical_names,
class_names=class_names)
fig = exp.explain_instance(x[0],
wrapped_model.predict_proba).as_pyplot_figure()
def test_regression_without_feature_names(self):
"""
Ensure lime.lime_tabular works when predict_fn = regressor.predict
and feature names are NOT passed
:return:
"""
interpretor = LimeTabularExplainer(self.X)
assert interpretor.explain_regressor_instance(self.example,
self.regressor.predict)
def test_classifier_with_proba_without_feature_names(self):
"""
Ensure lime.lime_tabular works when predict_fn = classifier.predict_proba
and feature names are NOT passed
:return:
"""
interpretor = LimeTabularExplainer(self.X)
assert interpretor.explain_instance(self.example,
self.classifier.predict_proba)
def test_classifier_no_proba_without_feature_names(self):
"""
Ensure lime.lime_tabular works when predict_fn = classifier.predict
and feature names are NOT passed
:return:
"""
interpretor = LimeTabularExplainer(self.X)
interpretor_func = partial(interpretor.explain_instance,
*[self.example, self.classifier.predict])
self.assertRaises(NotImplementedError, interpretor_func)
def on_value_change(change):
index = change['new']
exp = LimeTabularExplainer(X_test,
feature_names=features,
discretize_continuous=False,
class_names=['acdc', 'non_acdc'])
print("Model behavior at row: {}".format(index))
# Lets evaluate the prediction from the model and actual target label
print("prediction from the model:{}".format(estimator.predict(X_test[index].reshape(1, -1))))
print("Target Label on the row: {}".format(y_test.reshape(1,-1)[0][index]))
clear_output()
display(HTML(exp.explain_instance(X_test[index], models['ensemble'].predict_proba).as_html()))
def test_lime_coef_accuracy(self):
"""
Ensure that for a trivial example, the coefficients of a regressor explanation
are all similar to the true beta values of the generative process.
:return:
"""
error_epsilon = .1
explainer = LimeTabularExplainer(self.X, discretize_continuous=True)
explanation = explainer.explain_regressor_instance(
self.example,
self.regressor.predict,
model_regressor=self.model_regressor)
vals = dict(explanation.as_list())
keys = ['{} <= 0.00'.format(i) for i in [2, 1, 0]]
lime_coefs = np.array([vals[key] for key in keys])
assert (abs(self.regressor.coef_ - lime_coefs) < error_epsilon).all()
def __init__(self, mdl, test_x, test_z):
from hassbrain_algorithm.benchmark.interpretation import ModelWrapper
wrapped_model = ModelWrapper(mdl)
class_names = mdl.get_state_lbl_lst()
feature_names = mdl.get_obs_lbl_lst()
cat_idxs = [i for i in range(len(feature_names))]
categorical_names = {}
for i in cat_idxs:
categorical_names[i] = {}
categorical_names[i][0] = "off"
categorical_names[i][1] = "on"
from skater.core.local_interpretation.lime.lime_tabular import LimeTabularExplainer
self._exp = LimeTabularExplainer(test_x,
mode='classification',
training_labels=test_z,
feature_names=feature_names,
categorical_features=cat_idxs,
categorical_names=categorical_names,
class_names=class_names)
def run_explanations(csv_path, csv_columns, target_column, zero_value):
# Read the dataset from the provided CSV and print out information about it.
df = pd.read_csv(csv_path,
names=csv_columns,
skipinitialspace=True,
skiprows=1)
#df = df.drop('Target',axis=1)
input_features = [name for name in csv_columns if name != target_column]
#data, labels = shap.datasets.adult(display=True)
if target_column not in csv_columns:
print("target column error")
return ("target column error")
elif zero_value not in df[target_column].tolist():
if str.isdecimal(zero_value) and (
np.int64(zero_value) in df[target_column].tolist()
or np.float64(zero_value) in df[target_column].tolist()):
print("happy")
zero_value = np.int64(zero_value)
else:
print(zero_value, df[target_column].tolist(),
df[target_column].dtype)
return ("zero value error")
labels = df[target_column].tolist()
#labels = np.array([int(label) for label in labels])
labels2 = []
for label in labels:
if label == zero_value:
labels2.append(0)
else:
labels2.append(1)
labels = np.array(labels2)
data = df[input_features]
for feature in input_features:
if data[feature].dtype is not np.dtype(
np.int64) and data[feature].dtype is not np.dtype(
np.float64) and data[feature].dtype is not np.dtype(
np.float32):
data[feature] = data[feature].astype('category')
cat_cols = data.select_dtypes(['category']).columns
data[cat_cols] = data[cat_cols].apply(lambda x: x.cat.codes)
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(data,
labels,
test_size=0.3,
random_state=42)
data_disp, labels_disp = shap.datasets.adult(display=True)
X_train_disp, X_test_disp, y_train_disp, y_test_disp = train_test_split(
data_disp, labels_disp, test_size=0.3, random_state=42)
xgc = xgb.XGBClassifier(n_estimators=500,
max_depth=5,
base_score=0.5,
objective='binary:logistic',
random_state=42)
xgc.fit(X_train, y_train)
predictions = xgc.predict(X_test)
fig = plt.figure(figsize=(16, 12))
title = fig.suptitle("Default Feature Importances from XGBoost",
fontsize=14)
ax1 = fig.add_subplot(2, 2, 1)
xgb.plot_importance(xgc, importance_type='weight', ax=ax1)
t = ax1.set_title("Feature Importance - Feature Weight")
ax2 = fig.add_subplot(2, 2, 2)
xgb.plot_importance(xgc, importance_type='gain', ax=ax2)
t = ax2.set_title("Feature Importance - Split Mean Gain")
ax3 = fig.add_subplot(2, 2, 3)
xgb.plot_importance(xgc, importance_type='cover', ax=ax3)
t = ax3.set_title("Feature Importance - Sample Coverage")
#plt.savefig('static/explanations.png')
explanation = eli5.explain_weights(xgc.get_booster())
explanation_html = eli5.formatters.html.format_as_html(explanation)
print(explanation_html)
with open("templates/explanation.html", "a+") as file:
file.write(explanation_html)
doc_num = 0
print('Actual Label:', y_test[doc_num])
print('Predicted Label:', predictions[doc_num])
#eli5.show_prediction(xgc.get_booster(), X_test.iloc[doc_num],
# feature_names=list(data.columns) ,show_feature_values=True)
explanation2 = eli5.explain_prediction(xgc.get_booster(),
X_test.iloc[doc_num],
feature_names=list(data.columns))
explanation_html2 = eli5.formatters.html.format_as_html(explanation2)
with open("templates/explanation.html", "a") as file:
file.write(explanation_html2)
doc_num = 2
print('Actual Label:', y_test[doc_num])
print('Predicted Label:', predictions[doc_num])
#eli5.show_predicon(xgc.get_booster(), X_test.iloc[doc_num], feature_names=list(data.columns) ,show_feature_values=True)
explanation3 = eli5.explain_prediction(xgc.get_booster(),
X_test.iloc[doc_num],
feature_names=list(data.columns))
explanation_html3 = eli5.formatters.html.format_as_html(explanation3)
with open("templates/explanation.html", "a") as file:
file.write(explanation_html3)
#target_names = ['$50K or less', 'More than $50K']
interpreter = Interpretation(training_data=X_test,
training_labels=y_test,
feature_names=list(data.columns))
im_model = InMemoryModel(xgc.predict_proba, examples=X_train)
plots = interpreter.feature_importance.plot_feature_importance(
im_model, ascending=True, n_samples=23000)
plots[0].savefig('skater.png')
features_pdp = input_features
xgc_np = xgb.XGBClassifier(n_estimators=500,
max_depth=5,
base_score=0.5,
objective='binary:logistic',
random_state=42)
xgc_np.fit(X_train.values, y_train)
# In[ ]:
from skater.core.local_interpretation.lime.lime_tabular import LimeTabularExplainer
exp = LimeTabularExplainer(X_test.values,
feature_names=list(data.columns),
discretize_continuous=True)
doc_num = 0
print('Actual Label:', y_test[doc_num])
print('Predicted Label:', predictions[doc_num])
instance = exp.explain_instance(X_test.iloc[doc_num].values,
xgc_np.predict_proba)
instance.save_to_file('templates/lime.html', show_all=False)
doc_num = 2
print('Actual Label:', y_test[doc_num])
print('Predicted Label:', predictions[doc_num])
instance2 = exp.explain_instance(X_test.iloc[doc_num].values,
xgc_np.predict_proba)
instance2.save_to_file('templates/lime2.html', show_all=False)
explainer = shap.TreeExplainer(xgc)
shap_values = explainer.shap_values(X_test)
pd.DataFrame(shap_values).head()
#shap.force_plot(explainer.expected_value, shap_values[:,], X_test_disp.iloc[:,],show=False,matplotlib=True)
#plt.savefig("static/force_plot.png")
shap.summary_plot(shap_values, X_test, plot_type="bar", show=False)
plt.savefig("static/summary_plot.png")
shap.summary_plot(shap_values, X_test, show=False)
plt.savefig("static/summary_plot2.png")
return "Everyone Happy"
with_variance=True,
figsize=(10, 5))
model_feature_interaction = InMemoryModel(estimator.predict_proba, examples=X_train, target_names=['acdc',
'non_acdc'])
# Two-way interaction
interpreter.partial_dependence.plot_partial_dependence([("maxDeltaEta_tag_tag", "mass_higgsLikeDijet")],
model,
grid_resolution=10)
from skater.core.local_interpretation.lime.lime_tabular import LimeTabularExplainer
exp = LimeTabularExplainer(X_train,
feature_names=features,
class_names=['acdc', 'non_acdc'],
discretize_continuous=True)
plt.show()
# explain prediction for data point for background label 'non_acdc'
exp.explain_instance(X_test[1], estimator.predict_proba)
# Interactive slider for controlling grid resolution
def understanding_interaction():
pyint_model = InMemoryModel(estimator.predict_proba, examples=X_test, target_names=features)
# ['worst area', 'mean perimeter'] --> list(feature_selection.value)
# Two-way iteraction
interpreter.partial_dependence.plot_partial_dependence(["mass_tag_tag_max_mass", "maxDeltaEta_jet_jet"],
model,
grid_resolution=grid_resolution.value,
feature_indices],
train_labels=wqp_train_y,
plot_step=0.02,
cmap=plt.cm.RdYlBu,
markers=[',', 'd', '+'],
alphas=[1.0, 0.8, 0.5],
colors=['r', 'b', 'y'])
# ## Interpreting Model Predictions
# In[36]:
from skater.core.local_interpretation.lime.lime_tabular import LimeTabularExplainer
exp = LimeTabularExplainer(wqp_train_SX,
feature_names=wqp_feature_names,
discretize_continuous=True,
class_names=wqp_rf.classes_)
# In[80]:
exp.explain_instance(wqp_test_SX[10], wqp_rf.predict_proba,
top_labels=1).show_in_notebook()
# In[81]:
exp.explain_instance(wqp_test_SX[747], wqp_rf.predict_proba,
top_labels=1).show_in_notebook()
# ## Visualizing partial dependencies
# In[39]:
#feature_names = hmm_model.get_obs_lbl_lst()
#class_names = hmm_model.get_state_lbl_lst()
def boolean_arr2str(arr):
res = arr.astype(str)
return res
cat_idxs = [i for i in range(len(class_names))]
categorical_names = {}
for i in cat_idxs:
categorical_names[i] = {}
categorical_names[i][0] = "off"
categorical_names[i][1] = "on"
from skater.core.local_interpretation.lime.lime_tabular import LimeTabularExplainer
exp = LimeTabularExplainer(train_x,
mode='classification',
training_labels=train_z,
feature_names=feature_names,
categorical_features=cat_idxs,
categorical_names=categorical_names,
class_names=class_names)
#fig = exp.explain_instance(train_x[0], model.predict_proba).as_pyplot_figure()
#fig.show()
x = train_x[0]
lst = exp.explain_instance(x, model.predict_proba).as_list()
print(lst)
# In[203]:
p = interpreter.partial_dependence.plot_partial_dependence(['worst area'],
model,
grid_resolution=50,
with_variance=True,
figsize=(6, 4))
# ## Explaining Predictions
# In[195]:
from skater.core.local_interpretation.lime.lime_tabular import LimeTabularExplainer
exp = LimeTabularExplainer(X_train,
feature_names=data.feature_names,
discretize_continuous=True,
class_names=['0', '1'])
# In[204]:
exp.explain_instance(X_test[0], logistic.predict_proba).show_in_notebook()
# In[202]:
exp.explain_instance(X_test[1], logistic.predict_proba).show_in_notebook()
# # Model Deployment
# ## Persist model to disk
# In[207]: