def explain(skater_exp: Explanation, training_df, test_df, explanation_target,
prefix_target):
job = skater_exp.job
model = joblib.load(job.predictive_model.model_path)
model = model[0]
features = list(training_df.drop(['trace_id', 'label'], 1).columns.values)
interpreter = Interpretation(training_df, feature_names=features)
X_train = training_df.drop(['trace_id', 'label'], 1)
Y_train = training_df['label'].values
model_inst = InMemoryModel(model.predict,
examples=X_train,
model_type=model._estimator_type,
unique_values=[1, 2],
feature_names=features,
target_names=['label'])
surrogate_explainer = interpreter.tree_surrogate(model_inst, seed=5)
surrogate_explainer.fit(X_train,
Y_train,
use_oracle=True,
prune='post',
scorer_type='default')
surrogate_explainer.class_names = features
viz = dtreeviz(surrogate_explainer.estimator_,
X_train,
Y_train,
target_name='label',
feature_names=features,
orientation="TD",
class_names=list(surrogate_explainer.class_names),
fancy=True,
X=None,
label_fontsize=12,
ticks_fontsize=8,
fontname="Arial")
name = create_unique_name("skater_plot.svg")
viz.save(name)
if os.path.getsize(name) > 15000000:
return 'The file size is too big'
f = open(name, "r")
response = f.read()
os.remove(name)
if os.path.isfile(name.split('.svg')[0]):
os.remove(name.split('.svg')[0])
return response
def analyze(model_prediction, X_train, y_train):
skater_model = InMemoryModel(model_prediction, examples=X_train)
interpreter = Interpretation(X_train, feature_names=X_train.columns)
surrogate_explainer = interpreter.tree_surrogate(skater_model, seed=5)
surrogate_explainer.fit(X_train,
y_train,
use_oracle=True,
prune='post',
scorer_type='default')
surrogate_explainer.plot_global_decisions(
colors=['coral', 'lightsteelblue', 'darkkhaki'],
file_name='simple_tree_pre.png')
return Image(filename='simple_tree_pre.png')
def handle(self, *args, **kwargs):
# get model
TARGET_MODEL = 71
job = Job.objects.filter(pk=TARGET_MODEL)[0]
model = joblib.load(job.predictive_model.model_path)[0]
# load data
training_df, test_df = get_encoded_logs(job)
features = list(
training_df.drop(['trace_id', 'label'], 1).columns.values)
interpreter = Interpretation(training_df, feature_names=features)
X_train = training_df.drop(['trace_id', 'label'], 1)
Y_train = training_df['label'].values
model_inst = InMemoryModel(model.predict,
examples=X_train,
model_type='classifier',
unique_values=[1, 2],
feature_names=features,
target_names=['label'])
surrogate_explainer = interpreter.tree_surrogate(model_inst, seed=5)
surrogate_explainer.fit(X_train,
Y_train,
use_oracle=True,
prune='post',
scorer_type='default')
surrogate_explainer.class_names = features
viz = dtreeviz(surrogate_explainer.estimator_,
X_train,
Y_train,
target_name='label',
feature_names=features,
orientation="TD",
class_names=list(surrogate_explainer.class_names),
fancy=True,
X=None,
label_fontsize=12,
ticks_fontsize=8,
fontname="Arial")
viz.save("skater_plot_train_2_2.svg")
## To avoid clutter I only produce plots for gradient boosting and one fold only
if (fold == 2 and modelno == 5):
# Plot PDPs of variable "alm" since it is the most important feature, for 3 of the 4 models
## alm not the most important feature for Gaussian Naive bayes tho, explain that
# for other variables just change the name
# for other models just change the number
# interpreter.partial_dependence.plot_partial_dependence(["alm"],
# pyint_model, grid_resolution=30,
# with_variance=True)
# # PDP interaction between two variables, for each class
# interpreter.partial_dependence.plot_partial_dependence([("nuc", "mit")], pyint_model,
# grid_resolution=10)
surrogate_explainer = interpreter.tree_surrogate(
oracle=pyint_model, seed=5, max_depth=4)
surrogate_explainer.fit(train_data,
train_target,
use_oracle=True,
prune='pre',
scorer_type='default')
surrogate_explainer.plot_global_decisions(
file_name='mlp_tree_class_md4.png', fig_size=(8, 8))
#show_in_notebook('simple_tree_pre.png', width=400, height=300)
# This initialization, although showcased on the docs, does not work
# surrogate_explainer = interpreter.tree_surrogate(estimator_type_='classifier',
# feature_names=featureNames[1:9],
# class_names=["CYT", "ME3", "MIT", "NUC"], seed=5)
# y_hat_train = model.predict(train_data)