def main():
data = proc.read_data()
features, yfill = proc.features_yfill(data)
X_train, X_test, y_train, y_test = train_test_split(features,
yfill,
test_size=0.20,
random_state=42,
stratify=yfill)
X_train_over, y_train_over = proc.oversample(X_train, y_train, r=0.3)
#plot_roc(X_train, y_train, 'LogisticRegression', LogisticRegression(C=1e5,penalty='l2'))
'''
model_over = runLR(X_train_over, X_test, y_train_over, y_test)
test_results(model_over, X_test, y_test)
'''
model = runLR(X_train.values, X_test, y_train.values, y_test)
test_results(model, X_test, y_test)
def do_grid_search(data):
# Get the data from our function above
features, yfill = proc.features_yfill(data)
X_train, X_test, y_train, y_test = train_test_split(features,
yfill,
test_size=0.20,
random_state=42,
stratify=yfill)
# Initalize our model here
# original est = RandomForestClassifier()
est = RandomForestClassifier(bootstrap=True,
criterion="gini",
class_weight="balanced_subsample")
# Here are the params we are tuning, ie,
# if you look in the docs, all of these are 'nobs' within the GradientBoostingClassifier algo.
param_grid = {
"max_depth": [3, 5, 10, 30, 50, 100],
"max_features": [1, 3, 10, 30],
"min_samples_split": [2, 3, 10],
"min_samples_leaf": [2, 3, 10]
}
'''
{'max_depth': 10,
'max_features': 30,
'min_samples_leaf': 2,
'min_samples_split': 2}
'''
'''
param_grid = {"max_depth": [3, 5, 10, 30],
"max_features": [1, 3, 10, 30],
"min_samples_split": [2, 3, 10],
"min_samples_leaf": [2, 3, 10],
"bootstrap": [True, False],
"criterion": ["gini", "entropy"],
"class_weight":[None, "balanced_subsample"]}
Full param grid
'''
# Plug in our model, params dict, and the number of jobs, then .fit()
gs_cv = GridSearchCV(est, param_grid, n_jobs=2).fit(X_train, y_train)
# return the best score and the best params
return gs_cv.best_score_, gs_cv.best_params_
plt.plot(num_trees, recall)
#plt.ylim((0.8, 1))
plt.savefig('recall_vs_numtrees_{}.png'.format(graphid))
plt.close()
if __name__ == '__main__':
data = proc.read_data()
# bits, yfill = bits_yfill(data)
# X_train, X_test, y_train, y_test = train_test_split(bits, yfill, test_size=0.20, random_state=42, stratify =yfill)
# for num in range(10):
# rffit = RandomForestClass(X_train, X_test, y_train, y_test)
# feature_importance(bits, rffit)
# plot_features(bits, rffit, 20, 'bits', num)
features, yfill = proc.features_yfill(data)
X_train, X_test, y_train, y_test = train_test_split(features, yfill, test_size=0.20, random_state=1, stratify =yfill)
X_train_over, y_train_over = proc.oversample(X_train,y_train, r = 0.3)
rffit, y_predict = randomforest(X_train_over, X_test, y_train_over, y_test, num_est=50, cls_w = 'balanced_subsample')
precision, recall, median_recall_index, medianrecall_threshold = set_threshold(rffit, X_train, X_test, y_train, y_test)
print_threshold(rffit, X_train, X_test, y_train, y_test, medianrecall_threshold)
feature_importance(features, rffit)
'''
pprob = rffit.predict_proba(X_test)
pdf = pd.DataFrame(pprob)
print(pdf)
pdf['myH'] = pdf[1].map(lambda x: 1 if x>0.35 else 0)
my_pred = pdf['myH'].values
def uploaded_file(filename):
if request.method == 'POST':
# check if the post request has the file part
if 'file' not in request.files:
flash('No file part')
return redirect(request.url)
file = request.files['file']
# if user does not select file, browser also
# submit a empty part without filename
if file.filename == '':
flash('No selected file')
return redirect(request.url)
if file and allowed_file(file.filename):
filename = secure_filename(file.filename)
print(filename)
file.save(os.path.join(app.config['UPLOAD_FOLDER'], filename))
return redirect(url_for('uploaded_HTS', filename=filename))
elif request.method == 'GET':
# make a pd.dataframe of training data
df = pd.read_csv(os.path.join(app.config['UPLOAD_FOLDER'], filename))
# use all features and yfill (no NaNs, filled with 0)
features, yfill = proc.features_yfill(df)
#train test split at 20%
X_train, X_test, y_train, y_test = rf.train_test_split(features,
yfill,
test_size=0.20,
random_state=1,
stratify=yfill)
#Optional: oversampling of minority class for training purposes
#X_train_over, y_train_over = proc.oversample(X_train,y_train, r = 0.3)
#rffit, y_predict = rf.randomforest(X_train_over, X_test, y_train_over, y_test, num_est=50, cls_w = 'balanced_subsample')
#fit the Random Forest classifier: would like to add in a grid search
rffit, y_predict = rf.randomforest(X_train.values, X_test,
y_train.values, y_test)
# Use below to run a grid search .... takes to long to work right now
#rffit, y_predict = rf.randomforest(X_train.values, X_test, y_train.values, y_test, grid_search = 'small')
#pickle the fit model for use with test data
proc._pickle(rffit, 'RFC_fit.pkl')
#set_threshold_recall is a function which determines the threshold to set such that recall is optimized (the median of the available thresholds that return the second best recall (not 1.0))
precision_list, recall_list, median_recall_index, medianrecall_threshold = rf.set_threshold_recall(
rffit, X_train, X_test, y_train, y_test)
#print_threshold uses the trained model and the selected threshold (in this case recall optimized) to return listed statistics
precision, recall, fpr, fpr_test, tpr_test, cm = rf.print_threshold(
rffit, X_train, X_test, y_train, y_test, medianrecall_threshold)
r_cm = pd.DataFrame(cm)
proc._pickle(medianrecall_threshold, 'medianrecall_threshold.pkl')
#make a pd.dataframe of the stats for display
recall_opt_stats = pd.DataFrame([[
format(medianrecall_threshold, '.2f'),
format(recall, '.2f'),
format(fpr, '.2f'),
format(precision, '.2f'),
]],
columns=[
'Suggested Threshold',
'True Positive Rate (Recall)',
'False Positive Rate (Fall-out)',
'Precision'
])
# repeat the threshold selection process for precision optimization
p_precision, p_recall, p_median_precision, threshold_precision = rf.set_threshold_precision(
rffit, X_train, X_test, y_train, y_test)
p_precision, p_recall, p_fpr, p_fpr_test, p_tpr_test, p_cm = rf.print_threshold(
rffit, X_train, X_test, y_train, y_test, threshold_precision)
p_cm = pd.DataFrame(p_cm)
precision_opt_stats = pd.DataFrame(
[[
format(threshold_precision, '.2f'),
format(p_recall, '.2f'),
format(p_fpr, '.2f'),
format(p_precision, '.2f'),
]],
columns=[
'Suggested Threshold', 'True Positive Rate (Recall)',
'False Positive Rate (Fall-out)', 'Precision'
])
#produce a ROC plot
test_prob = rffit.predict_proba(X_test)
roc.plot_roc(X_train.values,
y_train.values,
y_test,
test_prob,
'Test',
RandomForestClassifier,
max_depth=10,
max_features=30,
min_samples_leaf=2,
min_samples_split=2)
feature_description = rf.plot_features(features,
rffit,
'Identifier',
n=10)
#option for overfit data
#roc.plot_roc(X_train_over, y_train_over, y_test, test_prob, 'Test', RandomForestClassifier, max_depth = 10, max_features= 30, min_samples_leaf= 2, min_samples_split = 2)
#roc.simple_roc(y_test, test_prob, 'ROC_RFC')
pd.set_option('display.max_colwidth', -1)
return render_template("rock.html",
data_recall_opt=recall_opt_stats.to_html(
index=False, classes="data_recall_opt"),
data_precision_opt=precision_opt_stats.to_html(
index=False, classes="data_precision_opt"),
rocname='Test',
f_descrip=feature_description.to_html(
index=False, classes="f_descrip"),
recall_cm=r_cm.to_html(classes="cm"),
precision_cm=p_cm.to_html(classes="cm"))