def main():
json_file_path = process_command_line_args()
json_manager = JsonManager(json_file_path)
feature_columns = json_manager.get_feature_columns()
categorical_features = json_manager.get_categorical_features()
binary_features = json_manager.get_binary_features()
hot_encoded_path = json_manager.get_hot_encoded_path()
normalized_folder = os.fsdecode(os.path.join(\
json_manager.get_normalized_path(), \
constants.NORMALIZED_CSV_FOLDER_NAME))
combined_csv_file = os.fsdecode(os.path.join(\
normalized_folder, \
constants.COMBINED_CSV_FILENAME))
features_data = pd.read_csv(combined_csv_file, usecols=feature_columns)
for binary_variable in binary_features:
features_data[binary_variable] = features_data[binary_variable].fillna(
value=-1)
features_data[binary_variable] = features_data[binary_variable] * 1
true_false_columns = features_data[binary_features]
true_false_columns_array = true_false_columns.to_numpy()
# true_false_features(features_data, true_false_features)
# hot encoded features
hot_encoded_array, hot_encoded_header = hot_encode_features(\
features_data, categorical_features)
# remove hot encoded features from features_data dataframe
features_data = features_data.drop(columns=categorical_features +
binary_features)
features_data_array = features_data.to_numpy()
# encode labels
labels_data = pd.read_csv(combined_csv_file, \
usecols = [constants.LABEL_COLUMN_NAME])
label_encoder, labels_column_array = encode_label_column(labels_data)
# add hot_encoded columns, than numerical columns, then encoded labels to one array
final_csv = np.concatenate(\
(hot_encoded_array, binary_columns_array, \
features_data_array, labels_column_array), \
axis = constants.COLUMN_AXIS)
hot_encoded_folder = constants.add_folder_to_directory(\
constants.HOT_ENCODED_CSV_FOLDER_NAME, hot_encoded_path)
hot_encoded_file_path = os.fsdecode(os.path.join(\
hot_encoded_folder, constants.HOT_ENCODED_CSV_FILENAME))
if os.path.exists(hot_encoded_file_path):
os.remove(hot_encoded_file_path)
# make_formatter_string(hot_encoded_header, numerical_columns, label_column)
hot_encode_fmt = "%i," * len(
hot_encoded_header +
binary_features) # format hot encoded columns to ints
feature_data_fmt = "%1.3f," * len(
features_data.columns) # format numerical columns to doubles
total_fmt = hot_encode_fmt + feature_data_fmt + "%i" # for label
final_header = ','.join(
str(i) for i in (hot_encoded_header + binary_features +
list(features_data.columns)))
final_header += "," + constants.LABEL_COLUMN_NAME # for label
np.savetxt(hot_encoded_file_path, final_csv, \
fmt = total_fmt, \
header = final_header, \
delimiter = constants.CSV_DELIMITER, \
comments='')
f = open(OUTPUT_LOG_FILE, "w")
f.write("{}\n".format(total_fmt))
f.write(str((label_encoder.classes_).tolist()))
f.close()
class Runner:
"""Main runner class for this file
Attributes:
json_manager (json_manager.JsonManager): JSON Manager for config.json
log_file (_io.TextIOWrapper): Log file used for formatting
"""
def __init__(self, json_file_path, log_file_path):
"""Summary
Args:
json_file_path (str): Full filepath to config.json
log_file_path (str): Full filepath to output.log file
"""
print(
f"BehaviorTree building started using {json_file_path} and {log_file_path}"
)
self.json_manager = JsonManager(json_file_path)
self.log_file = open(log_file_path, "r")
def get_file_fmt_and_label_encoding(self):
"""Summary
Returns:
tuple(str, list<str>): Tuple containg string output file format and list of label encodings
"""
fmt = self.log_file.readline()
label_encoding = eval(self.log_file.readline())
self.log_file.close()
return fmt, label_encoding
def get_supervised_data_csv_filepath(self):
"""Returns filepath of data, uses one hot encoded if upsample = false in config.json
Returns:
string: filepath to data csv
"""
data_folder = os.fsdecode(os.path.join(\
self.json_manager.get_hot_encoded_path(), constants.HOT_ENCODED_CSV_FOLDER_NAME))
filename = constants.HOT_ENCODED_CSV_FILENAME
if self.json_manager.get_upsample_status():
data_folder = os.fsdecode(os.path.join(\
self.json_manager.get_upsampled_path(), constants.UPSAMPLED_CSV_FOLDER_NAME))
filename = constants.UPSAMPLED_CSV_FILENAME
return os.fsdecode(os.path.join(\
data_folder, filename))
def get_output_folder(self, kFold, max_depth):
path = constants.combine_folder_and_working_dir(
constants.PIPELINE_OUTPUT_FOLDER_NAME,
self.json_manager.get_output_path())
return constants.combine_folder_and_working_dir(
"{}_kFold_{}_maxDepth".format(kFold, max_depth), path)
def create_output_folder(self, kFold, max_depth):
output_folder = constants.add_folder_to_directory(\
constants.PIPELINE_OUTPUT_FOLDER_NAME, self.json_manager.get_output_path())
folder_name = "{}_kFold_{}_maxDepth".format(kFold, max_depth)
return constants.add_folder_to_directory(folder_name, output_folder)
def format_float_list_to_precision(self, list_in, precision):
prec_str = "{0:0." + str(precision) + "f}"
return [prec_str.format(i) for i in list_in]
def k_fold_train_decision_tree_w_max_depth(self, num_k_folds, max_depth,
output_full_path):
kf = KFold(shuffle=True, n_splits=num_k_folds)
# build full tree on all data
full_tree = tree.DecisionTreeClassifier(random_state = self.json_manager.get_random_state(), \
max_depth = max_depth).fit(self.features_data, self.labels_data)
# get set of alphas from cost_complexity_pruning
prune_path = full_tree.cost_complexity_pruning_path(
self.features_data, self.labels_data)
ccp_alphas, impurities = prune_path.ccp_alphas, prune_path.impurities
self.train_scores = [0] * len(ccp_alphas)
self.test_scores = [0] * len(ccp_alphas)
# split data into train/test
for train_index, test_index in kf.split(self.features_data):
X_train, X_test = self.features_data.iloc[
train_index], self.features_data.iloc[test_index]
y_train, y_test = self.labels_data.iloc[
train_index], self.labels_data.iloc[test_index]
# create tree on each alpha
for i, alpha in enumerate(ccp_alphas):
clf = tree.DecisionTreeClassifier(\
random_state = self.json_manager.get_random_state(), \
max_depth = max_depth, \
ccp_alpha=alpha)
clf = clf.fit(X_train, y_train)
self.train_scores[i] += clf.score(X_train,
y_train) / num_k_folds
self.test_scores[i] += clf.score(X_test, y_test) / num_k_folds
def generate_full_binary_set(self):
bin_set = self.json_manager.get_binary_features()
# categrorical
cat_set = self.json_manager.get_categorical_features()
# LAT
def run(self):
"""Reads in data, trains, and reports results
"""
kFold = self.json_manager.get_kfold()
max_depth = self.json_manager.get_decision_tree_depth()
constants.remove_folder_if_exists(
self.get_output_folder(kFold, max_depth))
fmt, label_encoding = self.get_file_fmt_and_label_encoding()
self.supervised_learning_dataframe = pd.read_csv(
self.get_supervised_data_csv_filepath())
self.features_data = self.supervised_learning_dataframe.loc[:, self.
supervised_learning_dataframe
.columns !=
constants.
LABEL_COLUMN_NAME]
self.labels_data = self.supervised_learning_dataframe.loc[:, self.
supervised_learning_dataframe
.columns ==
constants.
LABEL_COLUMN_NAME]
output_full_path = self.create_output_folder(kFold, max_depth)
self.k_fold_train_decision_tree_w_max_depth(kFold, max_depth,
output_full_path)
report_file = "{}_kFold_{}_maxDepth.txt".format(kFold, max_depth)
dot_pdf_header = "{}_kFold_{}_maxDepth".format(kFold, max_depth)
report_file_path = os.path.join(output_full_path, report_file)
report_file_obj = open(report_file_path, "w")
report_file_obj.write("Decision Tree with max_depth: {}, and kFold: {}\n".format(\
max_depth, kFold))
# report_file_obj.write(" Average train error with {} fold: {}\n".format(\
# kFold, sum(self.train_scores)/len(self.train_accu)))
# report_file_obj.write(" Average test error with {} fold: {}\n".format(\
# kFold, sum(self.test_accu)/len(self.test_accu)))
report_file_obj.write(
" Decision Tree (DOT format) saved to: {}\n".format(
dot_pdf_header))
report_file_obj.write(
" Decision Tree (PDF format) saved to: {}.pdf\n".format(
dot_pdf_header))
report_file_obj.write("Check {} for appropriate pruning.\n\n\n".format(
PRUNING_GRAPH_FILENAME))
clf = tree.DecisionTreeClassifier(random_state = self.json_manager.get_random_state(), \
max_depth = max_depth)
clf = clf.fit(self.features_data, self.labels_data)
dot_pdf_full_path = os.fsdecode(
os.path.join(output_full_path, dot_pdf_header))
plot_decision_tree(clf, dot_pdf_full_path, self.features_data.columns)
prune_path = clf.cost_complexity_pruning_path(self.features_data,
self.labels_data)
ccp_alphas, impurities = prune_path.ccp_alphas, prune_path.impurities
pruning_folder = constants.add_folder_to_directory(\
constants.PRUNE_FOLDER_NAME, output_full_path)
clfs = []
train_scores = []
for i, ccp_alpha in enumerate(ccp_alphas):
clf = tree.DecisionTreeClassifier(random_state = self.json_manager.get_random_state(), \
max_depth = max_depth, ccp_alpha=ccp_alpha)
clf.fit(self.features_data, self.labels_data)
score = clf.score(self.features_data, self.labels_data)
clfs.append(clf)
train_scores.append(score)
newPrunePath = constants.add_folder_to_directory(
"Pruning_{0}_{1:.6g}".format(i, ccp_alpha), pruning_folder)
decision_tree_path = os.fsdecode(os.path.join(\
newPrunePath, "{0}_kFold_{1}_maxDepth_{2}_{3:.6g}_prune".format(kFold, max_depth,i, ccp_alpha)))
plot_decision_tree(clf, decision_tree_path,
self.features_data.columns)
decision_tree_obj = clf.tree_
# theoretical split to dump decision trees out to files
# TODO: should we include categorical features as binary?
full_binary_set = self.generate_full_binary_set()
behavior_tree_obj = btBuilder.bt_espresso_mod(\
decision_tree_obj,
self.features_data.columns,
label_encoding,
self.json_manager.get_binary_features())
behaviot_tree_full_path = os.fsdecode(os.path.join(\
newPrunePath, constants.BEHAVIOR_TREE_XML_FILENAME))
# btBuilder.save_tree(behavior_tree_obj, behaviot_tree_full_path)
btBuilder.save_tree(behavior_tree_obj, newPrunePath)
report_file_obj.write("prune: {} \n".format(i))
report_file_obj.write(" ccp_alpha: {}, train score: {}\n".format(
ccp_alpha, train_scores[i]))
report_file_obj.write(
" Decision Tree saved to {}\n".format(decision_tree_path))
report_file_obj.write(" Behavior Tree saved to {}\n\n".format(
behaviot_tree_full_path))
report_file_obj.write("")
fig, ax = plt.subplots()
ax.set_xlabel("alpha")
ax.set_ylabel("accuracy")
ax.set_title(
"Accuracy vs alpha for Final Tree Prunes (note: uses all data for final training)"
)
ax.plot(ccp_alphas,
self.train_scores,
marker='o',
label="train",
drawstyle="steps-post")
ax.plot(ccp_alphas,
self.test_scores,
marker='x',
label="test",
drawstyle="steps-post")
ax.legend()
graph_path = os.fsdecode(
os.path.join(output_full_path, PRUNING_GRAPH_FILENAME))
plt.savefig(graph_path)
results_txt_file = open(
os.fsdecode(os.path.join(output_full_path, RESULTS_TEXT_FILENAME)),
"w")
alist = ccp_alphas.flatten().tolist()
acc_diffs = [
a_i - b_i for a_i, b_i in zip(self.train_scores, self.test_scores)
]
float_precision = 6
results_txt_file.write(
f"alphas:\t\t{self.format_float_list_to_precision(alist, float_precision)}\n"
)
results_txt_file.write(
f"train acc:\t{self.format_float_list_to_precision(self.train_scores, float_precision)}\n"
)
results_txt_file.write(
f"test acc:\t{self.format_float_list_to_precision(self.test_scores, float_precision)}\n"
)
results_txt_file.write(
f"acc diff:\t{self.format_float_list_to_precision(acc_diffs, float_precision)}\n"
)
results_txt_file.close()
report_file_obj.close()
print(f"BehaviorTree buidling finished, results in {output_full_path}")
def run_hotencode(json_file_path):
global add_last_action_taken
print(f"Hot encoding started using {json_file_path}")
json_manager = JsonManager(json_file_path)
feature_list = json_manager.get_feature_columns()
categorical_features = json_manager.get_categorical_features()
add_last_action_taken = json_manager.get_add_last_action_taken()
if add_last_action_taken:
categorical_features.append(constants.LAST_ACTION_TAKEN_COLUMN_NAME)
binary_features = json_manager.get_binary_features()
hot_encoded_path = json_manager.get_hot_encoded_path()
constants.remove_folder_if_exists(\
constants.HOT_ENCODED_CSV_FOLDER_NAME, hot_encoded_path)
hot_encoded_folder = constants.add_folder_to_directory(\
constants.HOT_ENCODED_CSV_FOLDER_NAME, hot_encoded_path)
hot_encoded_file_path = os.fsdecode(os.path.join(\
hot_encoded_folder, constants.HOT_ENCODED_CSV_FILENAME))
normalized_folder = os.fsdecode(os.path.join(\
json_manager.get_normalized_path(), \
constants.NORMALIZED_CSV_FOLDER_NAME))
combined_csv_file = os.fsdecode(os.path.join(\
normalized_folder, \
constants.COMBINED_CSV_FILENAME))
feature_columns = generate_feature_col_dictionary(
get_header_row(combined_csv_file), feature_list, False)
features_data = pd.read_csv(combined_csv_file, usecols=feature_columns)
features_data[binary_features] = features_data[binary_features].fillna(0)
features_data[binary_features] = features_data[binary_features].astype(
bool)
binary_columns_array = features_data[binary_features].to_numpy()
# hot encoded features
hot_encoded_array, hot_encoded_header = hot_encode_features(\
features_data, categorical_features)
# remove hot encoded features from features_data dataframe
features_data = features_data.drop(columns=categorical_features +
binary_features)
features_data_array = features_data.to_numpy()
# encode labels
labels_data = pd.read_csv(combined_csv_file, \
usecols = [constants.LABEL_COLUMN_NAME])
label_encoder, labels_column_array = encode_label_column(labels_data)
# add hot_encoded columns, than numerical columns, then encoded labels to one array
final_csv = np.concatenate(\
(hot_encoded_array, binary_columns_array, \
features_data_array, labels_column_array), \
axis = constants.COLUMN_AXIS)
# make_formatter_string(hot_encoded_header, numerical_columns, label_column)
hot_encode_fmt = "%s," * len(
hot_encoded_header +
binary_features) # format hot encoded columns to binary features
feature_data_fmt = "%1.3f," * len(
features_data.columns) # format numerical columns to doubles
total_fmt = hot_encode_fmt + feature_data_fmt + "%i" # for label
final_header = ','.join(
str(i) for i in (hot_encoded_header + binary_features +
list(features_data.columns)))
final_header += "," + constants.LABEL_COLUMN_NAME # for label
np.savetxt(hot_encoded_file_path, final_csv, \
fmt = total_fmt, \
header = final_header, \
delimiter = constants.CSV_DELIMITER, \
comments='')
f = open(OUTPUT_LOG_FILE, "w")
f.write("{}\n".format(total_fmt))
f.write(str((label_encoder.classes_).tolist()))
f.close()
print(f"Hot Encoding finished, results in {hot_encoded_file_path}")