def reconstructRules(source):
#print("Reconstructing ",source)
file_name = source.split(".json")[0]
file_name = file_name + ".py"
functions.createFile(file_name,
"#This rule was reconstructed from " + source + "\n")
with open(source, 'r') as f:
rules = json.load(f)
#print(rules)
def padleft(rule, level):
for i in range(0, level):
rule = "\t" + rule
return rule
#print("def findDecision(obj):")
max_level = 0
rule_set = []
#json file might not store rules respectively
for instance in rules:
if len(instance) > 0:
rule = []
rule.append(instance["current_level"])
rule.append(instance["leaf_id"])
rule.append(instance["parents"])
rule.append(instance["rule"])
rule_set.append(rule)
#print(padleft(instance["rule"], instance["current_level"]))
df = np.array(rule_set)
def extractRules(df, parent='root', level=1):
level_raw = level * 1
parent_raw = copy.copy(parent)
for i in range(0, df.shape[0]):
leaf_id = df[i][1]
parent_id = df[i][2]
rule = df[i][3]
if parent_id == parent:
functions.storeRule(file_name, padleft(rule, level))
level = level + 1
parent = copy.copy(leaf_id)
extractRules(df, parent, level)
level = level_raw * 1
parent = copy.copy(parent_raw) #restore
functions.storeRule(file_name, "def findDecision(obj):")
extractRules(df)
def extractRules(df, parent='root', level=1):
level_raw = level * 1
parent_raw = copy.copy(parent)
else_rule = ""
leaf_idx = 0
for i in range(0, df.shape[0]):
leaf_id = df[i][1]
parent_id = df[i][2]
rule = df[i][3]
if parent_id == parent:
if_statement = False
if rule[0:2] == "if":
if_statement = True
else_statement = False
if rule[0:5] == "else:":
else_statement = True
else_rule = rule
#------------------------
if else_statement != True:
if if_statement == True and leaf_idx > 0:
rule = "el" + rule
#print(padleft(rule, level), "(", leaf_idx,")")
functions.storeRule(file_name, padleft(rule, level))
level = level + 1
parent = copy.copy(leaf_id)
extractRules(df, parent, level)
level = level_raw * 1
parent = copy.copy(parent_raw) #restore
leaf_idx = leaf_idx + 1
#add else statement
if else_rule != "":
#print(padleft(else_rule, level))
functions.storeRule(file_name, padleft(else_rule, level))
def extractRules(df, parent='root', level=1):
level_raw = level * 1
parent_raw = copy.copy(parent)
for i in range(0, df.shape[0]):
leaf_id = df[i][1]
parent_id = df[i][2]
rule = df[i][3]
if parent_id == parent:
functions.storeRule(file_name, padleft(rule, level))
level = level + 1
parent = copy.copy(leaf_id)
extractRules(df, parent, level)
level = level_raw * 1
parent = copy.copy(parent_raw) #restore
def apply(df, config, header, dataset_features, validation_df=None):
models = []
num_of_trees = config['num_of_trees']
pbar = tqdm(range(0, num_of_trees), desc='Bagging')
for i in pbar:
#for i in range(0, num_of_trees):
pbar.set_description("Sub decision tree %d is processing" % (i + 1))
subset = df.sample(frac=1 / num_of_trees)
root = 1
moduleName = "outputs/rules/rule_" + str(i)
file = moduleName + ".py"
json_file = moduleName + ".json"
functions.createFile(file, header)
functions.createFile(json_file, "[\n")
Training.buildDecisionTree(subset,
root,
file,
config,
dataset_features,
parent_level=0,
leaf_id=0,
parents='root')
functions.storeRule(json_file, "{}]")
#--------------------------------
fp, pathname, description = imp.find_module(moduleName)
myrules = imp.load_module(moduleName, fp, pathname, description)
models.append(myrules)
#-------------------------------
return models
def extractRules(df, parent = 'root', level=1):
level_raw = level * 1; parent_raw = copy.copy(parent)
else_rule = ""
leaf_idx = 0
for i in range(0 ,df.shape[0]):
current_level = int(df[i][0])
leaf_id = df[i][1]
parent_id = df[i][2]
rule = df[i][3]
feature_name = df[i][4]
instances = int(df[i][5])
metric = float(df[i][6])
return_statement = int(df[i][7])
if parent_id == parent:
if_statement = False
if rule[0:2] == "if":
if_statement = True
else_statement = False
if rule[0:5] == "else:":
else_statement = True
else_rule = rule
#------------------------
if else_statement != True:
if if_statement == True and leaf_idx > 0:
rule = "el"+rule
#print(padleft(rule, level), "(", leaf_idx,")")
if leaf_idx == 0 and return_statement == 0:
explainer = {}
explainer["feature"] = feature_name
explainer["instances"] = instances
explainer["metric_value"] = round(metric, 4)
explainer["depth"] = current_level
explainer = "# "+json.dumps(explainer)
functions.storeRule(file_name, padleft(explainer, level))
functions.storeRule(file_name, padleft(rule, level))
level = level + 1; parent = copy.copy(leaf_id)
extractRules(df, parent, level)
level = level_raw * 1; parent = copy.copy(parent_raw) #restore
leaf_idx = leaf_idx + 1
#add else statement
if else_rule != "":
#print(padleft(else_rule, level))
functions.storeRule(file_name, padleft(else_rule, level))
def addEpochAlpha(epoch, alpha):
file = "outputs/rules/alphas.py"
content = " if epoch == " + str(epoch) + ":\n"
content += " return " + str(alpha)
functions.storeRule(file, content)
def buildDecisionTree(df, root, file, config, dataset_features, parent_level = 0, leaf_id = 0, parents = 'root', validation_df = None):
models = []
feature_names = df.columns[0:-1]
enableParallelism = config['enableParallelism']
algorithm = config['algorithm']
json_file = file.split(".")[0]+".json"
if root == 1:
if config['enableRandomForest'] != True and config['enableGBM'] != True and config['enableAdaboost'] != True:
raw_df = df.copy()
#--------------------------------------
df_copy = df.copy()
winner_name, num_of_instances, metric, metric_name = findDecision(df, config)
#find winner index, this cannot be returned by find decision because columns dropped in previous steps
j = 0
for i in dataset_features:
if i == winner_name:
winner_index = j
j = j + 1
numericColumn = False
if dataset_features[winner_name] != 'object':
numericColumn = True
#restoration
columns = df.shape[1]
for i in range(0, columns-1):
column_name = df.columns[i]; column_type = df[column_name].dtypes
if column_type != 'object' and column_name != winner_name:
df[column_name] = df_copy[column_name]
classes = df[winner_name].value_counts().keys().tolist()
#-----------------------------------------------------
num_cores = config["num_cores"]
input_params = []
#serial approach
for i in range(0,len(classes)):
current_class = classes[i]
subdataset = df[df[winner_name] == current_class]
subdataset = subdataset.drop(columns=[winner_name])
branch_index = i * 1
#create branches serially
if enableParallelism != True:
if i == 0:
#descriptor = "# Feature: "+winner_name+", Instances: "+str(num_of_instances)+", "+metric_name+": "+str(round(metric, 4))
descriptor = {
"feature": winner_name,
"instances": num_of_instances,
#"metric_name": metric_name,
"metric_value": round(metric, 4),
"depth": parent_level + 1
}
descriptor = "# "+json.dumps(descriptor)
functions.storeRule(file, (functions.formatRule(root), "", descriptor))
createBranch(config, current_class, subdataset, numericColumn, branch_index
, winner_name, winner_index, root, parents, file, dataset_features, num_of_instances, metric)
else:
input_params.append((config, current_class, subdataset, numericColumn, branch_index
, winner_name, winner_index, root, parents, file, dataset_features, num_of_instances, metric))
#---------------------------
#add else condition in the decision tree
if df.Decision.dtypes == 'object': #classification
pivot = pd.DataFrame(subdataset.Decision.value_counts()).reset_index()
pivot = pivot.rename(columns = {"Decision": "Instances","index": "Decision"})
pivot = pivot.sort_values(by = ["Instances"], ascending = False).reset_index()
else_decision = "return '%s'" % (pivot.iloc[0].Decision)
if enableParallelism != True:
functions.storeRule(file,(functions.formatRule(root), "else:"))
functions.storeRule(file,(functions.formatRule(root+1), else_decision))
else: #parallelism
leaf_id = str(uuid.uuid1())
custom_rule_file = "outputs/rules/"+str(leaf_id)+".txt"
check_rule = "else: "+else_decision
sample_rule = {}
sample_rule["current_level"] = root
sample_rule["leaf_id"] = leaf_id
sample_rule["parents"] = parents
sample_rule["rule"] = check_rule
sample_rule["feature_idx"] = -1
sample_rule["feature_name"] = ""
sample_rule["instances"] = df.shape[0]
sample_rule["metric"] = 0
sample_rule["return_statement"] = 0
#json to string
sample_rule = json.dumps(sample_rule)
functions.createFile(custom_rule_file, "")
functions.storeRule(custom_rule_file, sample_rule)
else: #regression
else_decision = "return %s" % (subdataset.Decision.mean())
if enableParallelism != True:
functions.storeRule(file,(functions.formatRule(root), "else:"))
functions.storeRule(file,(functions.formatRule(root+1), else_decision))
else:
leaf_id = str(uuid.uuid1())
custom_rule_file = "outputs/rules/"+str(leaf_id)+".txt"
check_rule = "else: "+else_decision
sample_rule = " {\n"
sample_rule += " \"current_level\": "+str(root)+",\n"
sample_rule += " \"leaf_id\": \""+str(leaf_id)+"\",\n"
sample_rule += " \"parents\": \""+parents+"\",\n"
sample_rule += " \"rule\": \""+check_rule+"\"\n"
sample_rule += " }"
functions.createFile(custom_rule_file, "")
functions.storeRule(custom_rule_file, sample_rule)
#---------------------------
#create branches in parallel
if enableParallelism == True:
"""
#this usage causes trouble for recursive functions
with Pool(number_of_cpus) as pool:
pool.starmap(createBranch, input_params)
"""
pool = MyPool(num_cores)
results = pool.starmap(createBranch, input_params)
pool.close()
pool.join()
#---------------------------------------------
if root == 1:
if enableParallelism == True:
#custom rules are stored in .txt files. merge them all in a json file
functions.createFile(json_file, "[\n")
custom_rules = []
file_index = 0
for file in os.listdir(os.getcwd()+"/outputs/rules"):
if file.endswith(".txt"):
custom_rules.append(os.getcwd()+"/outputs/rules/"+file)
#print(file) #this file stores a custom rule
f = open(os.getcwd()+"/outputs/rules/"+file, "r")
custom_rule = f.read()
if file_index > 0:
custom_rule = ", "+custom_rule
functions.storeRule(json_file, custom_rule)
f.close()
file_index = file_index + 1
functions.storeRule(json_file, "]")
#-----------------------------------
#custom rules are already merged in a json file. clear messy custom rules
#TO-DO: if random forest trees are handled in parallel, this would be a problem. You cannot know the related tree of a rule. You should store a global tree id in a rule.
for file in custom_rules:
os.remove(file)
#-----------------------------------
reconstructRules(json_file, feature_names)
#feature importance should be calculated by demand?
feature_importance(json_file, dataset_features)
#-----------------------------------
#is regular decision tree
if config['enableRandomForest'] != True and config['enableGBM'] != True and config['enableAdaboost'] != True:
#this is reguler decision tree. find accuracy here.
moduleName = "outputs/rules/rules"
fp, pathname, description = imp.find_module(moduleName)
myrules = imp.load_module(moduleName, fp, pathname, description) #rules0
models.append(myrules)
return models
def createBranch(config, current_class, subdataset, numericColumn, branch_index
, winner_name, winner_index, root, parents, file, dataset_features, num_of_instances, metric):
algorithm = config['algorithm']
enableAdaboost = config['enableAdaboost']
enableGBM = config['enableGBM']
max_depth = config['max_depth']
enableParallelism = config['enableParallelism']
charForResp = "'"
if algorithm == 'Regression':
charForResp = ""
#---------------------------
json_file = file.split(".")[0]+".json"
tmp_root = root * 1
parents_raw = copy.copy(parents)
#---------------------------
if numericColumn == True:
compareTo = current_class #current class might be <=x or >x in this case
else:
compareTo = " == '"+str(current_class)+"'"
#print(subdataset)
terminateBuilding = False
#-----------------------------------------------
#can decision be made?
if enableGBM == True and root >= max_depth: #max depth
final_decision = subdataset['Decision'].mean()
terminateBuilding = True
elif enableAdaboost == True:
#final_decision = subdataset['Decision'].value_counts().idxmax()
final_decision = functions.sign(subdataset['Decision'].mean()) #get average
terminateBuilding = True
enableParallelism = False
elif len(subdataset['Decision'].value_counts().tolist()) == 1:
final_decision = subdataset['Decision'].value_counts().keys().tolist()[0] #all items are equal in this case
terminateBuilding = True
elif subdataset.shape[1] == 1: #if decision cannot be made even though all columns dropped
final_decision = subdataset['Decision'].value_counts().idxmax() #get the most frequent one
terminateBuilding = True
elif algorithm == 'Regression' and subdataset.shape[0] < 5: #pruning condition
#elif algorithm == 'Regression' and subdataset['Decision'].std(ddof=0)/global_stdev < 0.4: #pruning condition
final_decision = subdataset['Decision'].mean() #get average
terminateBuilding = True
#-----------------------------------------------
if enableParallelism == True:
check_condition = "if" #TODO: elif checks might be above than if statements in parallel
else:
if branch_index == 0:
check_condition = "if"
else:
check_condition = "elif"
check_rule = check_condition+" obj["+str(winner_index)+"]"+compareTo+":"
leaf_id = str(uuid.uuid1())
custom_rule_file = "outputs/rules/"+str(leaf_id)+".txt"
if enableParallelism != True:
#check_rule += " # feature: "+winner_name+", instances: "+str(num_of_instances)+", "+metric_name+": "+str(round(metric, 4))
functions.storeRule(file,(functions.formatRule(root),"",check_rule))
else:
sample_rule = {}
sample_rule["current_level"] = root
sample_rule["leaf_id"] = leaf_id
sample_rule["parents"] = parents
sample_rule["rule"] = check_rule
sample_rule["feature_idx"] = winner_index
sample_rule["feature_name"] = winner_name
sample_rule["instances"] = num_of_instances
sample_rule["metric"] = metric
sample_rule["return_statement"] = 0
#json to string
sample_rule = json.dumps(sample_rule)
functions.createFile(custom_rule_file, "")
functions.storeRule(custom_rule_file, sample_rule)
#-----------------------------------------------
if terminateBuilding == True: #check decision is made
parents = copy.copy(leaf_id)
leaf_id = str(uuid.uuid1())
decision_rule = "return "+charForResp+str(final_decision)+charForResp
if enableParallelism != True:
#serial
functions.storeRule(file,(functions.formatRule(root+1),decision_rule))
else:
#parallel
sample_rule = {}
sample_rule["current_level"] = root+1
sample_rule["leaf_id"] = leaf_id
sample_rule["parents"] = parents
sample_rule["rule"] = decision_rule
sample_rule["feature_idx"] = winner_index
sample_rule["feature_name"] = winner_name
sample_rule["instances"] = num_of_instances
sample_rule["metric"] = 0
sample_rule["return_statement"] = 1
#json to string
sample_rule = ", "+json.dumps(sample_rule)
functions.storeRule(custom_rule_file, sample_rule)
else: #decision is not made, continue to create branch and leafs
root = root + 1 #the following rule will be included by this rule. increase root
parents = copy.copy(leaf_id)
buildDecisionTree(subdataset, root, file, config, dataset_features
, root-1, leaf_id, parents)
root = tmp_root * 1
parents = copy.copy(parents_raw)
def reconstructRules(source):
#print("Reconstructing ",source)
file_name = source.split(".json")[0]
file_name = file_name + ".py"
functions.createFile(file_name,
"#This rule was reconstructed from " + source + "\n")
with open(source, 'r') as f:
rules = json.load(f)
#print(rules)
def padleft(rule, level):
for i in range(0, level):
rule = "\t" + rule
return rule
#print("def findDecision(obj):")
max_level = 0
rule_set = []
#json file might not store rules respectively
for instance in rules:
if len(instance) > 0:
rule = []
rule.append(instance["current_level"])
rule.append(instance["leaf_id"])
rule.append(instance["parents"])
rule.append(instance["rule"])
rule_set.append(rule)
#print(padleft(instance["rule"], instance["current_level"]))
df = np.array(rule_set)
def extractRules(df, parent='root', level=1):
level_raw = level * 1
parent_raw = copy.copy(parent)
else_rule = ""
leaf_idx = 0
for i in range(0, df.shape[0]):
leaf_id = df[i][1]
parent_id = df[i][2]
rule = df[i][3]
if parent_id == parent:
if_statement = False
if rule[0:2] == "if":
if_statement = True
else_statement = False
if rule[0:5] == "else:":
else_statement = True
else_rule = rule
#------------------------
if else_statement != True:
if if_statement == True and leaf_idx > 0:
rule = "el" + rule
#print(padleft(rule, level), "(", leaf_idx,")")
functions.storeRule(file_name, padleft(rule, level))
level = level + 1
parent = copy.copy(leaf_id)
extractRules(df, parent, level)
level = level_raw * 1
parent = copy.copy(parent_raw) #restore
leaf_idx = leaf_idx + 1
#add else statement
if else_rule != "":
#print(padleft(else_rule, level))
functions.storeRule(file_name, padleft(else_rule, level))
#------------------------------------
#print("def findDecision(obj):")
functions.storeRule(file_name, "def findDecision(obj):")
extractRules(df)
def buildDecisionTree(df,
root,
file,
config,
dataset_features,
parent_level=0,
leaf_id=0,
parents='root'):
models = []
enableParallelism = config['enableParallelism']
algorithm = config['algorithm']
json_file = file.split(".")[0] + ".json"
if root == 1:
if config['enableRandomForest'] != True and config[
'enableGBM'] != True and config['enableAdaboost'] != True:
raw_df = df.copy()
#--------------------------------------
df_copy = df.copy()
winner_name = findDecision(df, config)
#find winner index, this cannot be returned by find decision because columns dropped in previous steps
j = 0
for i in dataset_features:
if i == winner_name:
winner_index = j
j = j + 1
numericColumn = False
if dataset_features[winner_name] != 'object':
numericColumn = True
#restoration
columns = df.shape[1]
for i in range(0, columns - 1):
column_name = df.columns[i]
column_type = df[column_name].dtypes
if column_type != 'object' and column_name != winner_name:
df[column_name] = df_copy[column_name]
classes = df[winner_name].value_counts().keys().tolist()
#-----------------------------------------------------
#TO-DO: you should specify the number of cores in config
num_cores = int(multiprocessing.cpu_count() /
2) #allocate half of your total cores
input_params = []
#serial approach
for i in range(0, len(classes)):
current_class = classes[i]
subdataset = df[df[winner_name] == current_class]
subdataset = subdataset.drop(columns=[winner_name])
branch_index = i * 1
#create branches serially
if enableParallelism != True:
createBranch(config, current_class, subdataset, numericColumn,
branch_index, winner_index, root, parents, file,
dataset_features)
else:
input_params.append((config, current_class, subdataset,
numericColumn, branch_index, winner_index,
root, parents, file, dataset_features))
#---------------------------
#add else condition in the decision tree
if df.Decision.dtypes == 'object': #classification
pivot = pd.DataFrame(subdataset.Decision.value_counts()).reset_index()
pivot = pivot.rename(columns={
"Decision": "Instances",
"index": "Decision"
})
pivot = pivot.sort_values(by=["Instances"],
ascending=False).reset_index()
else_decision = "return '%s'" % (pivot.iloc[0].Decision)
if enableParallelism != True:
functions.storeRule(file, (functions.formatRule(root), "else:"))
functions.storeRule(
file, (functions.formatRule(root + 1), else_decision))
else: #parallelism
leaf_id = str(uuid.uuid1())
custom_rule_file = "outputs/rules/" + str(leaf_id) + ".txt"
check_rule = "else: " + else_decision
sample_rule = " {\n"
sample_rule += " \"current_level\": " + str(root) + ",\n"
sample_rule += " \"leaf_id\": \"" + str(leaf_id) + "\",\n"
sample_rule += " \"parents\": \"" + parents + "\",\n"
sample_rule += " \"rule\": \"" + check_rule + "\"\n"
sample_rule += " }"
functions.createFile(custom_rule_file, "")
functions.storeRule(custom_rule_file, sample_rule)
else: #regression
else_decision = "return %s" % (subdataset.Decision.mean())
if enableParallelism != True:
functions.storeRule(file, (functions.formatRule(root), "else:"))
functions.storeRule(
file, (functions.formatRule(root + 1), else_decision))
else:
leaf_id = str(uuid.uuid1())
custom_rule_file = "outputs/rules/" + str(leaf_id) + ".txt"
check_rule = "else: " + else_decision
sample_rule = " {\n"
sample_rule += " \"current_level\": " + str(root) + ",\n"
sample_rule += " \"leaf_id\": \"" + str(leaf_id) + "\",\n"
sample_rule += " \"parents\": \"" + parents + "\",\n"
sample_rule += " \"rule\": \"" + check_rule + "\"\n"
sample_rule += " }"
functions.createFile(custom_rule_file, "")
functions.storeRule(custom_rule_file, sample_rule)
#---------------------------
#create branches in parallel
if enableParallelism == True:
"""
#this usage causes trouble for recursive functions
with Pool(number_of_cpus) as pool:
pool.starmap(createBranch, input_params)
"""
pool = MyPool(num_cores)
results = pool.starmap(createBranch, input_params)
pool.close()
pool.join()
#---------------------------------------------
#calculate accuracy metrics
if root == 1:
if enableParallelism == True:
#custom rules are stored in .txt files. merge them all in a json file
functions.createFile(json_file, "[\n")
custom_rules = []
file_index = 0
for file in os.listdir(os.getcwd() + "/outputs/rules"):
if file.endswith(".txt"):
custom_rules.append(os.getcwd() + "/outputs/rules/" + file)
#print(file) #this file stores a custom rule
f = open(os.getcwd() + "/outputs/rules/" + file, "r")
custom_rule = f.read()
if file_index > 0:
custom_rule = ", " + custom_rule
functions.storeRule(json_file, custom_rule)
f.close()
file_index = file_index + 1
functions.storeRule(json_file, "]")
#-----------------------------------
#custom rules are already merged in a json file. clear messy custom rules
#TO-DO: if random forest trees are handled in parallel, this would be a problem. You cannot know the related tree of a rule. You should store a global tree id in a rule.
for file in custom_rules:
os.remove(file)
#-----------------------------------
reconstructRules(json_file)
#-----------------------------------
if config['enableRandomForest'] != True and config[
'enableGBM'] != True and config['enableAdaboost'] != True:
#this is reguler decision tree. find accuracy here.
moduleName = "outputs/rules/rules"
fp, pathname, description = imp.find_module(moduleName)
myrules = imp.load_module(moduleName, fp, pathname,
description) #rules0
models.append(myrules)
num_of_features = df.shape[1] - 1
instances = df.shape[0]
classified = 0
mae = 0
mse = 0
#instead of for loops, pandas functions perform well
raw_df['Prediction'] = raw_df.apply(findPrediction, axis=1)
if algorithm != 'Regression':
idx = raw_df[raw_df['Prediction'] == raw_df['Decision']].index
#raw_df['Classified'] = 0
#raw_df.loc[idx, 'Classified'] = 1
#print(raw_df)
accuracy = 100 * len(idx) / instances
print("Accuracy: ", accuracy, "% on ", instances, " instances")
else:
raw_df['Absolute_Error'] = abs(raw_df['Prediction'] -
raw_df['Decision'])
raw_df['Absolute_Error_Squared'] = raw_df[
'Absolute_Error'] * raw_df['Absolute_Error']
#print(raw_df)
mae = raw_df['Absolute_Error'].sum() / instances
print("MAE: ", mae)
mse = raw_df['Absolute_Error_Squared'].sum() / instances
rmse = math.sqrt(mse)
print("RMSE: ", rmse)
mean = raw_df['Decision'].mean()
print("Mean: ", mean)
if mean > 0:
print("MAE / Mean: ", 100 * mae / mean, "%")
print("RMSE / Mean: ", 100 * rmse / mean, "%")
return models
def buildDecisionTree(df,
root,
file,
config,
dataset_features,
parent_level=0,
leaf_id=0,
parents='root',
tree_id=0,
validation_df=None,
main_process_id=None):
models = []
decision_rules = []
feature_names = df.columns[0:-1]
enableParallelism = config['enableParallelism']
algorithm = config['algorithm']
json_file = file.split(".")[0] + ".json"
random_forest_enabled = config['enableRandomForest']
enableGBM = config['enableGBM']
enableAdaboost = config['enableAdaboost']
if root == 1:
if random_forest_enabled != True and enableGBM != True and enableAdaboost != True:
raw_df = df.copy()
#--------------------------------------
df_copy = df.copy()
winner_name, num_of_instances, metric, metric_name = findDecision(
df, config)
#find winner index, this cannot be returned by find decision because columns dropped in previous steps
j = 0
for i in dataset_features:
if i == winner_name:
winner_index = j
j = j + 1
numericColumn = False
if dataset_features[winner_name] != 'object':
numericColumn = True
#restoration
columns = df.shape[1]
for i in range(0, columns - 1):
#column_name = df.columns[i]; column_type = df[column_name].dtypes #numeric field already transformed to object. you cannot check it with df itself, you should check df_copy
column_name = df_copy.columns[i]
column_type = df_copy[column_name].dtypes
if column_type != 'object' and column_name != winner_name:
df[column_name] = df_copy[column_name]
classes = df[winner_name].value_counts().keys().tolist()
#print("classes: ",classes," in ", winner_name)
#-----------------------------------------------------
num_cores = config["num_cores"]
input_params = []
#serial approach
for i in range(0, len(classes)):
current_class = classes[i]
subdataset = df[df[winner_name] == current_class]
subdataset = subdataset.drop(columns=[winner_name])
branch_index = i * 1
#create branches serially
if enableParallelism != True:
if i == 0:
descriptor = {
"feature": winner_name,
"instances": num_of_instances,
#"metric_name": metric_name,
"metric_value": round(metric, 4),
"depth": parent_level + 1
}
descriptor = "# " + json.dumps(descriptor)
functions.storeRule(
file, (functions.formatRule(root), "", descriptor))
results = createBranch(config,
current_class,
subdataset,
numericColumn,
branch_index,
winner_name,
winner_index,
root,
parents,
file,
dataset_features,
num_of_instances,
metric,
tree_id=tree_id,
main_process_id=main_process_id)
decision_rules = decision_rules + results
else:
input_params.append(
(config, current_class, subdataset, numericColumn,
branch_index, winner_name, winner_index, root, parents, file,
dataset_features, num_of_instances, metric, tree_id,
main_process_id))
#---------------------------
#add else condition in the decision tree
if df.Decision.dtypes == 'object': #classification
pivot = pd.DataFrame(subdataset.Decision.value_counts()).reset_index()
pivot = pivot.rename(columns={
"Decision": "Instances",
"index": "Decision"
})
pivot = pivot.sort_values(by=["Instances"],
ascending=False).reset_index()
else_decision = "return '%s'" % (pivot.iloc[0].Decision)
if enableParallelism != True:
functions.storeRule(file, (functions.formatRule(root), "else:"))
functions.storeRule(
file, (functions.formatRule(root + 1), else_decision))
else: #parallelism
leaf_id = str(uuid.uuid1())
check_rule = "else: " + else_decision
sample_rule = {}
sample_rule["current_level"] = root
sample_rule["leaf_id"] = leaf_id
sample_rule["parents"] = parents
sample_rule["rule"] = check_rule
sample_rule["feature_idx"] = -1
sample_rule["feature_name"] = ""
sample_rule["instances"] = df.shape[0]
sample_rule["metric"] = 0
sample_rule["return_statement"] = 0
sample_rule["tree_id"] = tree_id
#json to string
sample_rule = json.dumps(sample_rule)
decision_rules.append(sample_rule)
else: #regression
else_decision = "return %s" % (subdataset.Decision.mean())
if enableParallelism != True:
functions.storeRule(file, (functions.formatRule(root), "else:"))
functions.storeRule(
file, (functions.formatRule(root + 1), else_decision))
else:
leaf_id = str(uuid.uuid1())
check_rule = "else: " + else_decision
sample_rule = {}
sample_rule["current_level"] = root
sample_rule["leaf_id"] = leaf_id
sample_rule["parents"] = parents
sample_rule["rule"] = check_rule
sample_rule["tree_id"] = tree_id
sample_rule["feature_name"] = ""
sample_rule["instances"] = 0
sample_rule["metric"] = 0
sample_rule["return_statement"] = 1
#json to string
sample_rule = json.dumps(sample_rule)
decision_rules.append(sample_rule)
#---------------------------
try:
main_process = psutil.Process(main_process_id)
children = main_process.children(recursive=True)
active_processes = len(children) + 1 #plus parent
#active_processes = len(children)
except:
active_processes = 100 #set a large initial value
results = []
#create branches in parallel
if enableParallelism == True:
required_threads = active_processes + len(classes)
#if parent_level == 0 and random_forest_enabled != True:
if main_process_id != None and num_cores >= required_threads: #len(classes) branches will be run in parallel
#POOL_SIZE = num_cores
POOL_SIZE = len(classes)
#with closing(multiprocessing.Pool(POOL_SIZE)) as pool:
with closing(MyPool(POOL_SIZE)) as pool:
funclist = []
for input_param in input_params:
f = pool.apply_async(createBranchWrapper,
[createBranch, input_param])
funclist.append(f)
#all functions registered here
for f in funclist:
branch_results = f.get(timeout=100000)
for branch_result in branch_results:
results.append(branch_result)
pool.close()
pool.terminate()
#--------------------------------
else: #serial
for input_param in input_params:
sub_results = createBranchWrapper(createBranch, input_param)
for sub_result in sub_results:
results.append(sub_result)
#--------------------------------
decision_rules = decision_rules + results
#--------------------------------
if root != 1: #return children results until the root node
return decision_rules
#---------------------------------------------
if root == 1:
if enableParallelism == True:
#custom rules are stored in decision_rules. merge them all in a json file first
json_rules = "[\n" #initialize
file_index = 0
for custom_rule in decision_rules:
json_rules += custom_rule
if file_index < len(decision_rules) - 1:
json_rules += ", "
json_rules += "\n"
file_index = file_index + 1
#-----------------------------------
json_rules += "]"
functions.createFile(json_file, json_rules)
#-----------------------------------
#reconstruct rules from json to py
reconstructRules(json_file, feature_names)
#-----------------------------------
#is regular decision tree
if config['enableRandomForest'] != True and config[
'enableGBM'] != True and config['enableAdaboost'] != True:
#this is reguler decision tree. find accuracy here.
moduleName = "outputs/rules/rules"
fp, pathname, description = imp.find_module(moduleName)
myrules = imp.load_module(moduleName, fp, pathname,
description) #rules0
models.append(myrules)
return models
def apply(df, config, header, dataset_features):
models = []
num_of_trees = config['num_of_trees']
pbar = tqdm(range(0, num_of_trees), desc='Bagging')
for i in pbar:
#for i in range(0, num_of_trees):
pbar.set_description("Sub decision tree %d is processing" % (i + 1))
subset = df.sample(frac=1 / num_of_trees)
root = 1
moduleName = "outputs/rules/rule_" + str(i)
file = moduleName + ".py"
json_file = moduleName + ".json"
functions.createFile(file, header)
functions.createFile(json_file, "[\n")
Training.buildDecisionTree(subset,
root,
file,
config,
dataset_features,
parent_level=0,
leaf_id=0,
parents='root')
functions.storeRule(json_file, "{}]")
#--------------------------------
fp, pathname, description = imp.find_module(moduleName)
myrules = imp.load_module(moduleName, fp, pathname, description)
models.append(myrules)
#-------------------------------
#check regression or classification
if df['Decision'].dtypes == 'object': problem_type = 'classification'
else: problem_type = 'regression'
actual_values = df['Decision'].values
num_of_features = df.shape[1] - 1 #discard Decision
number_of_instances = df.shape[0]
global_predictions = []
#if classification get the max number of prediction
if problem_type == 'classification':
for i in range(0, num_of_trees):
moduleName = "outputs/rules/rule_" + str(i)
fp, pathname, description = imp.find_module(moduleName)
myrules = imp.load_module(moduleName, fp, pathname, description)
predictions = []
for index, instance in df.iterrows():
params = []
for j in range(0, num_of_features):
params.append(instance[j])
#index row, i th column
prediction = myrules.findDecision(params)
predictions.append(prediction)
#print(i,"th tree prediction: ",prediction)
#print(predictions)
global_predictions.append(predictions)
#-------------------------------
classified = 0
for index in range(0, len(actual_values)):
actual = actual_values[index]
predictions = []
for i in range(0, num_of_trees):
prediction = global_predictions[i][index]
if prediction != None: #why None exists in some cases?
predictions.append(prediction)
predictions = np.array(predictions)
unique_values = np.unique(predictions)
if unique_values.shape[0] == 1:
prediction = unique_values[0]
else:
counts = []
for unique in unique_values:
count = 0
for j in predictions:
if unique == j:
count = count + 1
counts.append(count)
#print("unique: ",unique_values)
#print("counts: ",counts)
prediction = None
if len(counts) > 0:
max_index = np.argmax(np.array(counts))
prediction = unique_values[max_index]
#print(index,". actual: ",actual," - prediction: ", prediction)
if actual == prediction:
classified = classified + 1
print("Accuracy: ", 100 * classified / number_of_instances, "% on ",
number_of_instances, " instances")
return models
def buildDecisionTree(df, root, file, config, dataset_features):
models = []
if root == 1:
if config['enableRandomForest'] != True and config[
'enableGBM'] != True and config['enableAdaboost'] != True:
raw_df = df.copy()
algorithm = config['algorithm']
enableAdaboost = config['enableAdaboost']
#--------------------------------------
#print(df.shape)
charForResp = "'"
if algorithm == 'Regression':
charForResp = ""
tmp_root = root * 1
df_copy = df.copy()
winner_name = findDecision(df, config)
#find winner index, this cannot be returned by find decision because columns dropped in previous steps
j = 0
for i in dataset_features:
if i == winner_name:
winner_index = j
j = j + 1
numericColumn = False
if dataset_features[winner_name] != 'object':
numericColumn = True
#restoration
columns = df.shape[1]
for i in range(0, columns - 1):
column_name = df.columns[i]
column_type = df[column_name].dtypes
if column_type != 'object' and column_name != winner_name:
df[column_name] = df_copy[column_name]
classes = df[winner_name].value_counts().keys().tolist()
for i in range(0, len(classes)):
current_class = classes[i]
subdataset = df[df[winner_name] == current_class]
subdataset = subdataset.drop(columns=[winner_name])
if numericColumn == True:
compareTo = current_class #current class might be <=x or >x in this case
else:
compareTo = " == '" + str(current_class) + "'"
#print(subdataset)
terminateBuilding = False
#-----------------------------------------------
#can decision be made?
if enableAdaboost == True:
#final_decision = subdataset['Decision'].value_counts().idxmax()
final_decision = functions.sign(
subdataset['Decision'].mean()) #get average
terminateBuilding = True
elif len(subdataset['Decision'].value_counts().tolist()) == 1:
final_decision = subdataset['Decision'].value_counts().keys(
).tolist()[0] #all items are equal in this case
terminateBuilding = True
elif subdataset.shape[
1] == 1: #if decision cannot be made even though all columns dropped
final_decision = subdataset['Decision'].value_counts().idxmax(
) #get the most frequent one
terminateBuilding = True
elif algorithm == 'Regression' and subdataset.shape[
0] < 5: #pruning condition
#elif algorithm == 'Regression' and subdataset['Decision'].std(ddof=0)/global_stdev < 0.4: #pruning condition
final_decision = subdataset['Decision'].mean() #get average
terminateBuilding = True
#-----------------------------------------------
if i == 0:
check_condition = "if"
else:
check_condition = "elif"
functions.storeRule(file,
(functions.formatRule(root), "", check_condition,
" obj[", str(winner_index), "]", compareTo, ":"))
#-----------------------------------------------
if terminateBuilding == True: #check decision is made
functions.storeRule(
file, (functions.formatRule(root + 1), "return ",
charForResp + str(final_decision) + charForResp))
else: #decision is not made, continue to create branch and leafs
root = root + 1 #the following rule will be included by this rule. increase root
buildDecisionTree(subdataset, root, file, config, dataset_features)
root = tmp_root * 1
#---------------------------------------------
#calculate accuracy metrics
if root == 1:
if config['enableRandomForest'] != True and config[
'enableGBM'] != True and config['enableAdaboost'] != True:
#this is reguler decision tree. find accuracy here.
moduleName = "outputs/rules/rules"
fp, pathname, description = imp.find_module(moduleName)
myrules = imp.load_module(moduleName, fp, pathname,
description) #rules0
models.append(myrules)
num_of_features = df.shape[1] - 1
instances = df.shape[0]
classified = 0
mae = 0
mse = 0
#instead of for loops, pandas functions perform well
raw_df['Prediction'] = raw_df.apply(findPrediction, axis=1)
if algorithm != 'Regression':
idx = raw_df[raw_df['Prediction'] == raw_df['Decision']].index
#raw_df['Classified'] = 0
#raw_df.loc[idx, 'Classified'] = 1
#print(raw_df)
accuracy = 100 * len(idx) / instances
print("Accuracy: ", accuracy, "% on ", instances, " instances")
else:
raw_df['Absolute_Error'] = abs(raw_df['Prediction'] -
raw_df['Decision'])
raw_df['Absolute_Error_Squared'] = raw_df[
'Absolute_Error'] * raw_df['Absolute_Error']
#print(raw_df)
mae = raw_df['Absolute_Error'].sum() / instances
print("MAE: ", mae)
mse = raw_df['Absolute_Error_Squared'].sum() / instances
rmse = math.sqrt(mse)
print("RMSE: ", rmse)
mean = raw_df['Decision'].mean()
print("Mean: ", mean)
if mean > 0:
print("MAE / Mean: ", 100 * mae / mean, "%")
print("RMSE / Mean: ", 100 * rmse / mean, "%")
return models
