def impute_datawig_iterative(X, mask):
X_incomplete = X.copy()
X_incomplete[mask] = np.nan
df = pd.DataFrame(X_incomplete)
df.columns = [str(c) for c in df.columns]
df = SimpleImputer.complete(df, hpo=False, verbose=0, iterations=5)
mse = evaluate_mse(df.values, X, mask)
return mse
def get_imputer(cls):
"""Get the model object for this instance, loading it if it's not already loaded."""
if cls.imputer is None:
imputer = SimpleImputer.load(model_path)
print(imputer.input_columns)
imputer.load_hpo_model()
print(imputer.input_columns)
imputer.imputer.batch_size = 1
cls.imputer = imputer
return cls.imputer
def impute_datawig(X, mask):
X_incomplete = X.copy()
X_incomplete[mask] = np.nan
df = pd.DataFrame(X_incomplete)
df.columns = [str(c) for c in df.columns]
dw_dir = os.path.join(DIR_PATH,'datawig_imputers')
df = SimpleImputer.complete(df, output_path=dw_dir, hpo=True, verbose=0, iterations=1)
for d in glob.glob(os.path.join(dw_dir,'*')):
shutil.rmtree(d)
mse = evaluate_mse(df.values, X, mask)
return mse
def process(self, store):
"""
Runs check on provided columns
:return: result of the check
"""
if any(column not in store.column_names() for column in self.columns):
raise Exception(
"Not all defined columns are present in both data frames. "
"Defined: {}. Actual: {}".format(self.columns,
store.column_names()))
self.imputer = SimpleImputer(input_columns=self.columns,
output_column=self.output_column,
output_path=self.output_path)
df1 = store.df1[self.columns]
df2 = store.df2[self.columns]
df1_train, df1_test, df2_train, df2_test = self.prepare_dfs(df1, df2)
train_df = pd.concat([df1_train, df2_train], ignore_index=True)
test_df = pd.concat([df1_test, df2_test], ignore_index=True)
self.imputer.fit(train_df, test_df, num_epochs=self.num_epochs)
imputed = self.imputer.predict(test_df)
y_true, y_pred = imputed[self.output_column], imputed[
self.output_column + '_imputed']
base_accuracy, permuted_accuracies = self.calculate_permuted_accuracies(
df1_test, df2_test, self.columns)
result = {
'y_true': y_true,
'y_pred': y_pred,
'base_accuracy': base_accuracy,
'permuted_accuracies': permuted_accuracies
}
return self.columns, result
def impute_datawig(X):
df = pd.DataFrame(X)
df.columns = [str(c) for c in df.columns]
df_imputed = {}
for output_col in df.columns:
input_cols = sorted(list(set(df.columns) - set([output_col])))
idx_missing = df[output_col].isnull()
output_path = os.path.join(dir_path, output_col)
imputer = SimpleImputer(input_columns=input_cols,
output_column=output_col,
output_path=output_path).fit(
df.loc[~idx_missing, :],
num_epochs=50,
patience=5)
df_imputed[output_col] = imputer.predict(df.loc[idx_missing, :])
shutil.rmtree(output_path)
for output_col in df.columns:
idx_missing = df[output_col].isnull()
df.loc[idx_missing, output_col] = \
df_imputed[output_col].loc[idx_missing, output_col + "_imputed"]
return df.values
import pandas as pd
"""
Load Data
"""
df = pd.read_csv('mae_train_dataset.csv').sample(n=1000)
df_train, df_test = random_split(df, split_ratios=[0.8, 0.2])
# ------------------------------------------------------------------------------------
"""
Run default SimpleImputer
"""
# Initialize a SimpleImputer model
imputer = SimpleImputer(
input_columns=[
'title', 'text'
], # columns containing information about the column we want to impute
output_column='finish', # the column we'd like to impute values for
output_path='imputer_model' # stores model data and metrics
)
# Fit an imputer model on the train data
imputer.fit(train_df=df_train, num_epochs=5)
# Impute missing values and return original dataframe with predictions
predictions = imputer.predict(df_test)
# Calculate f1 score for true vs predicted values
f1 = f1_score(predictions['finish'],
predictions['finish_imputed'],
average='weighted')
# permissions and limitations under the License.
import pandas as pd
from datawig import SimpleImputer
from datawig.utils import random_split
'''
Text Data
'''
df = pd.read_csv('../finish_val_data_sample.csv')
df_train, df_test = random_split(df, split_ratios=[0.8, 0.2])
#Fit a Model Without HPO
imputer_text = SimpleImputer(input_columns=['title', 'text'],
output_column='finish',
output_path='imputer_text_model',
num_hash_buckets=2**15,
tokens='chars')
imputer_text.fit(train_df=df_train,
learning_rate=1e-4,
num_epochs=50,
final_fc_hidden_units=[512])
#Fit a Model With HPO
imputer_text = SimpleImputer(
input_columns=['title', 'text'],
output_column='finish',
output_path='imputer_model',
)
def impute_values(df, i):
df_model = df.loc[df.site == i]
#Produce anomaly labels
df_model.loc[df_model['anomaly'] == 1, 'anomaly'] = -1
df_model.loc[df_model['anomaly'] == 0, 'anomaly'] = 1
df_model = pivot(df_model)
scaler = StandardScaler()
scaler.fit_transform(df_model.usage)
df_model.loc[df_model.anomaly.isin([-1]).any(axis=1), 'anomaly'] = -1
labels = df_model.iloc[:, 0]
print(labels)
X_train, X_test, Y_train, Y_test = train_test_split(df_model.usage.values,
labels.values,
train_size=0.75,
shuffle=False)
names = df_model.usage.columns.values
df_train = pd.DataFrame(
data=X_train,
columns=names,
)
df_train["anomaly"] = Y_train
df_train.columns = df_train.columns.astype(str)
df_test = pd.DataFrame(
data=X_test,
columns=names,
)
df_test["anomaly"] = Y_test
df_test.columns = df_test.columns.astype(str)
names = df_train.columns.values.tolist()
names.remove("anomaly")
imputer_1 = SimpleImputer(input_columns=names,
output_column=names[0],
output_path='data/imputer_model_1')
imputer_2 = SimpleImputer(input_columns=names,
output_column=names[1],
output_path='data/imputer_model_2')
print(imputer_1.__class__.__name__)
EPOCHS = 20
imputer_1.fit(
train_df=df_train,
learning_rate=1e-4,
num_epochs=EPOCHS,
)
imputer_2.fit(
train_df=df_train,
learning_rate=1e-4,
num_epochs=EPOCHS,
)
clf = IsolationForest()
clf.fit(X_train, Y_train)
def make_prediction(title, X_test):
print(title)
Y_hat = clf.predict(X_test)
pred_results = {
'mse': mean_squared_error(Y_test, Y_hat),
'f1': f1_score(Y_test, Y_hat),
'balanced_accuracy': balanced_accuracy_score(Y_test, Y_hat),
}
print(pred_results)
make_prediction("Full data", X_test)
X_test_missing = X_test
p = 0.2
indices = np.random.choice(np.arange(X_test_missing.size),
replace=False,
size=int(X_test_missing.size * p))
X_test_missing[np.unravel_index(indices, X_test_missing.shape)] = -1
make_prediction("Missing data", X_test_missing)
predictions_1 = imputer_1.predict(df_test)
predictions_2 = imputer_2.predict(df_test)
X_test_imputed = np.concatenate((predictions_1.iloc[:, 0].values.reshape(
-1, 1), predictions_2.iloc[:, 1].values.reshape(-1, 1)),
axis=1)
make_prediction("Imputed data", X_test_imputed)
class DistinctionPrecalculation(Precalculation):
def __init__(self, columns, num_epochs=10):
if not isinstance(columns, Iterable) \
or any(not isinstance(column, str) for column in columns) \
or len(columns) < 1:
raise TypeError(
"columns should be a list of strings. Received: {}".format(
columns))
self.columns = list(columns)
if not isinstance(num_epochs, int):
raise TypeError("num_epochs should be a Number. "
"Received: {} ({})".format(
num_epochs, num_epochs.__class__.__name__))
if num_epochs < 1:
raise ValueError("num_epochs should be greater than 0. "
"Received: {}.".format(num_epochs))
self.num_epochs = num_epochs
self.output_column = '__shift_detector__dataset'
self.output_path = 'tmp/basicChecks_params'
self.imputer = None
def __eq__(self, other):
if not isinstance(other, self.__class__):
return False
return set(self.columns) == set(
other.columns) and self.num_epochs == other.num_epochs
def __hash__(self):
hash_items = sorted(self.columns) + [self.__class__, self.num_epochs]
return hash(tuple(hash_items))
def process(self, store):
"""
Runs check on provided columns
:return: result of the check
"""
if any(column not in store.column_names() for column in self.columns):
raise Exception(
"Not all defined columns are present in both data frames. "
"Defined: {}. Actual: {}".format(self.columns,
store.column_names()))
self.imputer = SimpleImputer(input_columns=self.columns,
output_column=self.output_column,
output_path=self.output_path)
df1 = store.df1[self.columns]
df2 = store.df2[self.columns]
df1_train, df1_test, df2_train, df2_test = self.prepare_dfs(df1, df2)
train_df = pd.concat([df1_train, df2_train], ignore_index=True)
test_df = pd.concat([df1_test, df2_test], ignore_index=True)
self.imputer.fit(train_df, test_df, num_epochs=self.num_epochs)
imputed = self.imputer.predict(test_df)
y_true, y_pred = imputed[self.output_column], imputed[
self.output_column + '_imputed']
base_accuracy, permuted_accuracies = self.calculate_permuted_accuracies(
df1_test, df2_test, self.columns)
result = {
'y_true': y_true,
'y_pred': y_pred,
'base_accuracy': base_accuracy,
'permuted_accuracies': permuted_accuracies
}
return self.columns, result
def label_dfs(self, df1: pd.DataFrame,
df2: pd.DataFrame) -> Tuple[pd.DataFrame, pd.DataFrame]:
"""
Set labels of the first dataframe to 'A' and those of the second dataframe to 'B'
:param df1: first DataFrame
:param df2: second DataFrame
:return: tuple of labeled DataFrames
"""
# Change the logging mode of pandas in order to not show the warning that shouldn't actually be shown.
mode = pd.options.mode.chained_assignment
pd.options.mode.chained_assignment = None
df1.loc[:, self.output_column] = 'A'
df2.loc[:, self.output_column] = 'B'
pd.options.mode.chained_assignment = mode
return df1, df2
@staticmethod
def sample_dfs(df1: pd.DataFrame,
df2: pd.DataFrame) -> Tuple[pd.DataFrame, pd.DataFrame]:
"""
Sample DataFrames to length of shorter DataFrame
:param df1: first DataFrame
:param df2: second DataFrame
:return: tuple of sampled DataFrame
"""
min_len = min(len(df1), len(df2))
return df1.sample(n=min_len), df2.sample(n=min_len)
def prepare_dfs(
self, df1: pd.DataFrame, df2: pd.DataFrame
) -> Tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame, pd.DataFrame]:
"""
Create a train and a test dataset, in which the number number of tuples
that come from the first and the number of those from the second dataset are equal
:param df1: first dataset
:param df2: second dataset
:return: tuple of train and test dataset
"""
df1, df2 = self.label_dfs(df1, df2)
df1_sampled, df2_sampled = self.sample_dfs(df1, df2)
df1_train, df1_test = random_split(df1_sampled)
df2_train, df2_test = random_split(df2_sampled)
return df1_train, df1_test, df2_train, df2_test
def get_accuracy(self, df):
"""
Predict the label for df and calculate the accuracy.
:param df: DataFrame
:return: accuracy
"""
imputed = self.imputer.predict(df)
y_true, y_pred = imputed[self.output_column], imputed[
self.output_column + '_imputed']
return accuracy_score(y_true, y_pred)
def permuted_accuracy(self, df1, df2, column):
"""
Shuffle the column of both dfs and then switch it.
Calculate the accuracy for the new DataFrame.
Do this multiple times to receive a meaningful average accuracy.
:param df1: first DataFrame
:param df2: second DataFrame
:param column: the column that will be permuted
:return: averaged accuracy
"""
accuracies = []
df = pd.concat([df1, df2], ignore_index=True)
for _ in range(5):
df1_col_rand = shuffle(df1[column])
df2_col_rand = shuffle(df2[column])
col_rand = pd.concat([df2_col_rand, df1_col_rand],
ignore_index=True)
df[column] = col_rand
accuracy = self.get_accuracy(df)
accuracies.append(accuracy)
return np.array(accuracies).mean()
def calculate_permuted_accuracies(self, df1, df2, columns):
"""
Calculate the base accuracy and the permuted accuracy for all columns.
:param df1: first DataFrame
:param df2: second DataFrame
:param columns: columns to calculate the permuted accuracy for
:return: base accuracy and the permuted accuracies as a dictionary from column to accuracy
"""
df = pd.concat([df1, df2], ignore_index=True)
base_accuracy = self.get_accuracy(df)
permuted_accuracies = {
col: self.permuted_accuracy(df1, df2, col)
for col in columns
}
return base_accuracy, permuted_accuracies
import pandas as pd
from datawig import SimpleImputer
from datawig.utils import random_split
import numpy as np
"""
Text Data
"""
df = pd.read_csv('mae_train_dataset.csv').sample(n=1000)
df_train, df_test = random_split(df, split_ratios=[0.8, 0.2])
# Fit a Model Without HPO
imputer_text = SimpleImputer(input_columns=['title', 'text'],
output_column='finish',
output_path='imputer_text_model',
num_hash_buckets=2**15,
tokens='chars')
imputer_text.fit(train_df=df_train,
learning_rate=1e-4,
num_epochs=5,
final_fc_hidden_units=[512])
# Fit a Model With HPO
imputer_text = SimpleImputer(
input_columns=['title', 'text'],
output_column='finish',
output_path='imputer_model',
)
def get_imputer(cls):
"""Get the model object for this instance, loading it if it's not already loaded."""
if cls.imputer == None:
cls.imputer = SimpleImputer.load(model_path)
return cls.imputer