def test_get_type_pandas(self):
d = {"col1": [1, 2, 3], "col2": ["a", "b", "c"]}
df = pd.DataFrame(data=d)
col1_type = PreprocessingUtils.get_type(df["col1"])
self.assertNotEqual(col1_type, PreprocessingUtils.CATEGORICAL)
col2_type = PreprocessingUtils.get_type(df["col2"])
self.assertNotEqual(col1_type, PreprocessingUtils.CATEGORICAL)
def _fit_na_fill(self, X):
for column in self._columns:
if np.sum(pd.isnull(X[column]) == True) == 0:
continue
self._na_fill_params[column] = self._get_fill_value(X[column])
if PreprocessingUtils.get_type(X[column]) == PreprocessingUtils.DATETIME:
self._datetime_columns += [column]
def _get_fill_value(self, x):
# categorical type
if PreprocessingUtils.get_type(x) == PreprocessingUtils.CATEGORICAL:
if self._na_fill_method == PreprocessingMissingValues.FILL_NA_MIN:
return (
PreprocessingMissingValues.MISSING_VALUE
) # add new categorical value
return PreprocessingUtils.get_most_frequent(x)
if PreprocessingUtils.get_type(x) == PreprocessingUtils.DATETIME:
return PreprocessingUtils.get_most_frequent(x)
# numerical type
if self._na_fill_method == PreprocessingMissingValues.FILL_NA_MIN:
return PreprocessingUtils.get_min(x) - 1.0
if self._na_fill_method == PreprocessingMissingValues.FILL_NA_MEAN:
return PreprocessingUtils.get_mean(x)
return PreprocessingUtils.get_median(x)
def _fit_categorical_convert(self, X):
for column in self._columns:
if PreprocessingUtils.get_type(
X[column]) != PreprocessingUtils.CATEGORICAL:
# no need to convert, already a number
continue
# limit categories - it is needed when doing one hot encoding
# this code is also used in predict.py file
# and transform_utils.py
# TODO it needs refactoring !!!
too_much_categories = len(np.unique(list(X[column].values))) > 200
lbl = None
if (self._convert_method
== PreprocessingCategorical.CONVERT_ONE_HOT
and not too_much_categories):
lbl = LabelBinarizer()
lbl.fit(X, column)
else:
lbl = LabelEncoder()
lbl.fit(X[column])
if lbl is not None:
self._convert_params[column] = lbl.to_json()
def compute(X_train, y_train, eda_path):
try:
# check if exists
if os.path.exists(eda_path):
# probably the EDA analysis is already done
# skip from here
return
else:
# need to create directory for EDA analysis
os.mkdir(eda_path)
inform = defaultdict(list)
if isinstance(y_train, pd.Series):
if PreprocessingUtils.get_type(y_train) in ("categorical"):
plt.figure(figsize=(5, 5))
sns.countplot(y_train, color=BLUE)
plt.title("Target class distribution")
plt.tight_layout(pad=2.0)
plot_path = os.path.join(eda_path, "target.png")
plt.savefig(plot_path)
plt.close("all")
else:
plt.figure(figsize=(5, 5))
sns.distplot(y_train, color=BLUE)
plt.title("Target class distribution")
plt.tight_layout(pad=2.0)
plot_path = os.path.join(eda_path, "target.png")
plt.savefig(plot_path)
plt.close("all")
inform["missing"].append(
pd.isnull(y_train).sum() / y_train.shape[0])
inform["unique"].append(y_train.nunique())
inform["feature_type"].append(
PreprocessingUtils.get_type(y_train))
inform["plot"].append("")
inform["feature"].append("target")
inform["desc"].append(y_train.describe().to_dict())
for col in X_train.columns:
inform["feature_type"].append(
PreprocessingUtils.get_type(X_train[col]))
if PreprocessingUtils.get_type(X_train[col]) in (
"categorical",
"discrete",
):
plt.figure(figsize=(5, 5))
chart = sns.countplot(
X_train[col],
order=X_train[col].value_counts().iloc[:10].index,
color=BLUE,
)
chart.set_xticklabels(chart.get_xticklabels(), rotation=90)
plt.title(f"{col} class distribution")
plt.tight_layout(pad=2.0)
plot_path = os.path.join(eda_path, f"{col}.png")
plt.savefig(plot_path)
plt.close("all")
elif PreprocessingUtils.get_type(
X_train[col]) in ("continous"):
plt.figure(figsize=(5, 5))
sns.distplot(X_train[col], color=BLUE)
plt.title(f"{col} value distribution")
plt.tight_layout(pad=2.0)
plot_path = os.path.join(eda_path, f"{col}.png")
plt.savefig(plot_path)
plt.close("all")
elif PreprocessingUtils.get_type(X_train[col]) in ("text"):
plt.figure(figsize=(10, 10), dpi=70)
word_string = " ".join(X_train[col].str.lower())
wordcloud = WordCloud(
width=500,
height=500,
stopwords=STOPWORDS,
background_color="white",
max_words=400,
max_font_size=None,
).generate(word_string)
plt.imshow(wordcloud,
aspect="auto",
interpolation="nearest")
plt.axis("off")
plot_path = os.path.join(eda_path, f"{col}.png")
plt.savefig(plot_path)
elif PreprocessingUtils.get_type(X_train[col]) in ("datetime"):
plt.figure(figsize=(5, 5))
pd.to_datetime(X_train[col]).plot(grid="True", color=BLUE)
plt.tight_layout(pad=2.0)
plot_path = os.path.join(eda_path, f"{col}.png")
plt.savefig(plot_path)
plt.close("all")
inform["missing"].append(
pd.isnull(X_train[col]).sum() * 100 / X_train.shape[0])
inform["unique"].append(int(X_train[col].nunique()))
inform["plot"].append(f"")
inform["feature"].append(str(col))
inform["desc"].append(X_train[col].describe().to_dict())
df = pd.DataFrame(inform)
with open(os.path.join(eda_path, "README.md"), "w") as fout:
for i, row in df.iterrows():
fout.write(f"## Feature : {row['feature']}\n")
fout.write(f"- **Feature type** : {row['feature_type']}\n")
fout.write(f"- **Missing** : {row['missing']}%\n")
fout.write(f"- **Unique** : {row['unique']}\n")
for key in row["desc"].keys():
if key in ("25%", "50%", "75%"):
fout.write(
f"- **{key.capitalize()}th Percentile** : {row['desc'][key]}\n"
)
else:
fout.write(
f"- **{key.capitalize()}** :{row['desc'][key]}\n"
)
fout.write(f"- {row['plot']}\n")
fout.close()
except Exception as e:
logger.error(f"There was an issue when running EDA. {str(e)}")
def extensive_eda(X, y, save_path):
# Check for empty dataframes in params
if not isinstance(X, pd.DataFrame):
raise ValueError("X should be a dataframe")
if X.shape[0] != len(y):
raise ValueError("X and y should have the same number of samples")
if X.shape[1] > MAXCOL:
X = X.iloc[:, :MAXCOL]
warnings.warn(
f"AutoML EDA column limit exceeded! running for first {MAXCOL} columns"
)
if save_path:
if not os.path.exists(save_path):
os.mkdir(save_path)
else:
raise ValueError("Please provide a valid path to save the Extensive EDA")
plt.style.use("ggplot")
try:
if PreprocessingUtils.get_type(y) in ("categorical", "discrete"):
for col in X.columns:
if PreprocessingUtils.get_type(X[col]) == "continous":
plt.figure(figsize=(5, 5))
for i in np.unique(y):
sns.kdeplot(
x=X.iloc[np.where(y == i)[0]][col],
label=f"class {i}",
shade=True,
)
plt.legend()
plt.gca().set_title(
f"Distribution of {col} for each class",
fontsize=11,
weight="bold",
alpha=0.75,
)
plt.savefig(EDA.plot_path(save_path, col + "_target"))
elif PreprocessingUtils.get_type(X[col]) in (
"categorical",
"discrete",
):
if X[col].nunique() > 7:
warnings.warn("Considering 7 the most frequent values")
values = X[col].value_counts().index[:7]
plt.figure(figsize=(5, 5))
sns.countplot(
x=X[X[col].isin(values)][col], hue=y[X[col].isin(values)]
)
plt.gca().set_title(
f"Count plot of each {col}",
fontsize=11,
weight="bold",
alpha=0.75,
)
plt.savefig(EDA.plot_path(save_path, col + "_target"))
elif PreprocessingUtils.get_type(y) == "continous":
for col in X.columns:
if PreprocessingUtils.get_type(X[col]) == "continous":
plt.figure(figsize=(5, 5))
plt.scatter(X[col].values, y)
plt.gca().set_xlabel(f"{col}")
plt.gca().set_ylabel("target")
plt.gca().set_title(
f"Scatter plot of {col} vs target",
fontsize=11,
weight="bold",
alpha=0.75,
)
plt.savefig(EDA.plot_path(save_path, col + "_target"))
elif PreprocessingUtils.get_type(X[col]) in (
"categorical",
"discrete",
):
if X[col].nunique() > 7:
warnings.warn("Considering 7 the most frequent values")
plt.figure(figsize=(5, 5))
for i in X[col].value_counts().index[:7]:
sns.kdeplot(
x=y[X[X[col] == i].index],
shade=True,
label=f"{col}_{i}",
)
plt.gca().set_title(
f"Distribution of target for each {col}",
fontsize=11,
weight="bold",
alpha=0.75,
)
plt.legend()
plt.savefig(EDA.plot_path(save_path, col + "_target"))
elif PreprocessingUtils.get_type(X[col]) == "datetime":
plt.figure(figsize=(5, 5))
plt.plot(X[col], y)
plt.gca().set_xticklabels(X[col].dt.date, rotation="45")
plt.gca().set_title(
f"Distribution of target over time",
fontsize=11,
weight="bold",
alpha=0.75,
)
plt.savefig(EDA.plot_path(save_path, col + "_target"))
cols = [
col
for col in X.columns
if PreprocessingUtils.get_type(X[col]) == "continous"
][:COLS]
if len(cols) > 0:
plt.figure(figsize=(10, 10))
sns.heatmap(X[cols].corr())
plt.gca().set_title("Heatmap", fontsize=11, weight="bold", alpha=0.75)
plt.savefig(os.path.join(save_path, "heatmap"))
with open(os.path.join(save_path, "Extensive_EDA.md"), "w") as fout:
for col in X.columns:
fout.write(f"## Bivariate analysis of {col} feature with target\n")
fout.write("\n\n".format(EDA.plot_fname(col + "_target")))
fout.write("\n")
fout.write(
"------------------------------------------------------\n"
)
if len(cols) > 0:
fout.write("## Heatmap\n")
fout.write("\n")
fout.write("\n")
fout.write(
"------------------------------------------------------\n"
)
except Exception as e:
AutoMLException(e)
def test_get_type_numpy_number(self):
tmp = np.array([1, 2, 3])
tmp_type = PreprocessingUtils.get_type(tmp)
self.assertNotEqual(tmp_type, PreprocessingUtils.CATEGORICAL)
def test_get_type_numpy_categorical(self):
tmp = np.array(["a", "b", "c"])
tmp_type = PreprocessingUtils.get_type(tmp)
self.assertEqual(tmp_type, PreprocessingUtils.CATEGORICAL)
