def main():
with mlflow.start_run():
# mlflow native APIs
mlflow.log_param("param", 0)
mlflow.log_metric("metric", 1.0)
# flatten dict
mlflow.log_params_flatten({"a": {"b": 0}})
mlflow.log_metrics_flatten({"a": {"b": 0.0}})
# dict
mlflow.log_dict({"a": 0}, "dict.json")
# numpy array
mlflow.log_numpy(np.array([0]), "array.npy")
# pandas dataframe
mlflow.log_df(pd.DataFrame({"a": [0]}), "df.csv")
# matplotlib figure
fig, ax = plt.subplots()
ax.plot([0, 1], [0, 1])
mlflow.log_figure(fig, "figure.png")
# confusion matrix
mlflow.log_confusion_matrix([[1, 2], [3, 4]])
def main():
config = {
"split": {
"test_size": 0.2,
"random_state": 42
},
"model": {
"objective": "binary",
"metric": "auc",
"seed": 42
},
"fit": {
"num_boost_round": 10,
"early_stopping_rounds": 3
},
}
# Prepare training data.
X, y = breast_cancer()
X_train, X_test, y_train, y_test = train_test_split(
X, y, **config["split"])
train_set = lgb.Dataset(X_train, label=y_train)
# Set experiment.
expr_name = "lightgbm"
mlflow.get_or_create_experiment(expr_name) # EX
mlflow.set_experiment(expr_name)
with mlflow.start_run():
# Log training configuration.
mlflow.log_params_flatten(config) # EX
mlflow.log_dict(config, "config.json") # EX
# Train model.
model = lgb.train(config["model"],
train_set,
valid_sets=[train_set],
valid_names=["train"],
**config["fit"])
# Log feature importance.
importance_type = "gain"
features = model.feature_name()
importances = model.feature_importance(importance_type)
mlflow.log_feature_importance(features, importances,
importance_type) # EX
# Log confusion metrics.
mlflow.log_metrics_flatten(model.best_score)
# Log confusion matrix.
y_proba = model.predict(X_test)
cm = confusion_matrix(y_test, y_proba > 0.5)
mlflow.log_confusion_matrix(cm) # EX
def main():
calendar, prices, sales, submission = read_data()
num_items = sales.shape[0]
pred_days = submission.shape[1] - 1 # 28
# encoder calendar and sales cols into numeric
calendar = encoder_category(
calendar, ["event_name_1", "event_type_1", "event_name_2", "event_type_2"
]).pipe(reduce_mem_usage, verbose)
sales = encoder_category(
sales,
["item_id", "dept_id", "cat_id", "store_id", "state_id"],
).pipe(reduce_mem_usage, verbose)
prices = encoder_category(prices,
["item_id", "store_id"]).pipe(reduce_mem_usage, verbose)
data = reshape_sales(sales, submission, pred_days, d_thresh=1941 -
int(365 * 2), verbose=False) # d_thresh why is this
del sales
calendar["d"] = extract_num(calendar["d"])
data = merge_calendar(data, calendar)
data = merge_prices(data, prices)
del calendar, prices
gc.collect()
data = reduce_mem_usage(data)
data = add_demand_features(data, pred_days).pipe(reduce_mem_usage)
data = add_price_features(data).pipe(reduce_mem_usage)
dt_col = "date"
data = add_time_features(data, dt_col).pipe(reduce_mem_usage)
data = data.sort_values("date")
print("start date:", data[dt_col].min())
print("end date:", data[dt_col].max())
print("data shape:", data.shape)
# stage 2
day_col = "d"
cv_params = {"n_splits": 3, "train_days": int(
365 * 1.5), "test_days": pred_days, "day_col": day_col, "pred_days": pred_days}
cv = CustomTimeSeriesSpliter(**cv_params)
sample = data.iloc[::1000][[day_col, dt_col]].reset_index(drop=True)
show_cv_days(cv, sample, dt_col, day_col)
plot_cv_indices(cv, sample, dt_col)
del sample
gc.collect()
features = [
"item_id",
"dept_id",
"cat_id",
"store_id",
"state_id",
"event_name_1",
"event_type_1",
"event_name_2",
"event_type_2",
"snap_CA",
"snap_TX",
"snap_WI",
"sell_price",
# demand features
"shift_t28",
"shift_t29",
"shift_t30",
# std
"rolling_std_t7",
"rolling_std_t30",
"rolling_std_t60",
"rolling_std_t90",
"rolling_std_t180",
# mean
"rolling_mean_t7",
"rolling_mean_t30",
"rolling_mean_t60",
"rolling_mean_t90",
"rolling_mean_t180",
# min
"rolling_min_t7",
"rolling_min_t30",
"rolling_min_t60",
# max
"rolling_max_t7",
"rolling_max_t30",
"rolling_max_t60",
# others
"rolling_skew_t30",
"rolling_kurt_t30",
# price features
"price_change_t1",
"price_change_t365",
"rolling_price_std_t7",
"rolling_price_std_t30",
# time features
"year",
"quarter",
"month",
"week",
"day",
"dayofweek",
"is_weekend",
]
is_train = data["d"] < 1914
# Attach "d" to X_train for cross validation.
x_train = data[is_train][[day_col] + features].reset_index(drop=True)
y_train = data[is_train]["demand"].reset_index(drop=True)
x_test = data[~is_train][features].reset_index(drop=True)
# keep these two columns to use later.
id_date = data[~is_train][["id", "date"]].reset_index(drop=True)
del data
gc.collect()
print("x_train shape:", x_train.shape)
print("x_test shape:", x_test.shape)
bst_params = {"boosting_type": "gbdt", "metric": "rmse",
"objective": "regression", "n_jobs": -1, "seed": 42, "learning_rate": 0.1,
"bagging_fraction": 0.75, "bagging_freq": 10, "colsample_bytree": 0.75}
fit_params = {"num_boost_round": 100_000,
"early_stopping_rounds": 50, "verbose_eval": 100}
models = train_lgb(bst_params, fit_params, x_train,
y_train, cv, drop_when_train=[day_col])
del x_train, y_train
gc.collect()
imp_type = "gain"
importances = np.zeros(x_test.shape[1])
preds = np.zeros(x_test.shape[0])
for model in models:
preds += model.predict(x_test)
importances += model.feature_importance(imp_type)
preds = preds / cv.get_splits()
importances = importances / cv.get_splits()
with mlflow.start_run():
mlflow.log_params_flatten(
{"bst": bst_params, "fit": fit_params, "cv": cv_params})
features = models[0].feature_name()
fig = mplt.feature_importance(features, importances, imp_type, limit=30)
plt.show()
make_submission(id_date.assign(demand=preds), submission, pred_days)
preds = np.zeros(X_test.shape[0])
for model in models:
preds += model.predict(X_test)
importances += model.feature_importance(imp_type)
preds = preds / cv.get_n_splits()
importances = importances / cv.get_n_splits()
#%%
from mlflow_extend import mlflow, plotting as mplt
with mlflow.start_run():
mlflow.log_params_flatten({
"bst": bst_params,
"fit": fit_params,
"cv": cv_params
})
features = models[0].feature_name()
_ = mplt.feature_importance(features, importances, imp_type, limit=30)
#%%
def make_submission(test, submission):
preds = test[["id", "date", "demand"]]
preds = preds.pivot(index="id", columns="date",
values="demand").reset_index()
preds.columns = ["id"] + ["F" + str(d + 1) for d in range(DAYS_PRED)]