def objective(self, trial):
# Extract optuna attribs from the input json
optuna_trn_params = {}
for key, val in self.params["trn_params"].items():
if type(val) != list:
optuna_trn_params[key] = val
else:
if type(val[0]) == float:
optuna_trn_params[key] = trial.suggest_uniform(
key, val[0], val[1])
elif type(val[0]) == int:
optuna_trn_params[key] = trial.suggest_int(
key, val[0], val[1])
else:
optuna_trn_params[key] = trial.suggest_categorical(
key, val)
start = time.time()
getLogger(get_version()).info(
"\t [OPTUNA] {}th optimization starts".format(
self.optimized_count))
send_message("\t [OPTUNA] :sushi: {} th optimization starts".format(
self.optimized_count))
# Classify
mtd_params = self.params["mtd_params"]
clf = lgb.train(
optuna_trn_params,
self.train_dataset,
mtd_params["num_boost_round"],
valid_sets=[self.train_dataset, self.valid_dataset],
feval=eval_auc,
verbose_eval=mtd_params["verbose_eval"],
early_stopping_rounds=mtd_params["early_stopping_rounds"])
getLogger(get_version()).info("\t {}".format(clf.params))
send_message("\t {}".format(clf.params))
for train_or_valid, metrics in clf.best_score.items():
for metric, score in metrics.items():
getLogger(get_version()).info("\t\t >> Best {} {}: {}".format(
train_or_valid, metric, score))
send_message("\t\t :star-struck: Best {} {}: {}".format(
train_or_valid, metric, score))
# Post-process this fold
elapsed_time = int(time.time() - start)
minutes, sec = divmod(elapsed_time, 60)
hour, minutes = divmod(minutes, 60)
getLogger(get_version()).info(
"\t [OPTUNA] >> {}th optimization finishes: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}"
.format(self.optimized_count, hour, minutes, sec))
send_message(
"\t [OPTUNA] :sushi: {}th optimiaztion finishes: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}"
.format(self.optimized_count, hour, minutes, sec))
self.optimized_count += 1
return clf.best_score["valid_1"]["binary_logloss"]
def save_feature_importance(feature_importance, directory_path):
feature_importance.set_index("feature", inplace=True)
feature_importance["median"] = feature_importance.median(axis='columns')
feature_importance.sort_values("median", ascending=False, inplace=True)
Path.mkdir(directory_path, exist_ok=True, parents=True)
feature_importance.to_csv(
Path(directory_path / "{}.csv".format(get_version())))
def train(self, feature_names):
self.feature_names = feature_names
fold = 0
valid = "valid{}".format(str(fold))
trn_x = super().get_feature_df(self.feature_names, valid, "train")
val_x = super().get_feature_df(self.feature_names, valid, "validate")
trn_x.set_index("MachineIdentifier", inplace=True)
val_x.set_index("MachineIdentifier", inplace=True)
trn_y = trn_x["HasDetections"].astype(np.int8)
val_y = val_x["HasDetections"].astype(np.int8)
del trn_x["HasDetections"], val_x["HasDetections"]
self.train_dataset = lgb.Dataset(trn_x, trn_y)
self.valid_dataset = lgb.Dataset(val_x, val_y)
self.optimized_count = 0
study = optuna.create_study()
study.optimize(self.objective, n_trials=self.params["n_trials"])
best_trn_params = study.best_params
getLogger(get_version()).info(
"\t >> Best params: {}".format(best_trn_params))
send_message("\t :youzyo: Best params: {}".format(best_trn_params))
for key, val in self.params["trn_params"].items():
if type(val) != list:
best_trn_params[key] = val
self.params["trn_params"] = best_trn_params
del self.train_dataset, self.valid_dataset
gc.collect()
self.best_lgbm_classifier = LGBMClassifier(self.params,
self.dataset_name)
return self.best_lgbm_classifier.train(feature_names)
def predict(self, feature_names):
"""
Input:
feature_names: directionary of features' names
Output:
predict_df: Dataframe(["MachineIdentifier", "HasDetections")
"""
model_directory_path = Path(__file__).absolute(
).parents[2] / "data" / "model" / str(get_version())
preds = None
FOLDS = 5
predict_df = None
for fold in range(FOLDS):
model_path = model_directory_path / "valid{}.model".format(fold)
model = torch.load(str(model_path))
valid = "valid{}".format(fold)
test_df = super().get_feature_df(feature_names, valid, "test")
if predict_df is None:
predict_df = test_df["MachineIdentifier"]
test_df = test_df.set_index("MachineIdentifier")
normalized_test_df = (test_df - test_df.min()) / (test_df.max() -
test_df.min())
if preds is None:
preds = predict_with_model(self.params, normalized_test_df,
model) / FOLDS
else:
preds += predict_with_model(self.params, normalized_test_df,
model) / FOLDS
predict_df = pd.DataFrame(predict_df)
predict_df["HasDetections"] = preds
return predict_df
def predict(self, feature_names):
"""
Input:
feature_names: directionary of features' names
Output:
predict_df: Dataframe(["MachineIdentifier", "HasDetections")
"""
model_directory_path = Path(__file__).absolute(
).parents[2] / "data" / "model" / str(get_version())
preds = None
FOLDS = 5
predict_df = None
for fold in range(FOLDS):
model_path = model_directory_path / "valid{}.model".format(fold)
clf = lgb.Booster(model_file=str(model_path))
valid = "valid{}".format(fold)
test_df = super().get_feature_df(feature_names, valid, "test")
if predict_df is None:
predict_df = test_df["MachineIdentifier"]
test_df = test_df.set_index("MachineIdentifier")
if preds is None:
preds = predict_chunk(clf, test_df) / FOLDS
else:
preds += predict_chunk(clf, test_df) / FOLDS
predict_df = pd.DataFrame(predict_df)
predict_df["HasDetections"] = preds
return predict_df
def __save_outputs(cls, outputs, is_train):
"""
is_train = True -> Save ground truth and validity of train dataset
is_train = False -> Save predictions of test dataset
"""
if is_train:
columns_order = ["MachineIdentifier", "HasDetections", "Predict"]
save_path = cls.ROOT_PATH / "data" / "oof"
else:
columns_order = ["MachineIdentifier", "HasDetections"]
save_path = cls.ROOT_PATH / "data" / "submit"
Path.mkdir(save_path, exist_ok=True, parents=True)
save_df = None
for output in outputs:
if save_df is None:
save_df = output
else:
save_df = pd.concat([save_df, output])
submission_df = pd.read_csv(cls.ROOT_PATH / "input" /
"sample_submission.csv")
del submission_df["HasDetections"]
if is_train is False:
predict_length = len(save_df)
sub_length = len(submission_df)
if predict_length != sub_length:
getLogger(get_version()).info(
"CAUSION: Length of predict_df ({}) is NOT equal to that of submisson_df ({})"
.format(predict_length, sub_length))
send_message(
":ghost: CAUSION: Length of predict_df ({}) is NOT equal to that of submisson_df ({})"
.format(predict_length, sub_length))
save_df = pd.merge(save_df,
submission_df,
on="MachineIdentifier",
how='right')
save_df.fillna(0, inplace=True)
save_df = save_df.sort_values("MachineIdentifier").reset_index(
drop=True).loc[:, columns_order]
filename = save_path / "{}.csv".format(get_version())
save_df.to_csv(filename, index=False, float_format="%.6f")
getLogger(get_version()).info("Output {}.".format(filename))
send_message("Output {}".format(filename))
return save_df
def model_exist():
model_path = Path(__file__).absolute().parents[1] / "data" / "model" / str(
get_version())
for fold in range(5):
if Path(model_path /
"valid{}.model".format(fold)).exists() is not True:
return False
return True
def output_cv(validity, stamp):
validity = validity.reset_index()
columns_order = ["MachineIdentifier", "HasDetections", "Predict"]
validity = validity.sort_values("MachineIdentifier").reset_index(
drop=True).loc[:, columns_order]
cv_auc = (fast_auc(validity["HasDetections"],
np.array(validity["Predict"])))
getLogger(get_version()).info("\t >> CV Score (AUC):{}".format(cv_auc))
send_message("\t {} CV Score (AUC):{}".format(stamp, cv_auc))
return validity
def get_feature_df(self, feature_names, valid_dir, part):
"""
Ex)
dataset_name : min, ...
feature_names : __preprocess() return
valid_dir : "valid0", "valid1", ...
part : "train", "validate", "test"
"""
feature_df = None
feature_set_path = Path(__file__).absolute(
).parents[2] / "data" / "features" / self.dataset_name / valid_dir
if dask_mode():
print("Using dask.dataframe.read_csv()")
for group, feature_list in feature_names.items():
getLogger(get_version()).info(
"\t \t \t Reading {}_{}.csv...".format(part, group))
send_message("\t \t \t Reading {}_{}.csv...".format(part, group))
if dask_mode():
df = dd.read_csv(feature_set_path /
"{}_{}.csv".format(part, group),
usecols=["MachineIdentifier"] + feature_list)
df = df.compute()
else:
df = pd.read_csv(feature_set_path /
"{}_{}.csv".format(part, group),
usecols=["MachineIdentifier"] + feature_list)
if feature_df is None:
feature_df = df
else:
feature_df = feature_df.merge(right=df,
how="inner",
on="MachineIdentifier")
if part in ["train", "validate"]:
if dask_mode():
HasDetections = dd.read_csv(
Path(__file__).absolute().parents[2] / "input" /
"train.csv",
usecols=["MachineIdentifier", "HasDetections"])
HasDetections = HasDetections.compute()
else:
HasDetections = pd.read_csv(
Path(__file__).absolute().parents[2] / "input" /
"train.csv",
usecols=["MachineIdentifier", "HasDetections"])
feature_df = feature_df.merge(right=HasDetections,
on="MachineIdentifier",
how="inner")
return feature_df
def __print_log(cls, name, clf_type, dataset_path):
version = get_version()
text = "Classifier Type: {}".format(clf_type)
getLogger(version).info(text)
send_message(text)
text = "Input Path: {}".format(dataset_path)
getLogger(version).info(text)
send_message(text)
text = "--------------------------------"
getLogger(version).info(text)
send_message(text)
def train(self, feature_names):
self.feature_names = feature_names
self.optimized_count = 0
study = optuna.create_study()
study.optimize(self.objective, n_trials=self.params["n_trials"])
best_trn_params = study.best_params
getLogger(get_version()).info("\t >> Best params: {}".format(best_trn_params))
send_message("\t :youzyo: Best params: {}".format(best_trn_params))
for key, val in self.params["trn_params"].items():
if type(val) != list:
best_trn_params[key] = val
self.params["trn_params"] = best_trn_params
self.best_lgbm_classifier = LGBMClassifier(self.params, self.dataset_name)
return self.best_lgbm_classifier.train(feature_names)
def process(cls, args):
"""
main method
"""
# Split the config file ({version}.json) into each variable
config_file = cls.__load_config(args.version)
with config_file.open() as f:
params_dict = json.load(f)
validities = []
predicts = []
trained_fully_before_processing = False
for name, params in params_dict.items():
feature_groups = params["Preprocess"]
clf_type = params["Classifier"]
clf_params = params["ClassifierParams"]
dataset_path = cls.ROOT_PATH / params["DatasetPath"]
ProcessorFactory.__print_log(name, clf_type, dataset_path)
processor = Processor(feature_groups,
ClassifierType.parseType(clf_type),
clf_params, dataset_path)
validity, predict = processor.process()
if validity is None:
trained_fully_before_processing = True
else:
validities.append(validity)
predicts.append(predict)
# Output total CV Score
if trained_fully_before_processing is False and only_prediction(
) is False:
valid_df = cls.__save_outputs(validities, is_train=True)
total_cv = fast_auc(valid_df["HasDetections"],
np.array(valid_df["Predict"]))
getLogger(get_version()).info(
"\t >> Total CV Score (AUC): {}".format(total_cv))
send_message(
"\t :youzyo: Total CV Score (AUC): {}".format(total_cv))
# Output submit file
if need_prediction() or only_prediction():
cls.__save_outputs(predicts, is_train=False)
def train(self, feature_names):
"""
FLow:
1. Initialize parameters
2. Process for each fold
2.1 Generate dataset
2.2 Classify
2.3 calculate feature importances
3. Output CV Score and features importances
4. Predict training data (validate all data)
Input:
feature_names: directionary of features' names
Output:
validity: Dataframe(["MachineIdentifier", "HasDetections", "Predict")
"""
# Initialize parameters
mtd_params = self.params["mtd_params"]
validity = None
model_path = Path(__file__).absolute(
).parents[2] / "data" / "model" / str(get_version())
Path.mkdir(model_path, exist_ok=True, parents=True)
feature_importance = pd.DataFrame()
START_FOLD = 0
if get_back_training():
START_FOLD = len(list(model_path.glob('**/*.model')))
END_FOLD = 5
if train_one_round():
START_FOLD = 0
END_FOLD = 1
if START_FOLD == END_FOLD:
return None
# Process for each fold
for fold in range(START_FOLD, END_FOLD):
# Measure start time of the classification of this fold
start = time.time()
getLogger(get_version()).info("\t >> {} folds start".format(fold))
send_message("\t :flashlight: {} folds start".format(fold))
# Generate dataset
getLogger(get_version()).info("\t \t Generating datasets...")
send_message("\t \t Generating datasets...")
valid = "valid{}".format(str(fold))
trn_x = super().get_feature_df(feature_names, valid, "train")
val_x = super().get_feature_df(feature_names, valid, "validate")
trn_x.set_index("MachineIdentifier", inplace=True)
val_x.set_index("MachineIdentifier", inplace=True)
trn_y = trn_x["HasDetections"].astype(np.int8)
val_y = val_x["HasDetections"].astype(np.int8)
train_dataset = lgb.Dataset(trn_x, trn_y)
valid_dataset = lgb.Dataset(val_x, val_y)
getLogger(get_version()).info("\t \t Datasets were generated.")
send_message("\t \t Datasets were generated.")
# Initialize variables for scoring
if validity is None:
validity = pd.DataFrame()
validity["HasDetections"] = pd.concat([trn_y, val_y])
validity["Predict"] = 0
# Delete needless features
del trn_x["HasDetections"], val_x["HasDetections"]
# Classify
callbacks = [
log_evaluation(get_training_logger(get_version()), fold)
]
clf = lgb.train(
self.params["trn_params"],
train_dataset,
mtd_params["num_boost_round"],
valid_sets=[train_dataset, valid_dataset],
feval=eval_auc,
verbose_eval=mtd_params["verbose_eval"],
early_stopping_rounds=mtd_params["early_stopping_rounds"],
callbacks=callbacks)
validity.loc[validity.index.isin(val_x.index),
"Predict"] = clf.predict(
val_x, num_iteration=clf.best_iteration)
for train_or_valid, metrics in clf.best_score.items():
for metric, score in metrics.items():
getLogger(get_version()).info(
"\t\t >> Best {} {}: {}".format(
train_or_valid, metric, score))
send_message("\t\t :star-struck: Best {} {}: {}".format(
train_or_valid, metric, score))
# Calculate feature importance per fold
if fold == 0:
feature_importance["feature"] = trn_x.columns
feature_importance["fold{}".format(fold)] = clf.feature_importance(
importance_type="gain")
# Measure finish time of the classification of this fold
elapsed_time = int(time.time() - start)
minutes, sec = divmod(elapsed_time, 60)
hour, minutes = divmod(minutes, 60)
getLogger(get_version()).info(
"\t >> {} folds finish: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}"
.format(fold, hour, minutes, sec))
send_message(
"\t :flashlight: {} folds finish: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}"
.format(fold, hour, minutes, sec))
# Post-process this fold
del train_dataset, valid_dataset
gc.collect()
clf.save_model(str(model_path / "valid{}.model".format(fold)))
# Output CV score
validity = output_cv(validity, ":flashlight:")
# Save importance
directory_path = Path(__file__).absolute().parents[2] / "importance"
save_feature_importance(feature_importance, directory_path)
# Post-process the training
del feature_importance
gc.collect()
return validity
def objective(self, trial):
# Extract optuna attribs from the input json
optuna_trn_params = {}
for key, val in self.params["trn_params"].items():
if type(val) != list:
optuna_trn_params[key] = val
else:
if type(val[0]) == float:
optuna_trn_params[key] = trial.suggest_uniform(key, val[0], val[1])
elif type(val[0]) == int:
optuna_trn_params[key] = trial.suggest_int(key, val[0], val[1])
else:
optuna_trn_params[key] = trial.suggest_categorical(key, val)
# Initialize parameters
mtd_params = self.params["mtd_params"]
validity = None
model_path = Path(__file__).absolute().parents[2] / "data" / "model" / str(get_version())
Path.mkdir(model_path, exist_ok=True, parents=True)
START_FOLD = 0
if get_back_training():
START_FOLD = len(list(model_path.glob('**/*.model')))
END_FOLD = 5
if train_one_round():
START_FOLD = 0
END_FOLD = 1
if START_FOLD == END_FOLD:
return None
start2 = time.time()
getLogger(get_version()).info("\t [OPTUNA] {}th optimization starts".format(self.optimized_count))
send_message("\t [OPTUNA] :sushi: {} th optimization starts".format(self.optimized_count))
# Process for each fold
for fold in range(START_FOLD, END_FOLD):
start = time.time()
getLogger(get_version()).info("\t [OPTUNA] >> {} folds start".format(fold))
send_message("\t [OPTUNA] :sushi: {} folds start".format(fold))
# Generate dataset
valid = "valid{}".format(str(fold))
trn_x = super().get_feature_df(self.feature_names, valid, "train")
val_x = super().get_feature_df(self.feature_names, valid, "validate")
trn_x.set_index("MachineIdentifier", inplace=True)
val_x.set_index("MachineIdentifier", inplace=True)
trn_y = trn_x["HasDetections"].astype(np.int8)
val_y = val_x["HasDetections"].astype(np.int8)
train_dataset = lgb.Dataset(trn_x, trn_y)
valid_dataset = lgb.Dataset(val_x, val_y)
# Initialize variables for scoring
if validity is None:
validity = pd.DataFrame()
validity["HasDetections"] = pd.concat([trn_y, val_y])
validity["Predict"] = 0
# Delete needless features
del trn_x["HasDetections"], val_x["HasDetections"]
# Classify
clf = lgb.train(optuna_trn_params,
train_dataset,
mtd_params["num_boost_round"],
valid_sets=[train_dataset, valid_dataset],
feval=eval_auc,
verbose_eval=mtd_params["verbose_eval"],
early_stopping_rounds=mtd_params["early_stopping_rounds"])
validity.loc[validity.index.isin(val_x.index), "Predict"] = clf.predict(val_x, num_iteration=clf.best_iteration)
if fold == START_FOLD:
getLogger(get_version()).info("\t {}".format(clf.params))
send_message("\t {}".format(clf.params))
for train_or_valid, metrics in clf.best_score.items():
for metric, score in metrics.items():
getLogger(get_version()).info("\t\t >> Best {} {}: {}".format(train_or_valid, metric, score))
send_message("\t\t :star-struck: Best {} {}: {}".format(train_or_valid, metric, score))
# Post-process this fold
del train_dataset, valid_dataset
gc.collect()
elapsed_time = int(time.time() - start)
minutes, sec = divmod(elapsed_time, 60)
hour, minutes = divmod(minutes, 60)
getLogger(get_version()).info(
"\t [OPTUNA] >> {} folds finish: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}"
.format(fold, hour, minutes, sec))
send_message("\t [OPTUNA] :sushi: {} folds finish: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}".format(fold, hour, minutes, sec))
elapsed_time = int(time.time() - start2)
minutes, sec = divmod(elapsed_time, 60)
hour, minutes = divmod(minutes, 60)
getLogger(get_version()).info(
"\t [OPTUNA] >> {}th optimization finishes: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}"
.format(self.optimized_count, hour, minutes, sec))
send_message("\t [OPTUNA] :sushi: {}th optimiaztion finishes: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}".format(self.optimized_count, hour, minutes, sec))
self.optimized_count += 1
# Output CV score
validity = validity.reset_index()
columns_order = ["MachineIdentifier", "HasDetections", "Predict"]
validity = validity.sort_values("MachineIdentifier").reset_index(drop=True).loc[:, columns_order]
cv_auc = (fast_auc(validity["HasDetections"], np.array(validity["Predict"])))
return 1 - cv_auc
def train(self, feature_names):
# Initialize parameters
validity = None
model_path = Path(__file__).absolute(
).parents[2] / "data" / "model" / str(get_version())
Path.mkdir(model_path, exist_ok=True, parents=True)
START_FOLD = 0
if get_back_training():
START_FOLD = len(list(model_path.glob('**/*.model')))
END_FOLD = 5
if train_one_round():
START_FOLD = 0
END_FOLD = 1
if START_FOLD == END_FOLD:
return None
get_training_logger(get_version()).debug("fold \t iteration \
\t train BCELoss \t valid BCELoss \
\t train AUC \t valid AUC")
# Process for each fold
for fold in range(START_FOLD, END_FOLD):
# Measure start time of the classification of this fold
start = time.time()
getLogger(get_version()).info("\t >> {} folds start".format(fold))
send_message("\t :fire: {} folds start".format(fold))
valid = "valid{}".format(str(fold))
# Generate train data
getLogger(get_version()).info("\t \t Generating datasets...")
send_message("\t \t Generating datasets...")
trn_x = super().get_feature_df(feature_names, valid, "train")
trn_x.set_index("MachineIdentifier", inplace=True)
trn_y = trn_x["HasDetections"].values.astype(np.float32)
val_x = super().get_feature_df(feature_names, valid, "validate")
val_x.set_index("MachineIdentifier", inplace=True)
val_y = val_x["HasDetections"].values.astype(np.float32)
# Initialize variables for scoring
if validity is None:
validity = pd.DataFrame()
validity["HasDetections"] = pd.concat([
trn_x["HasDetections"].astype(np.float32),
val_x["HasDetections"].astype(np.float32)
])
validity["Predict"] = 0
del trn_x["HasDetections"], val_x["HasDetections"]
trn_x.fillna(0, inplace=True)
val_x.fillna(0, inplace=True)
normalized_trn_x = (trn_x - trn_x.min()) / (trn_x.max() -
trn_x.min())
normalized_trn_x.fillna(normalized_trn_x.mean(), inplace=True)
normalized_val_x = (val_x - val_x.min()) / (val_x.max() -
val_x.min())
normalized_val_x.fillna(normalized_val_x.mean(), inplace=True)
train_loader, valid_loader, dataset_sizes = create_tensor_dataloader(
self.params, normalized_trn_x, trn_y, normalized_val_x, val_y)
data_loaders = {"train": train_loader, "valid": valid_loader}
getLogger(get_version()).info("\t \t Datasets were generated.")
send_message("\t \t Datasets were generated.")
# Define the Network
num_epochs = self.params["num_epochs"]
network_class = getattr(
import_module("classifier.pytorch_network." +
self.params["network"]), self.params["network"])
network = network_class(len(trn_x.columns),
self.params["network_params"])
criterion = nn.BCELoss() # Don't change the criterion function!
optimizer_class = getattr(import_module('torch.optim'),
self.params["optimizer"])
optimizer = optimizer_class(network.parameters(),
**self.params["optimizer_params"])
scheduler_class = getattr(
import_module('torch.optim.lr_scheduler'),
self.params["scheduler"])
scheduler = scheduler_class(optimizer,
**self.params["scheduler_params"])
model = train_model(
fold,
data_loaders,
dataset_sizes,
network,
criterion,
optimizer,
scheduler,
num_epochs=num_epochs,
early_stopping_rounds=self.params["early_stopping_rounds"],
verbose=self.params["verbose"])
# Classify
validity.loc[validity.index.isin(val_x.index),
"Predict"] = predict_with_model(
self.params, normalized_val_x, model)
# Measure finish time of the classification of this fold
elapsed_time = int(time.time() - start)
minutes, sec = divmod(elapsed_time, 60)
hour, minutes = divmod(minutes, 60)
getLogger(get_version()).info(
"\t >> {} folds finish: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}"
.format(fold, hour, minutes, sec))
send_message(
"\t :fire: {} folds finish: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}"
.format(fold, hour, minutes, sec))
# Post-process this fold
gc.collect()
torch.save(model, str(model_path / "valid{}.model".format(fold)))
# Output CV score
validity = validity.reset_index()
columns_order = ["MachineIdentifier", "HasDetections", "Predict"]
validity = validity.sort_values("MachineIdentifier").reset_index(
drop=True).loc[:, columns_order]
cv_auc = (fast_auc(validity["HasDetections"],
np.array(validity["Predict"])))
getLogger(get_version()).info("\t >> CV Score (AUC):{}".format(cv_auc))
send_message("\t :fire: CV Score (AUC):{}".format(cv_auc))
# Post-process the training
gc.collect()
return validity
from save_log import create_main_logger, create_train_logger, get_version, send_message
from processor_factory import ProcessorFactory
from exceptions import *
warnings.filterwarnings('ignore')
def main(args):
send_message(
":thinking_face: ============= {} ============= :thinking_face:".
format(str(datetime.now())))
ProcessorFactory.process(args)
if __name__ == "__main__":
gc.enable()
version = get_version()
create_main_logger(version)
create_train_logger(version)
try:
main(get_option())
except DuplicateVersionException:
send_message(":stop: Duplicate Version Exception Occurred.")
getLogger(version).exception("Duplicate Version Exception Occurred.")
except IrregularArgumentException:
send_message(":stop: Irregular Argument for Feature Extraction.")
getLogger(version).exception(
"Irregular Argument for Feature Extraction.")
except IrregularCalcBackException:
send_message(":stop: Irregular Dataframe back.")
getLogger(version).exception("Irregular Dataframe back.")
except Exception:
def train_model(fold, data_loaders, dataset_sizes, model,
criterion, optimizer, scheduler,
num_epochs, early_stopping_rounds=10, verbose=-1):
since = time.time()
best_score = {
"train BCELoss": 1.0,
"train AUC": 0.5,
"valid BCELoss": 1.0,
"valid AUC": 0.5}
best_model_wts = copy.deepcopy(model.state_dict())
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = model.to(device)
not_improve_round = 0
for epoch in range(num_epochs):
bce_dictionary = {"train": 0.0, "valid": 0.0}
auc_dictionary = {"train": 0.0, "valid": 0.0}
for phase in ["train", "valid"]:
if phase == "train":
scheduler.step()
model.train()
else:
model.eval()
running_bce = 0.0
running_auc = 0.0
# Iteration
for inputs, labels in data_loaders[phase]:
inputs = inputs.to(device)
labels = labels.to(device)
optimizer.zero_grad()
# forward
with torch.set_grad_enabled(phase == "train"):
outputs = model(inputs)
loss = criterion(outputs, labels)
# backward + optimize
if phase == "train":
loss.backward()
optimizer.step()
# statistics
running_bce += loss.item() * inputs.size(0)
if torch.cuda.is_available():
labels = labels.cpu()
outputs = outputs.cpu()
y_true = labels.numpy()
y_pred = list(chain.from_iterable(outputs.detach().numpy()))
running_auc += fast_auc(y_true, y_pred) * inputs.size(0)
bce_dictionary[phase] = running_bce / dataset_sizes[phase]
auc_dictionary[phase] = running_auc / dataset_sizes[phase]
# Update model
if best_score["train BCELoss"] > bce_dictionary["train"]:
best_score["train BCELoss"] = bce_dictionary["train"]
best_score["valid BCELoss"] = bce_dictionary["valid"]
best_score["train AUC"] = auc_dictionary["train"]
best_score["valid AUC"] = auc_dictionary["valid"]
best_model_wts = copy.deepcopy(model.state_dict())
not_improve_round = 0
else:
not_improve_round += 1
if not_improve_round >= early_stopping_rounds:
getLogger(get_version()).info(
"\t \t Epoch {}/{}: Early stopping".format(epoch, num_epochs))
send_message(
"\t \t :upura: Epoch {}/{}: Early stopping".format(epoch, num_epochs))
break
if epoch % verbose == 0 and verbose != -1:
getLogger(get_version()).info("{}\t{}\t{}\t{}\t{}\t{}".format(
fold, epoch, bce_dictionary["train"], bce_dictionary["valid"],
auc_dictionary["train"], auc_dictionary["valid"]))
send_message("{}\t{}\t{}\t{}\t{}\t{}".format(
fold, epoch, bce_dictionary["train"], bce_dictionary["valid"],
auc_dictionary["train"], auc_dictionary["valid"]))
get_training_logger(get_version()).debug("{}\t{}\t{}\t{}\t{}\t{}".format(
fold, epoch, bce_dictionary["train"], bce_dictionary["valid"],
auc_dictionary["train"], auc_dictionary["valid"]))
time_elapsed = time.time() - since
getLogger(get_version()).info(
"\t \t Training complete in {:.0f}m {:.0f}s".format(time_elapsed // 60, time_elapsed % 60))
send_message(
"\t \t Training complete in {:.0f}m {:.0f}s".format(time_elapsed // 60, time_elapsed % 60))
getLogger(get_version()).info(
"\t \t [Best train scores] BCELoss: {:.4f}, AUC: {:.4f}".format(
best_score["train BCELoss"], best_score["train AUC"]))
getLogger(get_version()).info(
"\t \t [Valid scores when train is the best] BCELoss: {:.4f}, AUC: {:.4f}".format(
best_score["valid BCELoss"], best_score["valid AUC"]))
send_message("\t \t :star-struck: Best train BCELoss: {:.4f}".format(best_score["train BCELoss"]))
send_message("\t \t :star-struck: Best train AUC: {:.4f}".format(best_score["train AUC"]))
send_message("\t \t :star-struck: Best valid BCELoss: {:.4f}".format(best_score["valid BCELoss"]))
send_message("\t \t :star-struck: Best valid AUC: {:.4f}".format(best_score["valid AUC"]))
model.load_state_dict(best_model_wts)
return model
def train(self, feature_names):
"""
Input:
feature_names: directionary of features' names
Output:
validity: Dataframe(["MachineIdentifier", "HasDetections", "Predict")
"""
# Initialize parameters
validity = None
model_path = Path(__file__).absolute().parents[2] / "data" / "model" / str(get_version())
Path.mkdir(model_path, exist_ok=True, parents=True)
feature_importance = pd.DataFrame()
START_FOLD = 0
if get_back_training():
START_FOLD = len(list(model_path.glob('**/*.model')))
END_FOLD = 5
if train_one_round():
START_FOLD = 0
END_FOLD = 1
if START_FOLD == END_FOLD:
return None
# Process for each fold
for fold in range(START_FOLD, END_FOLD):
log_path = Path(__file__).absolute().parents[2] / "log" / "train" / str(get_version()) / str("fold{}".format(fold))
Path.mkdir(log_path, exist_ok=True, parents=True)
# Measure start time of the classification of this fold
start = time.time()
getLogger(get_version()).info("\t >> {} folds start".format(fold))
send_message("\t :cat: {} folds start".format(fold))
# Generate dataset
getLogger(get_version()).info("\t \t Generating datasets...")
send_message("\t \t Generating datasets...")
valid = "valid{}".format(str(fold))
trn_x = super().get_feature_df(feature_names, valid, "train")
val_x = super().get_feature_df(feature_names, valid, "validate")
trn_x.set_index("MachineIdentifier", inplace=True)
val_x.set_index("MachineIdentifier", inplace=True)
trn_y = trn_x["HasDetections"].astype(np.int8)
val_y = val_x["HasDetections"].astype(np.int8)
getLogger(get_version()).info("\t \t Datasets were generated.")
send_message("\t \t Datasets were generated.")
# Initialize variables for scoring
if validity is None:
validity = pd.DataFrame()
validity["HasDetections"] = pd.concat([trn_y, val_y])
validity["Predict"] = 0
# Delete needless features
del trn_x["HasDetections"], val_x["HasDetections"]
# Classify
clf = CatBoostClassifier(iterations=self.params["iterations"],
verbose=self.params["verbose"],
early_stopping_rounds=self.params["early_stopping_rounds"],
random_seed=self.params["random_seed"],
max_depth=self.params["max_depth"],
loss_function=self.params["loss_function"],
custom_metric=self.params["custom_metric"],
eval_metric=self.params["eval_metric"],
rsm=self.params["rsm"],
train_dir=str(log_path))
clf.fit(trn_x.values, trn_y.values,
eval_set=(val_x.values, val_y.values))
for train_or_valid, metrics in clf.best_score_.items():
for metric, score in metrics.items():
getLogger(get_version()).info("\t\t >> Best {} {}: {}".format(train_or_valid, metric, score))
send_message("\t\t :star-struck: Best {} {}: {}".format(train_or_valid, metric, score))
validity.loc[validity.index.isin(val_x.index), "Predict"] = clf.predict_proba(val_x.values)[:, 1]
# Calculate feature importance per fold
if fold == 0:
feature_importance["feature"] = trn_x.columns
feature_importance["fold{}".format(fold)] = clf.get_feature_importance()
# Measure finish time of the classification of this fold
elapsed_time = int(time.time() - start)
minutes, sec = divmod(elapsed_time, 60)
hour, minutes = divmod(minutes, 60)
getLogger(get_version()).info(
"\t >> {} folds finish: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}"
.format(fold, hour, minutes, sec))
send_message("\t :cat: {} folds finish: [elapsed_time] >> {:0>2}:{:0>2}:{:0>2}".format(fold, hour, minutes, sec))
# Post-process this fold
clf.save_model(str(model_path / "valid{}.model".format(fold)))
# Output CV score
validity = output_cv(validity, ":cat:")
# Save importance
directory_path = Path(__file__).absolute().parents[2] / "importance"
save_feature_importance(feature_importance, directory_path)
# Post-process the training
del feature_importance
gc.collect()
return validity
