def _setup(self, config):
# print("config in set up is", config)
global_ft = ray.get(ft_id)
# global_model = ray.get(model_id)
self.trial_ft = deepcopy(global_ft)
self.trial_model = TimeSequenceModel(check_optional_config=False,
future_seq_len=future_seq_len)
# handling input
global_input_df = ray.get(input_df_id)
trial_input_df = deepcopy(global_input_df)
self.config = convert_bayes_configs(config).copy()
(self.x_train, self.y_train) = self.trial_ft.fit_transform(trial_input_df,
**self.config)
# trial_ft.fit(trial_input_df, **config)
# handling validation data
self.validation_data = None
if is_val_df_valid:
global_validation_df = ray.get(validation_df_id)
trial_validation_df = deepcopy(global_validation_df)
self.validation_data = self.trial_ft.transform(trial_validation_df)
# no need to call build since it is called the first time fit_eval is called.
# callbacks = [TuneCallback(tune_reporter)]
# fit model
self.best_reward_m = -999
self.reward_m = -999
self.ckpt_name = "pipeline.ckpt"
def train_func(config, tune_reporter):
# make a copy from global variables for trial to make changes
global_ft = ray.get(ft_id)
# global_model = ray.get(model_id)
trial_ft = deepcopy(global_ft)
# trial_model = deepcopy(global_model)
trial_model = TimeSequenceModel(check_optional_config=False,
future_seq_len=future_seq_len)
# handling input
global_input_df = ray.get(input_df_id)
trial_input_df = deepcopy(global_input_df)
config = convert_bayes_configs(config).copy()
# print("config is ", config)
(x_train, y_train) = trial_ft.fit_transform(trial_input_df, **config)
# trial_ft.fit(trial_input_df, **config)
# handling validation data
validation_data = None
if is_val_df_valid:
global_validation_df = ray.get(validation_df_id)
trial_validation_df = deepcopy(global_validation_df)
validation_data = trial_ft.transform(trial_validation_df)
# no need to call build since it is called the first time fit_eval is called.
# callbacks = [TuneCallback(tune_reporter)]
# fit model
best_reward_m = -999
reward_m = -999
for i in range(1, 101):
result = trial_model.fit_eval(x_train,
y_train,
validation_data=validation_data,
mc=mc,
# verbose=1,
**config)
reward_m = metric_op * result
# if metric == "mean_squared_error":
# reward_m = (-1) * result
# # print("running iteration: ",i)
# elif metric == "r_square":
# reward_m = result
# else:
# raise ValueError("metric can only be \"mean_squared_error\" or \"r_square\"")
ckpt_name = "best.ckpt"
if reward_m > best_reward_m:
best_reward_m = reward_m
save_zip(ckpt_name, trial_ft, trial_model, config)
if remote_dir is not None:
upload_ppl_hdfs(remote_dir, ckpt_name)
tune_reporter(
training_iteration=i,
reward_metric=reward_m,
checkpoint="best.ckpt"
)
def fit_with_fixed_configs(self,
input_df,
validation_df=None,
mc=False,
**user_configs):
"""
Fit pipeline with fixed configs. The model will be trained from initialization
with the hyper-parameter specified in configs. The configs contain both identity configs
(Eg. "future_seq_len", "dt_col", "target_col", "metric") and automl tunable configs
(Eg. "past_seq_len", "batch_size").
We recommend calling get_default_configs to see the name and default values of configs you
you can specify.
:param input_df: one data frame or a list of data frames
:param validation_df: one data frame or a list of data frames
:param user_configs: you can overwrite or add more configs with user_configs. Eg. "epochs"
:return:
"""
# self._check_configs()
if self.config is None:
self.config = self.get_default_configs()
if user_configs is not None:
self.config.update(user_configs)
ft_id_config_set = {
'future_seq_len', 'dt_col', 'target_col', 'extra_features_col',
'drop_missing'
}
ft_id_configs = {a: self.config[a] for a in ft_id_config_set}
self.feature_transformers = TimeSequenceFeatureTransformer(
**ft_id_configs)
model_id_config_set = {'future_seq_len'}
ft_id_configs = {a: self.config[a] for a in model_id_config_set}
self.model = TimeSequenceModel(check_optional_config=False,
**ft_id_configs)
all_available_features = self.feature_transformers.get_feature_list(
input_df)
self.config.update({"selected_features": all_available_features})
(x_train, y_train) = self.feature_transformers.fit_transform(
input_df, **self.config)
if self._is_val_df_valid(validation_df):
validation_data = self.feature_transformers.transform(
validation_df)
else:
validation_data = None
self.model.fit_eval(x_train,
y_train,
validation_data=validation_data,
mc=mc,
verbose=1,
**self.config)
def _hp_search(self,
input_df,
validation_df,
metric,
recipe,
mc,
resources_per_trial,
remote_dir):
ft = TimeSequenceFeatureTransformer(self.future_seq_len,
self.dt_col,
self.target_col,
self.extra_features_col,
self.drop_missing)
if isinstance(input_df, list):
feature_list = ft.get_feature_list(input_df[0])
else:
feature_list = ft.get_feature_list(input_df)
# model = VanillaLSTM(check_optional_config=False)
model = TimeSequenceModel(check_optional_config=False, future_seq_len=self.future_seq_len)
# prepare parameters for search engine
search_space = recipe.search_space(feature_list)
runtime_params = recipe.runtime_params()
num_samples = runtime_params['num_samples']
stop = dict(runtime_params)
search_algorithm_params = recipe.search_algorithm_params()
search_algorithm = recipe.search_algorithm()
fixed_params = recipe.fixed_params()
del stop['num_samples']
searcher = RayTuneSearchEngine(logs_dir=self.logs_dir,
resources_per_trial=resources_per_trial,
name=self.name,
remote_dir=remote_dir,
)
searcher.compile(input_df,
search_space=search_space,
stop=stop,
search_algorithm_params=search_algorithm_params,
search_algorithm=search_algorithm,
fixed_params=fixed_params,
# feature_transformers=TimeSequenceFeatures,
feature_transformers=ft,
# model=model,
future_seq_len=self.future_seq_len,
validation_df=validation_df,
metric=metric,
mc=mc,
num_samples=num_samples)
# searcher.test_run()
searcher.run()
best = searcher.get_best_trials(k=1)[0] # get the best one trial, later could be n
pipeline = self._make_pipeline(best,
feature_transformers=ft,
model=model,
remote_dir=remote_dir)
return pipeline
def load_ts_pipeline(file):
feature_transformers = TimeSequenceFeatureTransformer()
model = TimeSequenceModel(check_optional_config=False)
all_config = restore_zip(file, feature_transformers, model)
ts_pipeline = TimeSequencePipeline(feature_transformers=feature_transformers,
model=model,
config=all_config)
print("Restore pipeline from", file)
return ts_pipeline
class TimeSequencePipeline(Pipeline):
def __init__(self,
feature_transformers=None,
model=None,
config=None,
name=None):
"""
initialize a pipeline
:param model: the internal model
:param feature_transformers: the feature transformers
"""
self.feature_transformers = feature_transformers
self.model = model
self.config = config
self.name = name
self.time = time.strftime("%Y%m%d-%H%M%S")
def describe(self):
init_info = [
'future_seq_len', 'dt_col', 'target_col', 'extra_features_col',
'drop_missing'
]
print("**** Initialization info ****")
for info in init_info:
print(info + ":", self.config[info])
print("")
def fit(self, input_df, validation_df=None, mc=False, epoch_num=20):
x, y = self.feature_transformers.transform(input_df, is_train=True)
if validation_df is not None and not validation_df.empty:
validation_data = self.feature_transformers.transform(
validation_df)
else:
validation_data = None
new_config = {'epochs': epoch_num}
self.model.fit_eval(x,
y,
validation_data,
mc=mc,
verbose=1,
**new_config)
print('Fit done!')
def _is_val_df_valid(self, validation_df):
df_not_empty = isinstance(validation_df,
pd.DataFrame) and not validation_df.empty
df_list_not_empty = isinstance(validation_df, list) \
and validation_df and not all([d.empty for d in validation_df])
if validation_df is not None and (df_not_empty or df_list_not_empty):
return True
else:
return False
def _check_configs(self):
required_configs = {'future_seq_len'}
if not self.config.keys() & required_configs:
raise ValueError("Missing required parameters in configuration. " +
"Required parameters are: " +
str(required_configs))
default_config = {
'dt_col': 'datetime',
'target_col': 'value',
'extra_features_col': None,
'drop_missing': True,
'past_seq_len': 2,
'batch_size': 64,
'lr': 0.001,
'dropout': 0.2,
'epochs': 10,
'metric': 'mse'
}
for config, value in default_config.items():
if config not in self.config:
print('Config: \'{}\' is not specified. '
'A default value of {} will be used.'.format(
config, value))
def get_default_configs(self):
default_configs = {
'dt_col': 'datetime',
'target_col': 'value',
'extra_features_col': None,
'drop_missing': True,
'future_seq_len': 1,
'past_seq_len': 2,
'batch_size': 64,
'lr': 0.001,
'dropout': 0.2,
'epochs': 10,
'metric': 'mean_squared_error'
}
print("**** default config: ****")
for config in default_configs:
print(config + ":", default_configs[config])
print(
"You can change any fields in the default configs by passing into "
"fit_with_fixed_configs(). Otherwise, the default values will be used."
)
return default_configs
def fit_with_fixed_configs(self,
input_df,
validation_df=None,
mc=False,
**user_configs):
"""
Fit pipeline with fixed configs. The model will be trained from initialization
with the hyper-parameter specified in configs. The configs contain both identity configs
(Eg. "future_seq_len", "dt_col", "target_col", "metric") and automl tunable configs
(Eg. "past_seq_len", "batch_size").
We recommend calling get_default_configs to see the name and default values of configs you
you can specify.
:param input_df: one data frame or a list of data frames
:param validation_df: one data frame or a list of data frames
:param user_configs: you can overwrite or add more configs with user_configs. Eg. "epochs"
:return:
"""
# self._check_configs()
if self.config is None:
self.config = self.get_default_configs()
if user_configs is not None:
self.config.update(user_configs)
ft_id_config_set = {
'future_seq_len', 'dt_col', 'target_col', 'extra_features_col',
'drop_missing'
}
ft_id_configs = {a: self.config[a] for a in ft_id_config_set}
self.feature_transformers = TimeSequenceFeatureTransformer(
**ft_id_configs)
model_id_config_set = {'future_seq_len'}
ft_id_configs = {a: self.config[a] for a in model_id_config_set}
self.model = TimeSequenceModel(check_optional_config=False,
**ft_id_configs)
all_available_features = self.feature_transformers.get_feature_list(
input_df)
self.config.update({"selected_features": all_available_features})
(x_train, y_train) = self.feature_transformers.fit_transform(
input_df, **self.config)
if self._is_val_df_valid(validation_df):
validation_data = self.feature_transformers.transform(
validation_df)
else:
validation_data = None
self.model.fit_eval(x_train,
y_train,
validation_data=validation_data,
mc=mc,
verbose=1,
**self.config)
def evaluate(self, input_df, metrics=["mse"], multioutput='raw_values'):
"""
evaluate the pipeline
:param input_df:
:param metrics: subset of ['mean_squared_error', 'r_square', 'sMAPE']
:param multioutput: string in ['raw_values', 'uniform_average']
'raw_values' :
Returns a full set of errors in case of multioutput input.
'uniform_average' :
Errors of all outputs are averaged with uniform weight.
:return:
"""
if isinstance(metrics, str):
metrics = [metrics]
# if not isinstance(metrics, list):
# raise ValueError("Expected metrics to be a list!")
x, y = self.feature_transformers.transform(input_df, is_train=True)
y_pred = self.model.predict(x)
if y_pred.shape[1] == 1:
multioutput = 'uniform_average'
y_unscale, y_pred_unscale = self.feature_transformers.post_processing(
input_df, y_pred, is_train=True)
return [
Evaluator.evaluate(m,
y_unscale,
y_pred_unscale,
multioutput=multioutput) for m in metrics
]
def predict(self, input_df):
"""
predict test data with the pipeline fitted
:param input_df:
:return:
"""
x, _ = self.feature_transformers.transform(input_df, is_train=False)
y_pred = self.model.predict(x)
y_output = self.feature_transformers.post_processing(input_df,
y_pred,
is_train=False)
return y_output
def predict_with_uncertainty(self, input_df, n_iter=100):
x, _ = self.feature_transformers.transform(input_df, is_train=False)
y_pred, y_pred_uncertainty = self.model.predict_with_uncertainty(
x=x, n_iter=n_iter)
y_output = self.feature_transformers.post_processing(input_df,
y_pred,
is_train=False)
y_uncertainty = self.feature_transformers.unscale_uncertainty(
y_pred_uncertainty)
return y_output, y_uncertainty
def save(self, ppl_file=None):
"""
save pipeline to file, contains feature transformer, model, trial config.
:param ppl_file:
:return:
"""
ppl_file = ppl_file or os.path.join(
DEFAULT_PPL_DIR, "{}_{}.ppl".format(self.name, self.time))
save_zip(ppl_file, self.feature_transformers, self.model, self.config)
print("Pipeline is saved in", ppl_file)
return ppl_file
def config_save(self, config_file=None):
"""
save all configs to file.
:param config_file:
:return:
"""
config_file = config_file or os.path.join(
DEFAULT_CONFIG_DIR, "{}_{}.json".format(self.name, self.time))
save_config(config_file, self.config, replace=True)
return config_file
class TrainableClass(Trainable):
def _setup(self, config):
# print("config in set up is", config)
global_ft = ray.get(ft_id)
# global_model = ray.get(model_id)
self.trial_ft = deepcopy(global_ft)
self.trial_model = TimeSequenceModel(check_optional_config=False,
future_seq_len=future_seq_len)
# handling input
global_input_df = ray.get(input_df_id)
trial_input_df = deepcopy(global_input_df)
self.config = convert_bayes_configs(config).copy()
(self.x_train, self.y_train) = self.trial_ft.fit_transform(trial_input_df,
**self.config)
# trial_ft.fit(trial_input_df, **config)
# handling validation data
self.validation_data = None
if is_val_df_valid:
global_validation_df = ray.get(validation_df_id)
trial_validation_df = deepcopy(global_validation_df)
self.validation_data = self.trial_ft.transform(trial_validation_df)
# no need to call build since it is called the first time fit_eval is called.
# callbacks = [TuneCallback(tune_reporter)]
# fit model
self.best_reward_m = -999
self.reward_m = -999
self.ckpt_name = "pipeline.ckpt"
def _train(self):
# print("self.config in train is ", self.config)
result = self.trial_model.fit_eval(self.x_train, self.y_train,
validation_data=self.validation_data,
# verbose=1,
**self.config)
self.reward_m = metric_op * result
# if metric == "mean_squared_error":
# self.reward_m = (-1) * result
# # print("running iteration: ",i)
# elif metric == "r_square":
# self.reward_m = result
# else:
# raise ValueError("metric can only be \"mean_squared_error\" or \"r_square\"")
return {"reward_metric": self.reward_m, "checkpoint": self.ckpt_name}
def _save(self, checkpoint_dir):
# print("checkpoint dir is ", checkpoint_dir)
ckpt_name = self.ckpt_name
# save in the working dir (without "checkpoint_{}".format(training_iteration))
path = os.path.join(checkpoint_dir, "..", ckpt_name)
# path = os.path.join(checkpoint_dir, ckpt_name)
# print("checkpoint save path is ", checkpoint_dir)
if self.reward_m > self.best_reward_m:
self.best_reward_m = self.reward_m
print("****this reward is", self.reward_m)
print("*********saving checkpoint")
save_zip(ckpt_name, self.trial_ft, self.trial_model, self.config)
if remote_dir is not None:
upload_ppl_hdfs(remote_dir, ckpt_name)
return path
def _restore(self, checkpoint_path):
# print("checkpoint path in restore is ", checkpoint_path)
if remote_dir is not None:
restore_hdfs(checkpoint_path, remote_dir, self.trial_ft, self.trial_model)
else:
restore_zip(checkpoint_path, self.trial_ft, self.trial_model)
def model_create_func():
# model = VanillaLSTM(check_optional_config=False)
model = TimeSequenceModel(
check_optional_config=False,
future_seq_len=self.future_seq_len)
return model
