From 35c2290a9574fede4d0ceac5b0ac5ee524af29f3 Mon Sep 17 00:00:00 2001
From: Xiaojing Zhang <80235074+zhangxjohn@users.noreply.github.com>
Date: Fri, 11 Mar 2022 09:57:14 +0800
Subject: [PATCH] Rebuild TSFinalTrainStep.
---
hyperts/experiment.py | 4 +++-
hyperts/framework/compete.py | 30 +++++++++++++++++++++++++---
hyperts/framework/dl/models/_base.py | 9 +++++----
hyperts/utils/__init__.py | 2 +-
hyperts/utils/consts.py | 3 ++-
5 files changed, 38 insertions(+), 10 deletions(-)
diff --git a/hyperts/experiment.py b/hyperts/experiment.py
index a777c3c..38d2471 100644
--- a/hyperts/experiment.py
+++ b/hyperts/experiment.py
@@ -282,7 +282,7 @@ def append_early_stopping_callbacks(cbs):
raise ValueError("Forecast task 'timestamp' cannot be None.")
if task in consts.TASK_LIST_FORECAST and covariables is None:
- logger.warning('If the data contains covariables, specify the covariable column names.')
+ logger.info('If the data contains covariables, specify the covariable column names.')
if mode != consts.Mode_STATS:
try:
@@ -358,6 +358,8 @@ def append_early_stopping_callbacks(cbs):
if freq is None:
freq = tb.infer_ts_freq(X_train, ts_name=timestamp)
+ if freq is None:
+ raise RuntimeError('Unable to infer correct frequency, please check data or specify frequency.')
# 7. Covarite Transformer
if covariables is not None:
diff --git a/hyperts/framework/compete.py b/hyperts/framework/compete.py
index 713bc15..651c072 100644
--- a/hyperts/framework/compete.py
+++ b/hyperts/framework/compete.py
@@ -60,10 +60,10 @@ def fit_transform(self, hyper_model, X_train, y_train, X_test=None, X_eval=None,
# 4. eval variables data process
if X_eval is None or y_eval is None:
if self.task in consts.TASK_LIST_FORECAST:
- if X_train.shape[0] <= 2*consts.DEFAULT_FORECAST_EVAL_SIZE or isinstance(self.experiment.eval_size, int):
+ if int(X_train.shape[0]*consts.DEFAULT_MIN_EVAL_SIZE)<=10 or isinstance(self.experiment.eval_size, int):
eval_horizon = self.experiment.eval_size
else:
- eval_horizon = consts.DEFAULT_FORECAST_EVAL_SIZE
+ eval_horizon = consts.DEFAULT_MIN_EVAL_SIZE
X_train, X_eval, y_train, y_eval = \
tb.temporal_train_test_split(X_train, y_train, test_size=eval_horizon)
self.step_progress('split into train set and eval set')
@@ -284,6 +284,30 @@ def get_ensemble(self, estimators, X_train, y_train):
return tb.greedy_ensemble(ensemble_task, estimators, scoring=self.scorer, ensemble_size=self.ensemble_size)
+class TSFinalTrainStep(FinalTrainStep):
+ def __init__(self, experiment, name, mode=None, retrain_on_wholedata=False):
+ super().__init__(experiment, name)
+
+ self.mode = mode
+ self.retrain_on_wholedata = retrain_on_wholedata
+
+ def build_estimator(self, hyper_model, X_train, y_train, X_test=None, X_eval=None, y_eval=None, **kwargs):
+ if self.retrain_on_wholedata:
+ trial = hyper_model.get_best_trial()
+ tb = get_tool_box(X_train, X_eval)
+ X_all = tb.concat_df([X_train, X_eval], axis=0)
+ y_all = tb.concat_df([y_train, y_eval], axis=0)
+
+ if self.mode != consts.Mode_STATS:
+ kwargs.update({'epochs': consts.FINAL_TRAINING_EPOCHS})
+
+ estimator = hyper_model.final_train(trial.space_sample, X_all, y_all, **kwargs)
+ else:
+ estimator = hyper_model.load_estimator(hyper_model.get_best_trial().model_file)
+
+ return estimator
+
+
class TSPipeline:
"""Pipeline Extension for Time Series Analysis.
@@ -731,7 +755,7 @@ def __init__(self, hyper_model, X_train, y_train, X_eval=None, y_eval=None, X_te
# ensemble_size=ensemble_size))
# else:
# final train step
- steps.append(FinalTrainStep(self, consts.StepName_FINAL_TRAINING, retrain_on_wholedata=False))
+ steps.append(TSFinalTrainStep(self, consts.StepName_FINAL_TRAINING, retrain_on_wholedata=True))
# ignore warnings
import warnings
diff --git a/hyperts/framework/dl/models/_base.py b/hyperts/framework/dl/models/_base.py
index 4500b2e..2009301 100644
--- a/hyperts/framework/dl/models/_base.py
+++ b/hyperts/framework/dl/models/_base.py
@@ -191,7 +191,8 @@ def fit(self,
validation_freq=1,
max_queue_size=10,
workers=1,
- use_multiprocessing=False):
+ use_multiprocessing=False,
+ **kwargs):
"""Trains the model for a fixed number of epochs (iterations on a dataset).
Parameters
@@ -666,15 +667,15 @@ def _from_tensor_slices(self, X, y, batch_size, epochs=None, shuffle=False, drop
dataset = tf.data.Dataset.from_tensor_slices((data, y))
- if epochs is not None:
- dataset = dataset.repeat(epochs)
-
if shuffle:
dataset = dataset.shuffle(y.shape[0])
dataset = dataset.batch(batch_size, drop_remainder=drop_remainder and y.shape[0] >= batch_size)
dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)
+ if epochs is not None:
+ dataset = dataset.repeat(epochs+1)
+
return dataset
def _preprocessor(self, X, y):
diff --git a/hyperts/utils/__init__.py b/hyperts/utils/__init__.py
index 1203c3f..128f854 100644
--- a/hyperts/utils/__init__.py
+++ b/hyperts/utils/__init__.py
@@ -26,7 +26,7 @@ def set_random_state(seed=9527, mode=consts.Mode_STATS):
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
- os.environ['TF_DETERMINISTIC_OPS'] = '0'
+ os.environ['TF_DETERMINISTIC_OPS'] = '1'
np.random.seed(seed)
if mode == consts.Mode_DL:
diff --git a/hyperts/utils/consts.py b/hyperts/utils/consts.py
index c66ac53..e6fe4cc 100644
--- a/hyperts/utils/consts.py
+++ b/hyperts/utils/consts.py
@@ -3,8 +3,9 @@
TIMESTAMP = 'timestamp'
DEFAULT_EVAL_SIZE = 0.2
-DEFAULT_FORECAST_EVAL_SIZE = 10
+DEFAULT_MIN_EVAL_SIZE = 0.05
NAN_DROP_SIZE = 0.6
+FINAL_TRAINING_EPOCHS = 120
Task_UNIVARIATE_FORECAST = 'univariate-forecast'
Task_MULTIVARIATE_FORECAST = 'multivariate-forecast'