def _make_pipeline(self, analysis, feature_transformers, model_create_fn,
remote_dir):
metric = "reward_metric"
best_config = analysis.get_best_config(metric=metric, mode="max")
best_logdir = analysis.get_best_logdir(metric=metric, mode="max")
print("best log dir is ", best_logdir)
dataframe = analysis.dataframe(metric=metric, mode="max")
# print(dataframe)
model_path = os.path.join(best_logdir, dataframe["checkpoint"].iloc[0])
config = convert_bayes_configs(best_config).copy()
self._print_config(config)
model = model_create_fn(config=config)
if remote_dir is not None:
all_config = restore_hdfs(
model_path,
remote_dir,
feature_transformers,
model,
# config)
)
else:
all_config = restore_zip(
model_path,
feature_transformers,
model,
# config)
)
return TimeSequencePipeline(name=self.name,
feature_transformers=feature_transformers,
model=model,
config=all_config)
def _make_pipeline(self, trial, feature_transformers, model):
isinstance(trial, TrialOutput)
# TODO we need to save fitted parameters (not in config, e.g. min max for scalers, model weights)
# for both transformers and model
# temp restore from two files
self._print_config(trial.config)
dirname = tempfile.mkdtemp(prefix="automl_")
try:
with zipfile.ZipFile(trial.model_path) as zf:
zf.extractall(dirname)
# print("files are extracted into" + dirname)
# print(os.listdir(dirname))
model_path = os.path.join(dirname, "weights_tune.h5")
config_path = os.path.join(dirname, "local_config.json")
local_config = load_config(config_path)
all_config = trial.config.copy()
all_config.update(local_config)
model.restore(model_path, **all_config)
feature_transformers.restore(**all_config)
finally:
shutil.rmtree(dirname)
# model.restore(model_path)
# feature_transformers.restore(config_path, **trial.config)
return TimeSequencePipeline(feature_transformers=feature_transformers,
model=model,
config=all_config)
def _make_pipeline(self, trial, feature_transformers, model, remote_dir):
isinstance(trial, TrialOutput)
config = convert_bayes_configs(trial.config).copy()
self._print_config(config)
if remote_dir is not None:
all_config = restore_hdfs(trial.model_path,
remote_dir,
feature_transformers,
model,
# config)
)
else:
all_config = restore_zip(trial.model_path,
feature_transformers,
model,
# config)
)
return TimeSequencePipeline(name=self.name,
feature_transformers=feature_transformers,
model=model,
config=all_config)
def test_get_default_configs(self):
ppl = TimeSequencePipeline(name='test')
ppl.get_default_configs()
# print(test_df.describe())
tsp = TimeSequencePredictor(
dt_col="datetime",
target_col="value",
extra_features_col=None,
)
pipeline = tsp.fit(train_df,
validation_df=val_df,
metric="mean_squared_error")
print(
"evaluate:",
pipeline.evaluate(test_df, metric=["mean_squared_error", "r_square"]))
pred = pipeline.predict(test_df)
print("predict:", pred.shape)
save_pipeline_file = "../../../saved_pipeline/"
pipeline.save(save_pipeline_file)
new_pipeline = TimeSequencePipeline()
new_pipeline.restore(save_pipeline_file)
print(
"evaluate:",
new_pipeline.evaluate(test_df,
metric=["mean_squared_error", "r_square"]))
new_pred = new_pipeline.predict(test_df)
print("predict:", pred.shape)
np.testing.assert_allclose(pred["value"].values, new_pred["value"].values)