def predict_day(self, pipeline: str, dataset: str, target: str,
symbol: str, day: str, window: dict):
model = self.get_model(pipeline, dataset, target, symbol)
# Load dataset
ds = DatasetService()
d = ds.get_dataset(model.dataset, model.symbol)
# Get training data including the first training window
begin = sub_interval(timestamp=day, interval=window)
if from_timestamp(d.valid_index_min).timestamp() > from_timestamp(
begin).timestamp():
raise MessageException("Not enough data for training! [Pipeline: {} Dataset: {} Symbol: {} Window: {}]" \
.format(model.pipeline, model.dataset, model.symbol, window))
X = ds.get_features(model.dataset, model.symbol, begin=begin, end=day)
y = ds.get_target(model.target, model.symbol, begin=begin, end=day)
unique, counts = np.unique(y, return_counts=True)
if len(unique) < 2:
logging.error(
"[{}-{}-{}-{}]Training data contains less than 2 classes: {}".
format(model.symbol, model.dataset, model.target,
model.pipeline, unique))
raise MessageException(
"Training data contains less than 2 classes: {}".format(
unique))
# Load pipeline
pipeline_module = get_pipeline(model.pipeline)
# Slice testing interval in windows
df = predict_day(pipeline_module.estimator, model.parameters[-1], X, y,
day)
return df
def _get_dataset_and_pipeline(self, model: Model, mp: ModelParameters,
**kwargs):
if not model.id: # Make sure the task exists
model = self.model_repo.create(model)
if self.model_repo.exist_parameters(model.id, mp.task_key):
logging.info("Model {} Grid search {} already executed!".format(
model.id, mp.task_key))
return mp
# Load dataset
X = self.dataset_service.get_features(model.dataset,
model.symbol,
mp.cv_interval.begin,
mp.cv_interval.end,
columns=mp.features)
y = self.dataset_service.get_target(model.target, model.symbol,
mp.cv_interval.begin,
mp.cv_interval.end)
unique, counts = np.unique(y, return_counts=True)
if len(unique) < 2:
logging.error(
"[{}-{}-{}-{}]Training data contains less than 2 classes: {}".
format(model.symbol, model.dataset, model.target,
model.pipeline, unique))
raise MessageException(
"Training data contains less than 2 classes: {}".format(
unique))
logging.info("Dataset loaded: X {} y {} (unique: {})".format(
X.shape, y.shape, unique))
# Load pipeline
pipeline_module = get_pipeline(model.pipeline)
return pipeline_module, X, y
def test_model(self, model: Model, mt: ModelTest, **kwargs):
if not model.id:
model = self.model_repo.create(model)
if self.model_repo.exist_test(model.id, mt.task_key):
logging.info("Model {} test {} already executed!".format(
model.id, mt.task_key))
return mt
# Load dataset
ds = DatasetService()
d = ds.get_dataset(model.dataset, model.symbol)
# Get training data including the first training window
begin = sub_interval(timestamp=mt.test_interval.begin,
interval=mt.window)
end = add_interval(timestamp=mt.test_interval.end, interval=mt.step)
if from_timestamp(d.valid_index_min).timestamp() > from_timestamp(
begin).timestamp():
raise MessageException("Not enough data for training! [Pipeline: {} Dataset: {} Symbol: {} Window: {}]" \
.format(model.pipeline, model.dataset, model.symbol, mt.window))
X = ds.get_features(model.dataset, model.symbol, begin=begin, end=end)
y = ds.get_target(model.target, model.symbol, begin=begin, end=end)
unique, counts = np.unique(y, return_counts=True)
if len(unique) < 2:
logging.error(
"[{}-{}-{}-{}]Training data contains less than 2 classes: {}".
format(model.symbol, model.dataset, model.target,
model.pipeline, unique))
raise MessageException(
"Training data contains less than 2 classes: {}".format(
unique))
# Load pipeline
pipeline_module = get_pipeline(model.pipeline)
# Slice testing interval in windows
ranges = timestamp_windows(begin, end, mt.window, mt.step)
mt.start_at = get_timestamp()
df = test_windows(pipeline_module.estimator, mt.parameters, X, y,
ranges)
mt.end_at = get_timestamp()
mt.classification_results = df.to_dict()
clf_report = flattened_classification_report_imbalanced(
df.label, df.predicted)
roc_report = roc_auc_report(
df.label, df.predicted,
df[[c for c in df.columns if '_proba_' in c]])
clf_report.update(roc_report)
mt.classification_report = clf_report
self.model_repo.append_test(model.id, mt)
return mt
def select_from_model(X, y):
# Load pipeline
pipeline_module = get_pipeline("selection_xgboost", unlisted=True)
classes = np.unique(y)
print(f"unique y: {classes} count: {classes.size}")
pipeline_module.PARAMETERS.update({'num_class': classes.size})
pipeline = pipeline_module.estimator
pipeline.named_steps.c.set_params(**pipeline_module.PARAMETERS)
# Perform search
sfm = SelectFromModel(
pipeline,
threshold='mean',
importance_getter='named_steps.c.feature_importances_')
sfm.fit(X, y)
return sfm
def select_from_model_cv(X, y, sync=False):
# Load pipeline
pipeline_module = get_pipeline("selection_xgboost", unlisted=True)
gscv = GridSearchCV(estimator=pipeline_module.estimator,
param_grid=pipeline_module.PARAMETER_GRID,
scoring='precision',
cv=3,
n_jobs=8)
# Fit grid search
gscv.fit(X, y)
# Perform search
sfm = SelectFromModel(
gscv,
# prefit=True,
threshold='mean',
importance_getter='best_estimator_.named_steps.c.feature_importances_')
sfm.fit(X, y)
return sfm
def grid_search_new(self, symbol: str, dataset: str, target: str,
pipeline: str, split: float,
feature_selection_method: str, **kwargs):
# Check if a model exists and has same search method
existing_model = self.model_service.get_model(pipeline=pipeline,
dataset=dataset,
target=target,
symbol=symbol)
if existing_model:
mp_exists = ModelService.get_model_parameters(existing_model,
method='gridsearch')
if mp_exists:
if kwargs.get('replace'):
self.model_service.remove_parameters(model=existing_model,
method='gridsearch')
else:
if kwargs.get('save'):
raise MessageException(
f"Grid search already performed for {pipeline}({dataset}.{symbol}) -> {target}"
)
# Retrieve dataset to use
ds = self.dataset_service.get_dataset(dataset, symbol)
# Determine cv_splits=K for K-fold cross validation based on dataset's sample count
# Train-test split for each fold is 80% train, the lowest training window for accurate results is 30 samples
# so we need X samples where X is given by the proportion:
# 30/0.8 = X/1; X= 30/0.8 = 37.5 ~ 40 samples per fold
X = 40
cv_splits = 5
# If samples per fold with 5-fold CV are too low, use 3-folds
if ds.count / cv_splits < X:
cv_splits = 3
# If samples are still too low, raise a value error
if ds.count / cv_splits < X and not kwargs.get("permissive"):
raise ValueError("Not enough samples to perform cross validation!")
# Determine split indices based on dataset
splits = DatasetService.get_train_test_split_indices(ds, split)
cv_interval = splits['train']
# Load dataset features by applying a specified feature selection method
X = self.dataset_service.get_dataset_features(
ds=ds,
begin=cv_interval['begin'],
end=cv_interval['end'],
method=feature_selection_method,
target=target)
y = self.dataset_service.get_target(
name=target,
symbol=symbol,
begin=cv_interval['begin'],
end=cv_interval['end'],
)
# Check number of samples for each class in training data, if less than 3 instances are present for
# each class, we're going to get a very unstable model (or no model at all for k-NN based algos)
unique, counts = np.unique(y, return_counts=True)
if len(unique) < 2:
logging.error(
"[{}-{}-{}-{}]Training data contains less than 2 classes: {}".
format(symbol, dataset, target, pipeline, unique))
raise MessageException(
"Training data contains less than 2 classes: {}".format(
unique))
logging.info("Dataset loaded: X {} y {} (unique: {})".format(
X.shape, y.shape, unique))
# Load pipeline algorithm and parameter grid
pipeline_module = get_pipeline(pipeline)
# Perform search
gscv = GridSearchCV(
estimator=pipeline_module.estimator,
param_grid=kwargs.get('parameter_grid',
pipeline_module.PARAMETER_GRID),
# cv=BlockingTimeSeriesSplit(n_splits=mp.cv_splits),
cv=StratifiedKFold(n_splits=cv_splits),
scoring=get_precision_scorer(),
verbose=kwargs.get("verbose", 0),
n_jobs=kwargs.get("n_jobs", None),
refit=False)
mp = ModelParameters(cv_interval=splits['train'],
cv_splits=cv_splits,
task_key=kwargs.get('task_key', str(uuid4())),
features=[c for c in X.columns],
parameter_search_method='gridsearch')
mp.start_at = get_timestamp()
gscv.fit(X, y)
mp.end_at = get_timestamp()
# Collect results
results_df = pd.DataFrame(gscv.cv_results_)
mp.parameters = gscv.best_params_
mp.cv_results = results_df.loc[:,
results_df.columns != 'params'].to_dict(
'records')
tag = "{}-{}-{}-{}-{}".format(symbol, dataset, target, pipeline,
dict_hash(mp.parameters))
mp.result_file = 'cv_results-{}.csv'.format(tag)
# Is there an existing model for this search?
model = Model(pipeline=pipeline,
dataset=dataset,
target=target,
symbol=symbol,
features=feature_selection_method)
model.parameters.append(mp)
self.model_repo.create(model)
# Save grid search results on storage
if kwargs.get('save', True):
storage_service.upload_json_obj(mp.parameters,
'grid-search-results',
'parameters-{}.json'.format(tag))
storage_service.save_df(results_df, 'grid-search-results',
mp.result_file)
return mp
def fit_estimator_new(model: Model, mp: ModelParameters, features: str,
day: str, window: dict, X, y, b, e, **kwargs):
# Check if estimator exists
if exist_estimator(model=model,
parameters=mp.parameter_search_method,
features=features,
day=day,
window=window):
logging.info(
f"Estimator exists {model.pipeline}({model.dataset}.{model.symbol}) -> {model.target}"
f" Day: {day} Window: {window}")
existing = load_estimator(model=model,
parameters=mp.parameter_search_method,
features=features,
day=day,
window=window)
if existing and existing.is_fit:
return existing
X = X[b:e]
y = y[b:e]
X_train = X[:-1]
y_train = y[:-1]
pipeline_module = get_pipeline(model.pipeline)
y_unique, _, y_counts = np.unique(y_train,
return_index=True,
return_counts=True)
if (y_counts < 3).any():
logging.warning(
f"fit_estimator: y_train contains less than 3 samples for some class! \nUnique: {y_unique}\nCounts: {y_counts}"
)
est = pipeline_module.estimator
est.set_params(**mp.parameters)
try:
start_at = datetime.utcnow().timestamp()
est = est.fit(X_train, y_train)
dur = datetime.utcnow().timestamp() - start_at
except Exception as e:
logging.exception(f"Exception in estimator fit for day: {day}: {e}")
return None
# Save data as attributes of the fit estimator as well
est.fit_time = dur
est.fit_timestamp = get_timestamp()
est.is_fit = True
est.train_x = X_train
est.train_y = y_train
est.begin = b
est.end = e
est.skip_save = False
# Training parameters and Model tuple
est.day = day
est.pipeline = model.pipeline
est.dataset = model.dataset
est.target = model.target
est.symbol = model.symbol
est.train_begin = to_timestamp(X_train.first_valid_index().to_pydatetime())
est.train_end = to_timestamp(X_train.last_valid_index().to_pydatetime())
est.window = window
est.fit_timestamp = get_timestamp()
est.parameters = mp.parameter_search_method
est.features = features
return est