def main(dataset: str, target: str, symbol: str):
ds_service = DatasetService()
ds = ds_service.get_dataset(name=dataset, symbol=symbol)
fs = DatasetService.get_feature_selection(ds=ds, method='importances_shap', target='class')
# hierarchy = load_hierarchy(f"{dataset}_{target}_feature_hierarchy.yml", importances=fs.feature_importances)
# hdf = pd.DataFrame(hierarchy)
# fig = px.treemap(hdf, path=['category', 'subgroup', 'name'], values='importance')
# fig.show()
#
# fig = px.sunburst(hdf, path=['category', 'subgroup', 'name'], values='importance')
# fig.show()
shap_values, shap_expected_values = parse_shap_values(fs.shap_values)
X = ds_service.get_dataset_features(ds=ds, begin=fs.search_interval.begin, end=fs.search_interval.end)
y = ds_service.get_target(name='class', symbol=symbol, begin=fs.search_interval.begin, end=fs.search_interval.end)
fig = plt.figure()
plt.suptitle(f"Shap summary plot for {dataset}.{symbol} -> {target}")
shap.summary_plot(shap_values, X, class_names=["SELL", "HOLD", "BUY"], show=False, max_display=352, use_log_scale=True)
plt.tight_layout()
fig.show()
shap_dfs = []
for cls, arr in enumerate(shap_values):
class_df = pd.DataFrame(arr, columns=X.columns, index=X.index)
class_df.columns = [f"{c}_class{cls}" for c in class_df.columns]
shap_dfs.append(class_df)
shap_df = pd.concat(shap_dfs, axis='columns')
shap_df = shap_df.reindex(sorted(shap_df.columns), axis=1)
print(shap_df.head())
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 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 get_dataset(
symbol: str,
dataset: Optional[str] = None,
target: Optional[str] = None,
begin: Optional[str] = None,
end: Optional[str] = None,
service: DatasetService = Depends(DatasetService),
):
if not dataset and not target:
raise HTTPException(
status_code=400,
detail=
"At least one of 'dataset' or 'target' parameters must be specified!"
)
_name = dataset
if not _name:
_name = 'target'
d = service.get_dataset(name=_name, symbol=symbol)
# If begin/end not specified, use recorded.
# If auto use valid.
if not begin:
begin = d.index_min
elif begin == 'auto':
begin = d.valid_index_min
if not end:
end = d.index_max
elif end == 'auto':
end = d.valid_index_max
# Retrieve dataframes
dfs = []
if dataset:
df = service.get_features(name=dataset,
symbol=symbol,
begin=begin,
end=end)
dfs.append(df)
if target:
dfs.append(
service.get_target(name=target,
symbol=symbol,
begin=begin,
end=end))
# Concatenate dataframes and target
res = pd.concat(dfs, axis='columns') if len(dfs) > 1 else dfs[0]
# Return CSV
return res.to_csv(index_label='time')
class GridSearchService:
def __init__(self):
self.model_repo = ModelRepository()
self.model_service = ModelService()
self.dataset_service = DatasetService()
def create_parameters_search(self, model: Model, split: float,
**kwargs) -> ModelParameters:
ds = self.dataset_service.get_dataset(model.dataset, model.symbol)
splits = DatasetService.get_train_test_split_indices(ds, split)
# Features can either be a list of features to use, or a string
# If it is a string, and it is "latest", pick the latest
features = kwargs.get('features')
# if isinstance(features, str) and features == 'latest':
# if model.features:
# features = model.features[-1].features
# else:
# features = None
if features:
target = kwargs.get('target', 'class')
mf = DatasetService.get_feature_selection(
ds=ds, method=kwargs.get('features'), target=target)
if not mf:
raise MessageException(
f"Feature selection not found for {model.dataset}.{model.symbol} -> {target}!"
)
features = mf.features
# Determine 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
k = 5
# If samples per fold with 5-fold CV are too low, use 3-folds
if ds.count / k < X:
k = 3
# If samples are still too low, raise a value error
if ds.count / k < X and not kwargs.get("permissive"):
raise ValueError("Not enough samples to perform cross validation!")
result = ModelParameters(cv_interval=splits['train'],
cv_splits=k,
task_key=kwargs.get('task_key', str(uuid4())),
features=features or None)
return result
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 grid_search(self, model: Model, mp: ModelParameters,
**kwargs) -> ModelParameters:
pipeline_module, X, y = self._get_dataset_and_pipeline(model, mp)
tag = "{}-{}-{}-{}-{}" \
.format(model.symbol, model.dataset, model.target, model.pipeline, dict_hash(mp.parameters))
# Perform search
if not kwargs.get('halving'):
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=mp.cv_splits),
scoring=get_precision_scorer(),
verbose=kwargs.get("verbose", 0),
n_jobs=kwargs.get("n_jobs", None),
refit=False)
else:
gscv = HalvingGridSearchCV(
estimator=pipeline_module.estimator,
param_grid=kwargs.get('parameter_grid',
pipeline_module.PARAMETER_GRID),
factor=2,
cv=BlockingTimeSeriesSplit(n_splits=mp.cv_splits),
scoring=get_precision_scorer(),
verbose=kwargs.get("verbose", 0),
n_jobs=kwargs.get("n_jobs",
cpu_count() / 2),
refit=False,
random_state=0)
try:
mp.start_at = get_timestamp() # Log starting timestamp
gscv.fit(X, y)
mp.end_at = get_timestamp() # Log ending timestamp
except SplitException as e:
logging.exception(
"Model {} splitting yields single-class folds!\n{}".format(
tag, e.message))
return mp # Fit failed, don't save this.
except ValueError as e:
logging.exception("Model {} raised ValueError!\n{}".format(tag, e))
return mp # Fit failed, don't save this.
# Collect results
results_df = pd.DataFrame(gscv.cv_results_)
# Update search request with results
mp.parameter_search_method = 'halving_grid_search' if kwargs.get(
'halving') else 'gridsearch'
mp.parameters = gscv.best_params_
mp.cv_results = results_df.to_dict()
mp.result_file = 'cv_results-{}.csv'.format(tag)
# 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)
# Update model with the new results
self.model_repo.append_parameters(model.id, mp)
return mp
def random_search(self, model: Model, mp: ModelParameters,
**kwargs) -> ModelParameters:
pipeline_module, X, y = self._get_dataset_and_pipeline(model, mp)
tag = "{}-{}-{}-{}-{}" \
.format(model.symbol, model.dataset, model.target, model.pipeline, dict_hash(mp.parameters))
rscv = RandomizedSearchCV(estimator=pipeline_module.estimator,
param_distributions=kwargs.get(
'param_distributions',
pipeline_module.PARAMETER_DISTRIBUTION),
n_iter=kwargs.get('n_iter', 10),
cv=StratifiedKFold(n_splits=mp.cv_splits),
scoring=get_precision_scorer(),
verbose=kwargs.get("verbose", 0),
n_jobs=kwargs.get("n_jobs", None),
refit=False,
random_state=0)
try:
mp.start_at = get_timestamp() # Log starting timestamp
rscv.fit(X, y)
mp.end_at = get_timestamp() # Log ending timestamp
except SplitException as e:
logging.exception(
"Model {} splitting yields single-class folds!\n{}".format(
tag, e.message))
return mp # Fit failed, don't save this.
except ValueError as e:
logging.exception("Model {} raised ValueError!\n{}".format(tag, e))
return mp # Fit failed, don't save this.
# Collect results
results_df = pd.DataFrame(rscv.cv_results_)
# Update search request with results
mp.parameter_search_method = 'randomsearch'
mp.parameters = rscv.best_params_
mp.result_file = 'cv_results-{}.csv'.format(tag)
# Save grid search results on storage
if kwargs.get('save', True):
storage_service.upload_json_obj(mp.parameters,
'random-search-results',
'parameters-{}.json'.format(tag))
storage_service.save_df(results_df, 'random-search-results',
mp.result_file)
# Update model with the new results
self.model_repo.append_parameters(model.id, mp)
return mp
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
class FeatureSelectionService:
def __init__(self):
self.model_repo = ModelRepository()
self.dataset_service = DatasetService()
def create_features_search(self,
*,
symbol: str,
dataset: str,
target: str,
split: float,
method: str,
task_key: str = None) -> ModelFeatures:
ds = self.dataset_service.get_dataset(dataset, symbol)
splits = DatasetService.get_train_test_split_indices(ds, split)
result = ModelFeatures(dataset=dataset,
target=target,
symbol=symbol,
search_interval=splits['train'],
feature_selection_method=method,
task_key=task_key or str(uuid4()))
return result
def feature_selection(self, mf: ModelFeatures, **kwargs) -> ModelFeatures:
# Load dataset
X = self.dataset_service.get_features(mf.dataset,
mf.symbol,
mf.search_interval.begin,
mf.search_interval.end,
columns=mf.features)
y = self.dataset_service.get_target(mf.target, mf.symbol,
mf.search_interval.begin,
mf.search_interval.end)
unique, counts = np.unique(y, return_counts=True)
if len(unique) < 2:
logging.error(
"[{}-{}-{}]Training data contains less than 2 classes: {}".
format(mf.symbol, mf.dataset, mf.target, unique))
raise MessageException(
"Training data contains less than 2 classes: {}".format(
unique))
# Perform search
mf.start_at = get_timestamp() # Log starting timestamp
if not mf.feature_selection_method or mf.feature_selection_method == 'importances':
selector = select_from_model(X, y)
mf.feature_importances = label_feature_importances(
selector.estimator_, X.columns)
elif mf.feature_selection_method == 'importances_cv':
selector = select_from_model_cv(X, y)
mf.feature_importances = label_feature_importances(
selector.estimator_.best_estimator_, X.columns)
elif mf.feature_selection_method == 'fscore':
selector = select_percentile(X, y, percentile=10)
elif mf.feature_selection_method == 'relieff':
selector = select_relieff(X, y, percentile=10)
elif mf.feature_selection_method == 'multisurf':
selector = select_multisurf(X, y, percentile=10)
else:
raise NotFoundException(
"Cannot find feature selection method by {}".format(
mf.feature_selection_method))
mf.end_at = get_timestamp() # Log ending timestamp
# Update search request with results
mf.features = label_support(selector.get_support(), X.columns)
# Update model with the new results
if kwargs.get('save', True):
self.model_repo.append_features_query(
{
"dataset": mf.dataset,
"symbol": mf.symbol,
"target": mf.target
}, mf)
return mf
def get_available_symbols(self, dataset: str):
return self.dataset_service.get_dataset_symbols(name=dataset)
def feature_selection_new(self, *, symbol: str, dataset: str, target: str,
split: float, method: str,
**kwargs) -> ModelFeatures:
ds = self.dataset_service.get_dataset(dataset, symbol)
fs_exists = DatasetService.has_feature_selection(ds=ds,
method=method,
target=target)
if fs_exists:
if kwargs.get('replace'):
self.dataset_service.remove_feature_selection(ds=ds,
method=method,
target=target)
else:
if kwargs.get('save'):
raise MessageException(
f"Feature selection with method '{method}' alrady performed for '{dataset}.{symbol}' and target '{target}'"
)
splits = DatasetService.get_train_test_split_indices(ds, split)
fs = FeatureSelection(target=target,
method=method,
search_interval=splits['train'],
task_key=kwargs.get('task_key', str(uuid4())))
# Load dataset
X = self.dataset_service.get_dataset_features(
ds=ds, begin=fs.search_interval.begin, end=fs.search_interval.end)
y = self.dataset_service.get_dataset_target(
name=fs.target,
ds=ds,
begin=fs.search_interval.begin,
end=fs.search_interval.end)
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, unique))
raise MessageException(
"Training data contains less than 2 classes: {}".format(
unique))
# Perform search
fs.start_at = get_timestamp() # Log starting timestamp
if not fs.method or 'importances' in fs.method:
if '_cv' in fs.method:
selector = select_from_model_cv(X, y)
else:
selector = select_from_model(X, y)
fs.feature_importances = label_feature_importances(
selector.estimator_, X.columns)
if '_shap' in fs.method:
fs.shap_values = get_shap_values(
model=selector.estimator_.named_steps.c, X=X, X_train=X)
shap_values = parse_shap_values(fs.shap_values)
elif fs.method == 'fscore':
selector = select_percentile(X, y, percentile=10)
elif fs.method == 'relieff':
selector = select_relieff(X, y, percentile=10)
elif fs.method == 'multisurf':
selector = select_multisurf(X, y, percentile=10)
else:
raise NotFoundException(
"Cannot find feature selection method by {}".format(fs.method))
fs.end_at = get_timestamp() # Log ending timestamp
# Update search request with results
fs.features = label_support(selector.get_support(), X.columns)
if not kwargs.get('save'):
return fs
return self.dataset_service.append_feature_selection(ds, fs)