def test_save_load_db(self):
table = "sqlflow_models.test_model"
meta = {"model_params": {"n_classes": 3}}
m = Model(EstimatorType.XGBOOST, meta)
datasource = get_datasource()
# save mode
with temp_file.TemporaryDirectory() as d:
m.save_to_db(datasource, table, d)
# load model
with temp_file.TemporaryDirectory() as d:
m = Model.load_from_db(datasource, table, d)
self.assertEqual(m._meta, meta)
def predict(datasource,
select,
result_table,
label_name,
model,
pai_table="",
oss_model_path=""):
"""PAI XGBoost prediction wrapper
This function do some preparation for the local prediction, say,
download the model from OSS, extract metadata and so on.
Args:
datasource: the datasource from which to get data
select: data selection SQL statement
data_table: tmp table which holds the data from select
result_table: table to save prediction result
label_name: prediction label column
oss_model_path: the model path on OSS
"""
is_pai = True if pai_table != "" else False
if is_pai:
# NOTE(typhoonzero): the xgboost model file "my_model" is hard coded
# in xgboost/train.py
oss.load_file(oss_model_path, "my_model")
(estimator, model_params, train_params, feature_metas,
feature_column_names, train_label_desc,
fc_map_ir) = oss.load_metas(oss_model_path, "xgboost_model_desc")
else:
if isinstance(model, six.string_types):
model = Model.load_from_db(datasource, model)
else:
assert isinstance(
model, Model), "not supported model type %s" % type(model)
model_params = model.get_meta("attributes")
fc_map_ir = model.get_meta("features")
train_label_desc = model.get_meta("label").get_field_desc()[0]
feature_columns = compile_ir_feature_columns(fc_map_ir,
EstimatorType.XGBOOST)
field_descs = get_ordered_field_descs(fc_map_ir)
feature_column_names = [fd.name for fd in field_descs]
feature_metas = dict([(fd.name, fd.to_dict(dtype_to_string=True))
for fd in field_descs])
transform_fn = ComposedColumnTransformer(
feature_column_names, *feature_columns["feature_columns"])
bst = xgb.Booster()
bst.load_model("my_model")
conn = db.connect_with_data_source(datasource)
result_column_names, train_label_idx = create_predict_table(
conn, select, result_table, train_label_desc, label_name)
with temp_file.TemporaryDirectory() as tmp_dir_name:
pred_fn = os.path.join(tmp_dir_name, "predict.txt")
raw_data_dir = os.path.join(tmp_dir_name, "predict_raw_dir")
dpred = xgb_dataset(
datasource=datasource,
fn=pred_fn,
dataset_sql=select,
feature_metas=feature_metas,
feature_column_names=feature_column_names,
label_meta=None,
cache=True,
batch_size=10000,
transform_fn=transform_fn,
raw_data_dir=raw_data_dir) # NOTE: default to use external memory
print("Start predicting XGBoost model...")
for idx, pred_dmatrix in enumerate(dpred):
feature_file_name = os.path.join(
tmp_dir_name, "predict_raw_dir/predict.txt_%d" % idx)
preds = _calc_predict_result(bst, pred_dmatrix, model_params)
_store_predict_result(preds, result_table, result_column_names,
train_label_idx, feature_file_name, conn)
print("Done predicting. Predict table : %s" % result_table)
conn.close()
def local_train(original_sql,
model_image,
estimator_string,
datasource,
select,
validation_select,
model_params,
train_params,
feature_metas,
feature_column_names,
feature_column_map,
label_column,
transform_fn,
save,
load="",
is_pai=False,
oss_model_dir=""):
disk_cache = train_params.pop("disk_cache", False)
batch_size = train_params.pop("batch_size", None)
if batch_size is not None and batch_size < 0:
batch_size = None
epoch = train_params.pop("epoch", 1)
num_workers = train_params.pop("num_workers", 1)
label_meta_dict = label_column.get_field_desc()[0].to_dict(
dtype_to_string=True)
def build_dataset(fn, slct):
return xgb_dataset(datasource,
fn,
slct,
feature_metas,
feature_column_names,
label_meta_dict,
cache=disk_cache,
batch_size=batch_size,
epoch=epoch,
transform_fn=transform_fn)
file_name = "my_model"
if load:
Model.load_from_db(datasource, load)
bst = xgb.Booster()
bst.load_model(file_name)
else:
bst = None
with temp_file.TemporaryDirectory() as tmp_dir_name:
train_fn = os.path.join(tmp_dir_name, 'train.txt')
val_fn = os.path.join(tmp_dir_name, 'val.txt')
train_dataset = build_dataset(train_fn, select)
if validation_select:
val_dataset = build_dataset(val_fn, validation_select)
else:
val_dataset = None
eval_result = dict()
watchlist = [None]
if val_dataset:
# The `xgboost.train` API only accepts the XGBoost DMatrix
# object as the training or validation dataset, so we should
# convert the generator to DMatrix.
if isinstance(val_dataset, types.GeneratorType):
val_dataset = list(val_dataset)[0]
watchlist.append((val_dataset, "validate"))
for per_batch_dmatrix in train_dataset:
watchlist[0] = (per_batch_dmatrix, "train")
bst = xgb.train(model_params,
per_batch_dmatrix,
evals=watchlist,
evals_result=eval_result,
xgb_model=bst,
**train_params)
print("Evaluation result: %s" % eval_result)
meta = collect_metadata(original_sql=original_sql,
select=select,
validation_select=validation_select,
model_repo_image=model_image,
class_name=estimator_string,
attributes=model_params,
features=feature_column_map,
label=label_column,
evaluation=eval_result,
num_workers=num_workers)
save_model_to_local_file(bst, model_params, file_name)
model = Model(EstimatorType.XGBOOST, meta)
model.save_to_db(datasource, save)
if is_pai and len(oss_model_dir) > 0:
# TODO(typhoonzero): remove this since we are saving metas into db now.
save_model(oss_model_dir, "my_model", model_params, train_params,
feature_metas, feature_column_names, label_meta_dict,
feature_column_map)
return eval_result
def train(original_sql,
model_image,
estimator_string,
datasource,
select,
validation_select,
model_params,
train_params,
feature_column_map,
label_column,
save,
load=None):
"""
Train, evaluate and save the XGBoost model locally.
Args:
original_sql (str): the original SQL statement.
model_image (str): the model repo docker image.
estimator (str): the XGBoost booster type like xgboost.gbtree.
datasource (str): the database connection URI.
select (str): the SQL statement for training.
validation_select (str): the SQL statement for evaluation.
model_params (dict): the XGBoost model parameters.
train_params (dict): the training parameters, can have
disk_cache(bool), batch_size(int), epoch(int)
settings in the dict.
feature_column_map (dict): the feature column map to do derivation.
label_column (FeatureColumn): the label column.
save (str): the table name to save the trained model and meta.
load (str): the table name to load the pretrained model.
Returns:
A dict which indicates the evaluation result.
"""
conn = db.connect_with_data_source(datasource)
fc_map_ir, fc_label_ir = infer_feature_columns(conn,
select,
feature_column_map,
label_column,
n=1000)
fc_map = compile_ir_feature_columns(fc_map_ir, EstimatorType.XGBOOST)
feature_column_list = fc_map["feature_columns"]
field_descs = get_ordered_field_descs(fc_map_ir)
feature_column_names = [fd.name for fd in field_descs]
feature_metas = dict([(fd.name, fd.to_dict()) for fd in field_descs])
label_meta = label_column.get_field_desc()[0].to_dict()
# NOTE: in the current implementation, we are generating a transform_fn
# from the COLUMN clause. The transform_fn is executed during the process
# of dumping the original data into DMatrix SVM file.
transform_fn = xgboost_extended.feature_column.ComposedColumnTransformer(
feature_column_names, *feature_column_list)
disk_cache = False
batch_size = None
epoch = 1
if "disk_cache" in train_params:
disk_cache = train_params.pop("disk_cache")
if "batch_size" in train_params:
batch_size = train_params.pop("batch_size")
if "epoch" in train_params:
epoch = train_params.pop("epoch")
def build_dataset(fn, slct):
return xgb_dataset(datasource,
fn,
slct,
feature_metas,
feature_column_names,
label_meta,
cache=disk_cache,
batch_size=batch_size,
epoch=epoch,
transform_fn=transform_fn)
file_name = "my_model"
if load:
Model.load_from_db(datasource, load)
bst = xgb.Booster()
bst.load_model(file_name)
else:
bst = None
with temp_file.TemporaryDirectory() as tmp_dir_name:
train_fn = os.path.join(tmp_dir_name, 'train.txt')
val_fn = os.path.join(tmp_dir_name, 'val.txt')
train_dataset = build_dataset(train_fn, select)
if validation_select:
val_dataset = build_dataset(val_fn, validation_select)
else:
val_dataset = None
eval_result = dict()
watchlist = [None]
if val_dataset:
# The `xgboost.train` API only accepts the XGBoost DMatrix
# object as the training or validation dataset, so we should
# convert the generator to DMatrix.
if isinstance(val_dataset, types.GeneratorType):
val_dataset = list(val_dataset)[0]
watchlist.append((val_dataset, "validate"))
for per_batch_dmatrix in train_dataset:
watchlist[0] = (per_batch_dmatrix, "train")
bst = xgb.train(model_params,
per_batch_dmatrix,
evals=watchlist,
evals_result=eval_result,
xgb_model=bst,
**train_params)
print("Evaluation result: %s" % eval_result)
meta = collect_metadata(original_sql=original_sql,
select=select,
validation_select=validation_select,
model_repo_image=model_image,
class_name=estimator_string,
attributes=model_params,
features=fc_map_ir,
label=fc_label_ir,
evaluation=eval_result,
num_workers=1)
save_model_to_local_file(bst, model_params, file_name)
model = Model(EstimatorType.XGBOOST, meta)
model.save_to_db(datasource, save)
return eval_result
def predict(datasource,
select,
result_table,
result_column_names,
train_label_idx,
model,
extra_result_cols=[],
pai_table=None):
"""TBD
"""
bst = xgb.Booster()
if isinstance(model, six.string_types):
# NOTE(typhoonzero): must run Model.load_from_db in a temp
# directory, calling pyodps in current directory on PAI
# workers will cause paiio fails.
with temp_file.TemporaryDirectory(as_cwd=True):
model = Model.load_from_db(datasource, model)
bst.load_model("my_model")
else:
assert isinstance(model,
Model), "not supported model type %s" % type(model)
bst.load_model("my_model")
model_params = model.get_meta("attributes")
fc_map_ir = model.get_meta("features")
feature_columns = compile_ir_feature_columns(fc_map_ir,
EstimatorType.XGBOOST)
field_descs = get_ordered_field_descs(fc_map_ir)
feature_column_names = [fd.name for fd in field_descs]
feature_metas = dict([(fd.name, fd.to_dict(dtype_to_string=True))
for fd in field_descs])
transform_fn = ComposedColumnTransformer(
feature_column_names, *feature_columns["feature_columns"])
is_pai = True if pai_table else False
if is_pai:
conn = PaiIOConnection.from_table(pai_table)
else:
conn = db.connect_with_data_source(datasource)
with temp_file.TemporaryDirectory() as tmp_dir_name:
pred_fn = os.path.join(tmp_dir_name, "predict.txt")
raw_data_dir = os.path.join(tmp_dir_name, "predict_raw_dir")
dpred = xgb_dataset(datasource=datasource,
fn=pred_fn,
dataset_sql=select,
feature_metas=feature_metas,
feature_column_names=feature_column_names,
label_meta=None,
cache=True,
batch_size=10000,
transform_fn=transform_fn,
raw_data_dir=raw_data_dir,
is_pai=is_pai,
pai_table=pai_table,
pai_single_file=True,
feature_column_code=fc_map_ir)
print("Start predicting XGBoost model...")
for idx, pred_dmatrix in enumerate(dpred):
if is_pai:
feature_file_name = os.path.join(tmp_dir_name,
"predict.txt.raw")
else:
feature_file_name = os.path.join(
tmp_dir_name, "predict_raw_dir/predict.txt_%d" % idx)
preds = _calc_predict_result(bst, pred_dmatrix, model_params)
_store_predict_result(preds, result_table, result_column_names,
train_label_idx, feature_file_name, conn)
print("Done predicting. Predict table : %s" % result_table)
conn.close()
def train_step(original_sql,
model_image,
estimator_string,
datasource,
select,
validation_select,
model_params,
train_params,
validation_params,
feature_column_map,
label_column,
save,
load=None,
pai_table=None,
pai_val_table=None):
if model_params is None:
model_params = {}
if train_params is None:
train_params = {}
if validation_params is None:
validation_params = {}
if load:
Model.load_from_db(datasource, load)
load = "model_save"
else:
load = None
is_pai = True if pai_table else False
fc_map = compile_ir_feature_columns(feature_column_map,
EstimatorType.TENSORFLOW)
field_descs = get_ordered_field_descs(feature_column_map)
feature_column_names = [fd.name for fd in field_descs]
feature_metas = dict([(fd.name, fd.to_dict(dtype_to_string=True))
for fd in field_descs])
# no label for clustering model
label_meta = None
if label_column:
label_meta = label_column.get_field_desc()[0].to_dict(
dtype_to_string=True)
feature_column_names_map = dict()
for target in feature_column_map:
fclist = feature_column_map[target]
feature_column_names_map[target] = [
fc.get_field_desc()[0].name for fc in fclist
]
# Construct optimizer objects to pass to model initializer.
# The original model_params is serializable (do not have tf.xxx objects).
model_params_constructed = copy.deepcopy(model_params)
for optimizer_arg in ["optimizer", "dnn_optimizer", "linear_optimizer"]:
if optimizer_arg in model_params_constructed:
model_params_constructed[optimizer_arg] = get_tf_optimizer(
model_params_constructed[optimizer_arg])
if "loss" in model_params_constructed:
model_params_constructed["loss"] = get_tf_loss(
model_params_constructed["loss"])
# extract params for training.
verbose = train_params.get("verbose", 1)
batch_size = train_params.get("batch_size", 1)
epoch = train_params.get("epoch", 1)
save_checkpoints_steps = train_params.get("save_checkpoints_steps", 100)
max_steps = train_params.get("max_steps", None)
if max_steps is not None and max_steps <= 0:
max_steps = None
validation_metrics = validation_params.get("metrics", "Accuracy")
validation_metrics = [v.strip() for v in validation_metrics.split(",")]
validation_steps = validation_params.get("steps", 1)
validation_start_delay_secs = validation_params.get("start_delay_secs", 0)
validation_throttle_secs = validation_params.get("throttle_secs", 0)
estimator = import_model(estimator_string)
is_estimator = is_tf_estimator(estimator)
# always use verbose == 1 when using PAI to get more logs
if verbose < 1:
verbose = 1
set_log_level(verbose, is_estimator)
model_params_constructed.update(fc_map)
FLAGS = define_tf_flags()
set_oss_environs(FLAGS)
num_workers = len(FLAGS.worker_hosts.split(","))
worker_id = FLAGS.task_index
train_dataset_fn = get_dataset_fn(select,
datasource,
feature_column_names,
feature_metas,
label_meta,
is_pai,
pai_table,
batch_size,
epochs=epoch,
shuffle_size=1000,
num_workers=num_workers,
worker_id=worker_id)
val_dataset_fn = None
if validation_select or pai_val_table:
val_dataset_fn = get_dataset_fn(validation_select, datasource,
feature_column_names, feature_metas,
label_meta, is_pai, pai_val_table,
batch_size)
model_meta = collect_metadata(original_sql=original_sql,
select=select,
validation_select=validation_select,
model_repo_image=model_image,
class_name=estimator_string,
attributes=model_params,
features=feature_column_map,
label=label_column)
# FIXME(typhoonzero): avoid save model_meta twice, keras_train_and_save,
# estimator_train_and_save also dumps model_meta to a file under cwd.
# should only keep the model.save_to_db part.
save_dir = "model_save"
if not is_estimator:
if isinstance(estimator, types.FunctionType):
# functional model need field_metas parameter
model_params_constructed["field_metas"] = feature_metas
keras_train_and_save(estimator, model_params_constructed, save_dir,
FLAGS, train_dataset_fn, val_dataset_fn,
label_meta, epoch, verbose, validation_metrics,
validation_steps, load, model_meta, is_pai)
else:
estimator_train_and_save(estimator, model_params_constructed, save_dir,
FLAGS, train_dataset_fn, val_dataset_fn,
max_steps, validation_start_delay_secs,
validation_throttle_secs,
save_checkpoints_steps, validation_metrics,
load, model_meta)
# save model to DB/OSS
model = Model(EstimatorType.TENSORFLOW, model_meta)
if num_workers == 1 or worker_id == 0:
saved = model.save_to_db(datasource,
save,
oss_model_dir=FLAGS.sqlflow_oss_modeldir)
print("Model saved to DB: %s" % saved)
print("Done training")