def eval_input_fn(batch_size, cache=False):
feature_types = []
for name in feature_column_names:
# NOTE: vector columns like 23,21,3,2,0,0 should use shape None
if feature_metas[name]["is_sparse"]:
feature_types.append((tf.int64, tf.int32, tf.int64))
else:
feature_types.append(get_dtype(feature_metas[name]["dtype"]))
if is_pai:
pai_table_parts = pai_table.split(".")
formatted_pai_table = "odps://%s/tables/%s" % (pai_table_parts[0],
pai_table_parts[1])
gen = db.pai_maxcompute_db_generator(formatted_pai_table,
feature_column_names, None,
feature_metas)
selected_cols = feature_column_names
else:
gen = db.db_generator(driver, conn, select, feature_column_names,
None, feature_metas)
selected_cols = db.selected_cols(driver, conn, select)
tf_gen = tf_generator(gen, selected_cols, feature_column_names,
feature_metas)
dataset = tf.data.Dataset.from_generator(tf_gen,
(tuple(feature_types), ))
ds_mapper = functools.partial(
parse_sparse_feature_predict,
feature_column_names=feature_column_names,
feature_metas=feature_metas)
dataset = dataset.map(ds_mapper).batch(batch_size)
if cache:
dataset = dataset.cache()
return dataset
def pred(datasource,
select,
feature_metas,
feature_column_names,
label_meta,
result_table,
is_pai=False,
hdfs_namenode_addr="",
hive_location="",
hdfs_user="",
hdfs_pass="",
pai_table="",
model_params=None,
train_params=None,
transform_fn=None,
feature_column_code=""):
if not is_pai:
conn = db.connect_with_data_source(datasource)
else:
conn = None
label_name = label_meta["feature_name"]
dpred = xgb_dataset(
datasource=datasource,
fn='predict.txt',
dataset_sql=select,
feature_specs=feature_metas,
feature_column_names=feature_column_names,
label_spec=None,
is_pai=is_pai,
pai_table=pai_table,
pai_single_file=True,
cache=True,
batch_size=DEFAULT_PREDICT_BATCH_SIZE,
transform_fn=transform_fn,
feature_column_code=feature_column_code,
raw_data_dir="predict.raw.dir") # NOTE: default to use external memory
bst = xgb.Booster({'nthread': 4}) # init model
bst.load_model("my_model") # load data
print("Start predicting XGBoost model...")
if is_pai:
pai_table = "odps://{}/tables/{}".format(*pai_table.split("."))
selected_cols = db.pai_selected_cols(pai_table)
else:
selected_cols = db.selected_cols(conn.driver, conn, select)
feature_file_id = 0
for pred_dmatrix in dpred:
predict_and_store_result(bst, pred_dmatrix, feature_file_id,
model_params, selected_cols, label_name,
is_pai, conn, result_table,
hdfs_namenode_addr, hive_location, hdfs_user,
hdfs_pass)
feature_file_id += 1
print("Done predicting. Predict table : %s" % result_table)
def input_fn(select,
datasource,
feature_column_names,
feature_metas,
label_meta,
is_pai=False,
pai_table="",
num_workers=1,
worker_id=0):
feature_types = []
shapes = []
for name in feature_column_names:
# NOTE: vector columns like 23,21,3,2,0,0 should use shape None
if feature_metas[name]["is_sparse"]:
feature_types.append((tf.int64, tf.int32, tf.int64))
shapes.append((None, None, None))
else:
feature_types.append(get_dtype(feature_metas[name]["dtype"]))
shapes.append(feature_metas[name]["shape"])
if is_pai:
pai_table = "odps://{}/tables/{}".format(*pai_table.split("."))
return pai_dataset(pai_table,
feature_column_names,
label_meta,
feature_metas,
slice_id=worker_id,
slice_count=num_workers)
selected_cols = db.pai_selected_cols(pai_table)
else:
conn = db.connect_with_data_source(datasource)
gen = db.db_generator(conn.driver, conn, select, feature_column_names,
label_meta, feature_metas)
selected_cols = db.selected_cols(conn.driver, conn, select)
gen = tf_generator(gen, selected_cols, feature_column_names, feature_metas)
# Clustering model do not have label
if not label_meta or label_meta["feature_name"] == "":
dataset = tf.data.Dataset.from_generator(gen, (tuple(feature_types), ),
(tuple(shapes), ))
ds_mapper = functools.partial(
parse_sparse_feature_predict,
feature_column_names=feature_column_names,
feature_metas=feature_metas)
else:
dataset = tf.data.Dataset.from_generator(
gen, (tuple(feature_types), eval("tf.%s" % label_meta["dtype"])),
(tuple(shapes), label_meta["shape"]))
ds_mapper = functools.partial(
parse_sparse_feature,
feature_column_names=feature_column_names,
feature_metas=feature_metas)
return dataset.map(ds_mapper)
def xgb_dataset(datasource,
fn,
dataset_sql,
feature_specs,
feature_column_names,
label_spec,
is_pai=False,
pai_table="",
pai_single_file=False,
cache=False,
batch_size=None,
epoch=1,
rank=0,
nworkers=1):
if is_pai:
for dmatrix in pai_dataset(
fn,
feature_specs,
feature_column_names,
label_spec,
"odps://{}/tables/{}".format(*pai_table.split(".")),
pai_single_file,
cache,
rank,
nworkers,
batch_size=batch_size):
yield dmatrix
return
conn = db.connect_with_data_source(datasource)
gen = db.db_generator(conn.driver, conn, dataset_sql, feature_column_names,
label_spec, feature_specs)()
selected_cols = db.selected_cols(conn.driver, conn, dataset_sql)
for i in range(epoch):
step = 0
# the filename per batch is [filename]_[step]
step_file_name = "%s_%d" % (fn, step)
written_rows = dump_dmatrix(step_file_name, gen, feature_column_names,
feature_specs, label_spec, selected_cols)
while written_rows > 0:
yield load_dmatrix('{0}#{0}.cache'.format(step_file_name)
if cache else step_file_name)
os.remove(step_file_name)
step += 1
step_file_name = "%s_%d" % (fn, step)
written_rows = dump_dmatrix(step_file_name, gen,
feature_column_names, feature_specs,
label_spec, selected_cols)
def xgb_shap_dataset(datasource, select, feature_column_names, label_spec,
feature_specs, is_pai, pai_explain_table):
label_column_name = label_spec["feature_name"]
if is_pai:
pai_table_parts = pai_explain_table.split(".")
formatted_pai_table = "odps://%s/tables/%s" % (pai_table_parts[0],
pai_table_parts[1])
stream = db.pai_maxcompute_db_generator(formatted_pai_table,
feature_column_names,
label_column_name,
feature_specs)
selected_cols = feature_column_names[:]
else:
conn = db.connect_with_data_source(datasource)
stream = db.db_generator(conn.driver, conn, select,
feature_column_names, label_spec,
feature_specs)
selected_cols = db.selected_cols(conn.driver, conn, select)
xs = pd.DataFrame(columns=feature_column_names)
i = 0
for row, label in stream():
features = db.read_features_from_row(row, selected_cols,
feature_column_names,
feature_specs)
xs.loc[i] = [item[0] for item in features]
i += 1
# NOTE(typhoonzero): set dtype to the feature's actual type, or the dtype
# may be "object". Use below code to reproduce:
# import pandas as pd
# feature_column_names=["a", "b"]
# xs = pd.DataFrame(columns=feature_column_names)
# for i in range(10):
# xs.loc[i] = [int(j) for j in range(2)]
# print(xs.dtypes)
for fname in feature_column_names:
dtype = feature_specs[fname]["dtype"]
xs[fname] = xs[fname].astype(dtype)
return xs
def xgb_shap_dataset(datasource,
select,
feature_column_names,
label_spec,
feature_specs,
is_pai,
pai_explain_table,
transform_fn=None,
feature_column_code=""):
label_column_name = label_spec["feature_name"]
if is_pai:
pai_table_parts = pai_explain_table.split(".")
formatted_pai_table = "odps://%s/tables/%s" % (pai_table_parts[0],
pai_table_parts[1])
stream = db.pai_maxcompute_db_generator(formatted_pai_table,
feature_column_names,
label_column_name,
feature_specs)
selected_cols = db.pai_selected_cols(formatted_pai_table)
else:
conn = db.connect_with_data_source(datasource)
stream = db.db_generator(conn.driver, conn, select,
feature_column_names, label_spec,
feature_specs)
selected_cols = db.selected_cols(conn.driver, conn, select)
if transform_fn:
column_names = transform_fn.get_column_names()
else:
column_names = feature_column_names
# NOTE(sneaxiy): pandas.DataFrame does not support Tensor whose rank is larger than 2.
# But `INDICATOR` would generate one hot vector for each element, and pandas.DataFrame
# would not accept `INDICATOR` results as its input. In a word, we do not support
# `TO EXPLAIN` when using `INDICATOR`.
xs = pd.DataFrame(columns=column_names)
dtypes = []
i = 0
for row, label in stream():
features = db.read_features_from_row(row, selected_cols,
feature_column_names,
feature_specs)
if transform_fn:
features = transform_fn(features)
# TODO(sneaxiy): support sparse features in `TO EXPLAIN`
features = [item[0] for item in features]
xs.loc[i] = features
if i == 0:
for f in features:
if isinstance(f, np.ndarray):
if f.dtype == np.float32 or f.dtype == np.float64:
dtypes.append('float32')
elif f.dtype == np.int32 or f.dtype == np.int64:
dtypes.append('int64')
else:
raise ValueError('Not supported data type {}'.format(
f.dtype))
elif isinstance(f, (np.float32, np.float64, float)):
dtypes.append('float32')
elif isinstance(f, (np.int32, np.int64, six.integer_types)):
dtypes.append('int64')
else:
raise ValueError('Not supported data type {}'.format(
type(f)))
i += 1
# NOTE(typhoonzero): set dtype to the feature's actual type, or the dtype
# may be "object". Use below code to reproduce:
# import pandas as pd
# feature_column_names=["a", "b"]
# xs = pd.DataFrame(columns=feature_column_names)
# for i in range(10):
# xs.loc[i] = [int(j) for j in range(2)]
# print(xs.dtypes)
for dtype, name in zip(dtypes, column_names):
xs[name] = xs[name].astype(dtype)
return xs
def estimator_predict(estimator, model_params, save, result_table,
feature_column_names, feature_column_names_map,
feature_columns, feature_metas, result_col_name,
datasource, select, hdfs_namenode_addr, hive_location,
hdfs_user, hdfs_pass, is_pai, pai_table):
if not is_pai:
conn = db.connect_with_data_source(datasource)
column_names = feature_column_names[:]
column_names.append(result_col_name)
if is_pai:
driver = "pai_maxcompute"
conn = None
pai_table_parts = pai_table.split(".")
formatted_pai_table = "odps://%s/tables/%s" % (pai_table_parts[0],
pai_table_parts[1])
selected_cols = db.pai_selected_cols(formatted_pai_table)
predict_generator = db.pai_maxcompute_db_generator(
formatted_pai_table, feature_column_names, None, feature_metas)()
else:
driver = conn.driver
# bypass all selected cols to the prediction result table
selected_cols = db.selected_cols(conn.driver, conn, select)
predict_generator = db.db_generator(conn.driver, conn, select,
feature_column_names, None,
feature_metas)()
write_cols, target_col_index = write_cols_from_selected(
result_col_name, selected_cols)
# load from the exported model
with open("exported_path", "r") as fn:
export_path = fn.read()
if tf_is_version2():
imported = tf.saved_model.load(export_path)
else:
imported = tf.saved_model.load_v2(export_path)
def add_to_example(example, x, i):
feature_name = feature_column_names[i]
dtype_str = feature_metas[feature_name]["dtype"]
if feature_metas[feature_name]["delimiter"] != "":
if feature_metas[feature_name]["is_sparse"]:
# NOTE(typhoonzero): sparse feature will get (indices,values,shape) here, use indices only
values = x[0][i][0].flatten()
else:
values = x[0][i].flatten()
if dtype_str == "float32" or dtype_str == "float64":
example.features.feature[feature_name].float_list.value.extend(
list(values))
elif dtype_str == "int32" or dtype_str == "int64":
example.features.feature[feature_name].int64_list.value.extend(
list(values))
else:
if "feature_columns" in feature_columns:
idx = feature_column_names.index(feature_name)
fc = feature_columns["feature_columns"][idx]
else:
# DNNLinearCombinedXXX have dnn_feature_columns and linear_feature_columns param.
idx = -1
try:
idx = feature_column_names_map[
"dnn_feature_columns"].index(feature_name)
fc = feature_columns["dnn_feature_columns"][idx]
except:
try:
idx = feature_column_names_map[
"linear_feature_columns"].index(feature_name)
fc = feature_columns["linear_feature_columns"][idx]
except:
pass
if idx == -1:
raise ValueError(
"can not found feature %s in all feature columns")
if dtype_str == "float32" or dtype_str == "float64":
# need to pass a tuple(float, )
example.features.feature[feature_name].float_list.value.extend(
(float(x[0][i][0]), ))
elif dtype_str == "int32" or dtype_str == "int64":
numeric_type = type(tf.feature_column.numeric_column("tmp"))
if type(fc) == numeric_type:
example.features.feature[
feature_name].float_list.value.extend(
(float(x[0][i][0]), ))
else:
example.features.feature[
feature_name].int64_list.value.extend(
(int(x[0][i][0]), ))
elif dtype_str == "string":
example.features.feature[feature_name].bytes_list.value.extend(
x[0][i])
def predict(x):
example = tf.train.Example()
for i in range(len(feature_column_names)):
add_to_example(example, x, i)
return imported.signatures["predict"](
examples=tf.constant([example.SerializeToString()]))
with db.buffered_db_writer(driver, conn, result_table, write_cols, 100,
hdfs_namenode_addr, hive_location, hdfs_user,
hdfs_pass) as w:
for row, _ in predict_generator:
features = db.read_features_from_row(row, selected_cols,
feature_column_names,
feature_metas)
result = predict((features, ))
if target_col_index != -1:
del row[target_col_index]
if "class_ids" in result:
row.append(str(result["class_ids"].numpy()[0][0]))
else:
# regression predictions
row.append(str(result["predictions"].numpy()[0][0]))
w.write(row)
def xgb_dataset(datasource,
fn,
dataset_sql,
feature_specs,
feature_column_names,
label_spec,
is_pai=False,
pai_table="",
pai_single_file=False,
cache=False,
batch_size=None,
epoch=1,
rank=0,
nworkers=1,
transform_fn=None,
feature_column_code="",
raw_data_dir=None):
if raw_data_dir:
# raw_data_dir is needed when predicting. Because we
# should write the raw data from the source db into
# the dest db, instead of the transformed data after
# `transform_fn(features)` . If raw_data_dir is not
# None, the raw data from the source db would be written
# into another file.
if os.path.exists(raw_data_dir):
shutil.rmtree(raw_data_dir, ignore_errors=True)
os.mkdir(raw_data_dir)
if is_pai:
for dmatrix in pai_dataset(
fn,
feature_specs,
feature_column_names,
label_spec,
"odps://{}/tables/{}".format(*pai_table.split(".")),
pai_single_file,
cache,
rank,
nworkers,
batch_size=batch_size,
feature_column_code=feature_column_code,
raw_data_dir=raw_data_dir):
yield dmatrix
return
conn = db.connect_with_data_source(datasource)
gen = db.db_generator(conn.driver, conn, dataset_sql, feature_column_names,
label_spec, feature_specs)()
selected_cols = db.selected_cols(conn.driver, conn, dataset_sql)
for _ in six.moves.range(epoch):
step = 0
# the filename per batch is [filename]_[step]
step_file_name = "%s_%d" % (fn, step)
written_rows = dump_dmatrix(step_file_name,
gen,
feature_column_names,
feature_specs,
label_spec,
selected_cols,
transform_fn=transform_fn,
raw_data_dir=raw_data_dir)
while written_rows > 0:
yield load_dmatrix('{0}#{0}.cache'.format(step_file_name)
if cache else step_file_name)
os.remove(step_file_name)
step += 1
step_file_name = "%s_%d" % (fn, step)
written_rows = dump_dmatrix(step_file_name,
gen,
feature_column_names,
feature_specs,
label_spec,
selected_cols,
transform_fn=transform_fn,
raw_data_dir=raw_data_dir)