def evaluate(self,
data_creator,
verbose=1,
sample_weight=None,
steps=None,
callbacks=None,
data_config=None,
feature_cols=None,
label_cols=None):
"""Evaluates the model on the validation data set."""
logger.info("Starting validation step.")
params = dict(
verbose=verbose,
sample_weight=sample_weight,
steps=steps,
callbacks=callbacks,
data_config=data_config,
)
from zoo.orca.data import SparkXShards
from pyspark.sql import DataFrame
if isinstance(data_creator, DataFrame):
assert feature_cols is not None,\
"feature_col must be provided if data_creator is a spark dataframe"
assert label_cols is not None,\
"label_cols must be provided if data_creator is a spark dataframe"
schema = data_creator.schema
numpy_rdd = data_creator.rdd.map(lambda row: convert_row_to_numpy(
row, schema, feature_cols, label_cols))
shard_rdd = numpy_rdd.mapPartitions(
lambda x: arrays2dict(x, feature_cols, label_cols))
data_creator = SparkXShards(shard_rdd)
if isinstance(data_creator, SparkXShards):
data = data_creator
if data.num_partitions() != self.num_workers:
data = data.repartition(self.num_workers)
ray_xshards = RayXShards.from_spark_xshards(data)
def transform_func(worker, shards_ref):
params["data_creator"] = shards_ref_to_creator(shards_ref)
return worker.validate.remote(**params)
stats_shards = ray_xshards.transform_shards_with_actors(
self.remote_workers, transform_func, gang_scheduling=True)
worker_stats = stats_shards.collect()
else: # data_creator functions; should return Iter or DataLoader
params["data_creator"] = data_creator
params_list = [params] * self.num_workers
worker_stats = ray.get([
w.validate.remote(**params_list[i])
for i, w in enumerate(self.remote_workers)
])
worker_stats = list(itertools.chain.from_iterable(worker_stats))
stats = worker_stats[0].copy()
return stats
def update_predict_xshards(xshard, pred_xshards):
def updates(d1_d2):
d1, d2 = d1_d2
d1.update(d2)
return d1
result = SparkXShards(xshard.rdd.zip(pred_xshards.rdd).map(updates))
return result
def load(self, model_path, minPartitions=None):
"""
restore model from model file and config.
:param model_path: the model file
:return: the restored model
"""
self.internal = SparkXShards.load_pickle(model_path,
minPartitions=minPartitions)
def get_pred_xshards(key):
rdd = self.sc.range(0, 110).map(lambda x: np.array([x] * 50))
shards = rdd.mapPartitions(lambda iter: chunks(iter, 5)).map(
lambda x: {
key: np.stack(x)
}).map(lambda x: {key: [x[key][:, :24], x[key][:, 24:]]})
shards = SparkXShards(shards)
return shards
def test_openvino_predict_xshards(self):
input_data_list = [np.array([self.input] * 4), np.array([self.input] * 2)]
sc = init_nncontext()
rdd = sc.parallelize(input_data_list, numSlices=2)
shards = SparkXShards(rdd)
def pre_processing(images):
return {"x": images}
shards = shards.transform_shard(pre_processing)
result = self.est.predict(shards)
result_c = result.collect()
assert isinstance(result, SparkXShards)
assert result_c[0]["prediction"].shape == (4, 1000)
assert result_c[1]["prediction"].shape == (2, 1000)
assert self.check_result(result_c[0]["prediction"], 4)
assert self.check_result(result_c[1]["prediction"], 2)
def test_transform_with_repartition(self):
# shards of pandas dataframe
file_path = os.path.join(self.resource_path, "orca/data/csv")
data_shard = zoo.orca.data.pandas.read_csv(file_path)
partitions = data_shard.rdd.glom().collect()
for par in partitions:
assert len(par) <= 1
def negative(df, column_name):
df[column_name] = df[column_name] * (-1)
return df
shard2 = data_shard.transform_shard(negative, "sale_price")
shard3 = shard2.repartition(4)
partitions3 = shard3.rdd.glom().collect()
for par in partitions3:
assert len(par) <= 1
shard4 = shard2.repartition(1)
partitions4 = shard4.rdd.glom().collect()
for par in partitions4:
assert len(par) <= 1
shard5 = shard4.transform_shard(negative, "sale_price")
partitions5 = shard5.rdd.glom().collect()
for par in partitions5:
assert len(par) <= 1
# shards of list
data = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]
sc = init_nncontext()
rdd = sc.parallelize(data)
data_shard = SparkXShards(rdd)
shard2 = data_shard.repartition(6)
partitions2 = shard2.rdd.glom().collect()
for par in partitions2:
assert len(par) <= 1
shard3 = data_shard.repartition(1)
partitions2 = shard3.rdd.glom().collect()
for par in partitions2:
assert len(par) <= 1
# shards of numpy array
data = [np.array([1, 2, 3, 4]), np.array([5, 6, 7, 8]),
np.array([9, 10, 11, 12]), np.array([13, 14, 15, 16])]
sc = init_nncontext()
rdd = sc.parallelize(data)
data_shard = SparkXShards(rdd)
shard2 = data_shard.repartition(6)
partitions2 = shard2.rdd.glom().collect()
for par in partitions2:
assert len(par) <= 1
shard3 = data_shard.repartition(1)
partitions2 = shard3.rdd.glom().collect()
for par in partitions2:
assert len(par) <= 1
def test_spark_xshards(self):
from zoo import init_nncontext
from zoo.orca.data import SparkXShards
estimator = get_estimator(workers_per_node=1)
sc = init_nncontext()
x_rdd = sc.parallelize(np.random.rand(4000, 1, 50).astype(np.float32))
# torch 1.7.1+ requires target size same as output size, which is (batch, 1)
y_rdd = sc.parallelize(
np.random.randint(0, 2, size=(4000, 1, 1)).astype(np.float32))
rdd = x_rdd.zip(y_rdd).map(lambda x_y: {'x': x_y[0], 'y': x_y[1]})
train_rdd, val_rdd = rdd.randomSplit([0.9, 0.1])
train_xshards = SparkXShards(train_rdd)
val_xshards = SparkXShards(val_rdd)
train_stats = estimator.fit(train_xshards, batch_size=256, epochs=2)
print(train_stats)
val_stats = estimator.evaluate(val_xshards, batch_size=128)
print(val_stats)
estimator.shutdown()
def _dataframe_to_xshards(data, feature_cols, label_cols=None):
from zoo.orca import OrcaContext
schema = data.schema
shard_size = OrcaContext._shard_size
numpy_rdd = data.rdd.map(lambda row: convert_row_to_numpy(
row, schema, feature_cols, label_cols))
shard_rdd = numpy_rdd.mapPartitions(
lambda x: arrays2dict(x, feature_cols, label_cols, shard_size))
return SparkXShards(shard_rdd)
def test_openvino_predict_xshards(self):
input_data_list = [
np.random.random([1, 4, 3, 224, 224]),
np.random.random([2, 4, 3, 224, 224])
]
sc = init_nncontext()
rdd = sc.parallelize(input_data_list, numSlices=2)
shards = SparkXShards(rdd)
def pre_processing(images):
return {"x": images}
shards = shards.transform_shard(pre_processing)
result = self.est.predict(shards)
result_c = result.collect()
assert isinstance(result, SparkXShards)
assert result_c[0]["prediction"].shape == (1, 4, 1000)
assert result_c[1]["prediction"].shape == (2, 4, 1000)
def test_zip(self):
def negative(df, column_name, minus_val):
df[column_name] = df[column_name] * (-1)
df[column_name] = df[column_name] - minus_val
return df
file_path = os.path.join(self.resource_path, "orca/data/json")
data_shard = zoo.orca.data.pandas.read_json(file_path,
orient='columns',
lines=True)
data_shard = data_shard.repartition(2)
data_shard.cache()
transformed_shard = data_shard.transform_shard(negative, "value", 2)
zipped_shard = data_shard.zip(transformed_shard)
assert not transformed_shard.is_cached(
), "transformed_shard should be uncached."
data = zipped_shard.collect()
assert data[0][0]["value"].values[0] + data[0][1]["value"].values[0] == -2, \
"value should be -2"
list1 = list([1, 2, 3])
with self.assertRaises(Exception) as context:
data_shard.zip(list1)
self.assertTrue(
'other should be a SparkXShards' in str(context.exception))
transformed_shard = transformed_shard.repartition(
data_shard.num_partitions() - 1)
with self.assertRaises(Exception) as context:
data_shard.zip(transformed_shard)
self.assertTrue(
'The two SparkXShards should have the same number of partitions' in
str(context.exception))
dict_data = [{"x": 1, "y": 2}, {"x": 2, "y": 3}]
sc = init_nncontext()
rdd = sc.parallelize(dict_data)
dict_shard = SparkXShards(rdd)
dict_shard = dict_shard.repartition(1)
with self.assertRaises(Exception) as context:
transformed_shard.zip(dict_shard)
self.assertTrue(
'The two SparkXShards should have the same number of elements in '
'each partition' in str(context.exception))
def test_spark_xshards(self):
from zoo import init_nncontext
from zoo.orca.data import SparkXShards
estimator = Estimator.from_torch(model=get_model,
optimizer=get_optimizer,
loss=nn.BCELoss(),
config={"lr": 1e-1},
backend="torch_distributed")
sc = init_nncontext()
x_rdd = sc.parallelize(np.random.rand(4000, 1, 50).astype(np.float32))
y_rdd = sc.parallelize(
np.random.randint(0, 2, size=(4000, 1)).astype(np.float32))
rdd = x_rdd.zip(y_rdd).map(lambda x_y: {'x': x_y[0], 'y': x_y[1]})
train_rdd, val_rdd = rdd.randomSplit([0.9, 0.1])
train_xshards = SparkXShards(train_rdd)
val_xshards = SparkXShards(val_rdd)
train_stats = estimator.fit(train_xshards, batch_size=256, epochs=2)
print(train_stats)
val_stats = estimator.evaluate(val_xshards, batch_size=128)
print(val_stats)
estimator.shutdown()
def to_spark_xshards(self):
from zoo.orca.data import SparkXShards
ray_ctx = RayContext.get()
sc = ray_ctx.sc
address = ray_ctx.redis_address
password = ray_ctx.redis_password
num_parts = self.num_partitions()
partition2store = self.partition2store_name
rdd = sc.parallelize([0] * num_parts * 10, num_parts)\
.mapPartitionsWithIndex(
lambda idx, _: get_from_ray(idx, address, password, partition2store))
spark_xshards = SparkXShards(rdd)
return spark_xshards
def convert_predict_to_xshard(prediction_rdd):
def transform_predict(iter):
predictions = list(iter)
# list of np array
if isinstance(predictions[0], list):
predictions = np.array(predictions).T.tolist()
result = [np.array(predict) for predict in predictions]
return [{'prediction': result}]
# np array
else:
return [{'prediction': np.array(predictions)}]
return SparkXShards(prediction_rdd.mapPartitions(transform_predict))
def test_convert_predict_rdd_to_xshard(self):
rdd = self.sc.range(0, 110).map(lambda x: np.array([x] * 50))
shards = rdd.mapPartitions(lambda iter: chunks(iter, 5)).map(
lambda x: {"x": np.stack(x)})
shards = SparkXShards(shards)
pred_rdd = self.sc.range(0, 110).map(lambda x: np.array([x] * 50))
result_shards = convert_predict_rdd_to_xshard(shards, pred_rdd)
result = np.concatenate(
[shard["prediction"] for shard in result_shards.collect()])
expected_result = np.concatenate(
[shard["x"] for shard in result_shards.collect()])
assert np.array_equal(result, expected_result)
def test_xshards_predict(self):
sc = init_nncontext()
rdd = sc.range(0, 110).map(lambda x: np.array([x] * 50))
shards = rdd.mapPartitions(lambda iter: chunks(iter, 5)).map(
lambda x: {"x": np.stack(x)})
shards = SparkXShards(shards)
estimator = get_estimator(workers_per_node=2,
model_fn=lambda config: IdentityNet())
result_shards = estimator.predict(shards, batch_size=4)
result = np.concatenate(
[shard["prediction"] for shard in result_shards.collect()])
expected_result = np.concatenate(
[shard["x"] for shard in result_shards.collect()])
assert np.array_equal(result, expected_result)
def convert_predict_rdd_to_xshard(data, prediction_rdd):
import numpy as np
from zoo.orca.data import SparkXShards
def group_index(iter):
for data in iter:
size = get_size(data["x"])
for i in range(size):
yield size
def transform_predict(predictions):
# list of np array
if isinstance(predictions[0], list):
predictions = np.array(predictions).T.tolist()
result = [np.array(predict) for predict in predictions]
return result
# np array
else:
return np.array(predictions)
def group(iter):
this_index = 0
buffer = []
this_count = None
for (count, pred) in iter:
if this_index == 0:
this_count = count
if this_index < this_count:
buffer.append(pred)
this_index += 1
if this_index == this_count:
yield transform_predict(buffer)
buffer.clear()
this_index = 0
def add_pred(shard_pred):
shard, pred = shard_pred
shard["prediction"] = pred
return shard
indexed_rdd = data.rdd.mapPartitions(group_index)
grouped_pred = indexed_rdd.zip(prediction_rdd).mapPartitions(group)
result_rdd = data.rdd.zip(grouped_pred).map(add_pred)
return SparkXShards(result_rdd)
def to_spark_xshards(self):
from zoo.orca.data import SparkXShards
ray_ctx = RayContext.get()
sc = ray_ctx.sc
address = ray_ctx.redis_address
password = ray_ctx.redis_password
num_parts = self.num_partitions()
partition2store = self.partition2store_name
rdd = sc.parallelize([0] * num_parts * 10, num_parts)\
.mapPartitionsWithIndex(
lambda idx, _: get_from_ray(idx, address, password, partition2store))
# the reason why we trigger computation here is to ensure we get the data
# from ray before the RayXShards goes out of scope and the data get garbage collected
from pyspark.storagelevel import StorageLevel
rdd = rdd.cache()
result_rdd = rdd.map(lambda x: x) # sparkxshards will uncache the rdd when gc
spark_xshards = SparkXShards(result_rdd)
return spark_xshards
def predict(self,
data_creator,
batch_size=None,
verbose=1,
steps=None,
callbacks=None,
data_config=None,
feature_cols=None):
"""Evaluates the model on the validation data set."""
logger.info("Starting predict step.")
params = dict(
verbose=verbose,
batch_size=batch_size,
steps=steps,
callbacks=callbacks,
data_config=data_config,
)
from zoo.orca.data import SparkXShards
from pyspark.sql import DataFrame
if isinstance(data_creator, DataFrame):
assert feature_cols is not None,\
"feature_col must be provided if data_creator is a spark dataframe"
schema = data_creator.schema
numpy_rdd = data_creator.rdd.map(lambda row: convert_row_to_numpy(
row, schema, feature_cols, None))
shard_rdd = numpy_rdd.mapPartitions(
lambda x: arrays2dict(x, feature_cols, None))
data_creator = SparkXShards(shard_rdd)
if isinstance(data_creator, SparkXShards):
ray_xshards = RayXShards.from_spark_xshards(data_creator)
def transform_func(worker, shards_ref):
params["data_creator"] = shards_ref_to_creator(shards_ref)
return worker.predict.remote(**params)
stats_shards = ray_xshards.transform_shards_with_actors(
self.remote_workers, transform_func, gang_scheduling=False)
spark_xshards = stats_shards.to_spark_xshards()
else:
raise ValueError("Only xshards is supported for predict")
return spark_xshards
def _read_as_xshards(path):
rdd, schema = ParquetDataset._read_as_dict_rdd(path)
def merge_records(schema, iter):
l = list(iter)
result = {}
for k in schema.keys():
result[k] = []
for i, rec in enumerate(l):
for k in schema.keys():
result[k].append(rec[k])
for k, v in schema.items():
if not v.feature_type == FeatureType.IMAGE:
result[k] = np.stack(result[k])
return [result]
result_rdd = rdd.mapPartitions(lambda iter: merge_records(schema, iter))
xshards = SparkXShards(result_rdd)
return xshards
def read_parquet(file_path, columns=None, schema=None, **options):
"""
Read parquet files to SparkXShards of pandas DataFrames.
:param file_path: Parquet file path, a list of multiple parquet file paths, or a directory
containing parquet files. Local file system, HDFS, and AWS S3 are supported.
:param columns: list of column name, default=None.
If not None, only these columns will be read from the file.
:param schema: pyspark.sql.types.StructType for the input schema or
a DDL-formatted string (For example col0 INT, col1 DOUBLE).
:param options: other options for reading parquet.
:return: An instance of SparkXShards.
"""
sc = init_nncontext()
from pyspark.sql import SQLContext
sqlContext = SQLContext.getOrCreate(sc)
spark = sqlContext.sparkSession
# df = spark.read.parquet(file_path)
df = spark.read.load(file_path, "parquet", schema=schema, **options)
if columns:
df = df.select(*columns)
def to_pandas(columns):
def f(iter):
import pandas as pd
data = list(iter)
pd_df = pd.DataFrame(data, columns=columns)
return [pd_df]
return f
pd_rdd = df.rdd.mapPartitions(to_pandas(df.columns))
try:
data_shards = SparkXShards(pd_rdd)
except Exception as e:
print("An error occurred when reading parquet files")
raise e
return data_shards
def read_parquet(file_path, columns=None, **kwargs):
"""
Read parquet files to SparkXShards of pandas DataFrames.
:param file_path: Parquet file path, a list of multiple parquet file paths, or a directory
containing parquet files. Local file system, HDFS, and AWS S3 are supported.
:param columns: list of column name, default=None.
If not None, only these columns will be read from the file.
:param kwargs: Any additional kwargs.
:return: An instance of SparkXShards.
"""
sc = init_nncontext()
from pyspark.sql import SQLContext
sqlContext = SQLContext.getOrCreate(sc)
spark = sqlContext.sparkSession
df = spark.read.parquet(file_path)
if columns:
df = df.select(*columns)
def to_pandas(columns):
def f(iter):
import pandas as pd
data = list(iter)
pd_df = pd.DataFrame(data, columns=columns)
return [pd_df]
return f
pd_rdd = df.rdd.mapPartitions(to_pandas(df.columns))
try:
data_shards = SparkXShards(pd_rdd)
except Exception as e:
print("An error occurred when reading parquet files")
raise e
return data_shards
def predict(self, data, feature_cols=None, batch_size=4):
"""
Predict input data
:param batch_size: Int. Set batch Size, default is 4.
:param data: data to be predicted. XShards, Spark DataFrame, numpy array and list of numpy
arrays are supported. If data is XShards, each partition is a dictionary of {'x':
feature}, where feature(label) is a numpy array or a list of numpy arrays.
:param feature_cols: Feature column name(s) of data. Only used when data is a Spark
DataFrame. Default: None.
:return: predicted result.
If the input data is XShards, the predict result is a XShards, each partition
of the XShards is a dictionary of {'prediction': result}, where the result is a
numpy array or a list of numpy arrays.
If the input data is numpy arrays or list of numpy arrays, the predict result is
a numpy array or a list of numpy arrays.
"""
sc = init_nncontext()
model_bytes_broadcast = sc.broadcast(self.model_bytes)
weight_bytes_broadcast = sc.broadcast(self.weight_bytes)
def partition_inference(partition):
model_bytes = model_bytes_broadcast.value
weight_bytes = weight_bytes_broadcast.value
partition = list(partition)
data_num = len(partition)
ie = IECore()
config = {'CPU_THREADS_NUM': str(self.core_num)}
ie.set_config(config, 'CPU')
net = ie.read_network(model=model_bytes,
weights=weight_bytes,
init_from_buffer=True)
net.batch_size = batch_size
local_model = ie.load_network(network=net,
device_name="CPU",
num_requests=data_num)
inputs = list(iter(local_model.requests[0].input_blobs))
outputs = list(iter(local_model.requests[0].output_blobs))
assert len(
outputs) != 0, "The number of model outputs should not be 0."
def add_elem(d):
d_len = len(d)
if d_len < batch_size:
rep_time = [1] * (d_len - 1)
rep_time.append(batch_size - d_len + 1)
return np.repeat(d, rep_time, axis=0), d_len
else:
return d, d_len
results = []
for idx, batch_data in enumerate(partition):
infer_request = local_model.requests[idx]
input_dict = dict()
elem_num = 0
if isinstance(batch_data, list):
for i, input in enumerate(inputs):
input_dict[input], elem_num = add_elem(batch_data[i])
else:
input_dict[inputs[0]], elem_num = add_elem(batch_data)
infer_request.infer(input_dict)
if len(outputs) == 1:
results.append(infer_request.output_blobs[
outputs[0]].buffer[:elem_num])
else:
results.append(
list(
map(
lambda output: infer_request.output_blobs[
output].buffer[:elem_num], outputs)))
return results
def predict_transform(dict_data, batch_size):
assert isinstance(dict_data, dict), "each shard should be an dict"
assert "x" in dict_data, "key x should in each shard"
feature_data = dict_data["x"]
if isinstance(feature_data, np.ndarray):
assert feature_data.shape[0] <= batch_size, \
"The batch size of input data (the second dim) should be less than the model " \
"batch size, otherwise some inputs will be ignored."
elif isinstance(feature_data, list):
for elem in feature_data:
assert isinstance(elem, np.ndarray), "Each element in the x list should be " \
"a ndarray, but get " + \
elem.__class__.__name__
assert elem.shape[0] <= batch_size, "The batch size of each input data (the " \
"second dim) should be less than the " \
"model batch size, otherwise some inputs " \
"will be ignored."
else:
raise ValueError(
"x in each shard should be a ndarray or a list of ndarray."
)
return feature_data
if isinstance(data, DataFrame):
from zoo.orca.learn.utils import dataframe_to_xshards, convert_predict_rdd_to_dataframe
xshards, _ = dataframe_to_xshards(data,
validation_data=None,
feature_cols=feature_cols,
label_cols=None,
mode="predict")
transformed_data = xshards.transform_shard(predict_transform,
batch_size)
result_rdd = transformed_data.rdd.mapPartitions(
lambda iter: partition_inference(iter))
return convert_predict_rdd_to_dataframe(
data, result_rdd.flatMap(lambda data: data))
elif isinstance(data, SparkXShards):
transformed_data = data.transform_shard(predict_transform,
batch_size)
result_rdd = transformed_data.rdd.mapPartitions(
lambda iter: partition_inference(iter))
def update_result_shard(data):
shard, y = data
shard["prediction"] = y
return shard
return SparkXShards(
data.rdd.zip(result_rdd).map(update_result_shard))
elif isinstance(data, (np.ndarray, list)):
if isinstance(data, np.ndarray):
split_num = math.ceil(len(data) / batch_size)
arrays = np.array_split(data, split_num)
num_slices = min(split_num, self.node_num)
data_rdd = sc.parallelize(arrays, numSlices=num_slices)
elif isinstance(data, list):
flattened = nest.flatten(data)
data_length = len(flattened[0])
data_to_be_rdd = []
split_num = math.ceil(flattened[0].shape[0] / batch_size)
num_slices = min(split_num, self.node_num)
for i in range(split_num):
data_to_be_rdd.append([])
for x in flattened:
assert isinstance(x, np.ndarray), "the data in the data list should be " \
"ndarrays, but get " + \
x.__class__.__name__
assert len(x) == data_length, \
"the ndarrays in data must all have the same size in first dimension" \
", got first ndarray of size {} and another {}".format(data_length, len(x))
x_parts = np.array_split(x, split_num)
for idx, x_part in enumerate(x_parts):
data_to_be_rdd[idx].append(x_part)
data_to_be_rdd = [
nest.pack_sequence_as(data, shard)
for shard in data_to_be_rdd
]
data_rdd = sc.parallelize(data_to_be_rdd, numSlices=num_slices)
print("Partition number: ", data_rdd.getNumPartitions())
result_rdd = data_rdd.mapPartitions(
lambda iter: partition_inference(iter))
result_arr_list = result_rdd.collect()
result_arr = None
if isinstance(result_arr_list[0], list):
result_arr = [
np.concatenate([r[i] for r in result_arr_list], axis=0)
for i in range(len(result_arr_list[0]))
]
elif isinstance(result_arr_list[0], np.ndarray):
result_arr = np.concatenate(result_arr_list, axis=0)
return result_arr
else:
raise ValueError(
"Only XShards, Spark DataFrame, a numpy array and a list of numpy arr"
"ays are supported as input data, but get " +
data.__class__.__name__)
def read_file_spark(file_path, file_type, **kwargs):
sc = init_nncontext()
file_url_splits = file_path.split("://")
prefix = file_url_splits[0]
node_num, core_num = get_node_and_core_number()
file_paths = []
if isinstance(file_path, list):
[
file_paths.extend(extract_one_path(path, file_type, os.environ))
for path in file_path
]
else:
file_paths = extract_one_path(file_path, file_type, os.environ)
if not file_paths:
raise Exception(
"The file path is invalid/empty or does not include csv/json files"
)
if ZooContext.orca_pandas_read_backend == "pandas":
num_files = len(file_paths)
total_cores = node_num * core_num
num_partitions = num_files if num_files < total_cores else total_cores
rdd = sc.parallelize(file_paths, num_partitions)
if prefix == "hdfs":
pd_rdd = rdd.mapPartitions(
lambda iter: read_pd_hdfs_file_list(iter, file_type, **kwargs))
elif prefix == "s3":
pd_rdd = rdd.mapPartitions(
lambda iter: read_pd_s3_file_list(iter, file_type, **kwargs))
else:
def loadFile(iterator):
for x in iterator:
df = read_pd_file(x, file_type, **kwargs)
yield df
pd_rdd = rdd.mapPartitions(loadFile)
else:
from pyspark.sql import SQLContext
sqlContext = SQLContext.getOrCreate(sc)
spark = sqlContext.sparkSession
# TODO: add S3 confidentials
if file_type == "json":
df = spark.read.json(file_paths, **kwargs)
elif file_type == "csv":
df = spark.read.csv(file_paths, **kwargs)
else:
raise Exception("Unsupported file type")
if df.rdd.getNumPartitions() < node_num:
df = df.repartition(node_num)
def to_pandas(columns):
def f(iter):
import pandas as pd
data = list(iter)
yield pd.DataFrame(data, columns=columns)
return f
pd_rdd = df.rdd.mapPartitions(to_pandas(df.columns))
data_shards = SparkXShards(pd_rdd)
return data_shards
def fit(
self,
data_creator,
epochs=1,
verbose=1,
callbacks=None,
validation_data_creator=None,
class_weight=None,
steps_per_epoch=None,
validation_steps=None,
validation_freq=1,
data_config=None,
feature_cols=None,
label_cols=None,
):
"""Runs a training epoch."""
params = dict(epochs=epochs,
verbose=verbose,
callbacks=callbacks,
class_weight=class_weight,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps,
validation_freq=validation_freq,
data_config=data_config)
from zoo.orca.data import SparkXShards
from pyspark.sql import DataFrame
if isinstance(data_creator, DataFrame):
assert feature_cols is not None,\
"feature_col must be provided if data_creator is a spark dataframe"
assert label_cols is not None,\
"label_cols must be provided if data_creator is a spark dataframe"
schema = data_creator.schema
numpy_rdd = data_creator.rdd.map(lambda row: convert_row_to_numpy(
row, schema, feature_cols, label_cols))
shard_rdd = numpy_rdd.mapPartitions(
lambda x: arrays2dict(x, feature_cols, label_cols))
data_creator = SparkXShards(shard_rdd)
if isinstance(data_creator, SparkXShards):
max_length, ray_xshards = process_spark_xshards(
data_creator, self.num_workers)
if validation_data_creator is None:
def transform_func(worker, shards_ref):
params["data_creator"] = shards_ref_to_creator(shards_ref)
return worker.step.remote(**params)
stats_shards = ray_xshards.transform_shards_with_actors(
self.remote_workers, transform_func, gang_scheduling=True)
else:
val_max_length, val_ray_xshards = process_spark_xshards(
validation_data_creator, self.num_workers)
def zip_func(worker, this_shards_ref, that_shards_ref):
params["data_creator"] = shards_ref_to_creator(
this_shards_ref)
params["validation_data_creator"] =\
shards_ref_to_creator(that_shards_ref)
return worker.step.remote(**params)
stats_shards = ray_xshards.zip_shards_with_actors(
val_ray_xshards,
self.remote_workers,
zip_func,
gang_scheduling=True)
worker_stats = stats_shards.collect()
else:
params["data_creator"] = data_creator
params["validation_data_creator"] = validation_data_creator
params_list = [params] * self.num_workers
worker_stats = ray.get([
self.remote_workers[i].step.remote(**params_list[i])
for i in range(self.num_workers)
])
worker_stats = list(itertools.chain.from_iterable(worker_stats))
stats = worker_stats[0].copy()
return stats
def test_nnEstimator(self):
from zoo.pipeline.nnframes import NNModel
linear_model = Sequential().add(Linear(2, 2))
mse_criterion = MSECriterion()
df = self.get_estimator_df()
est = Estimator.from_bigdl(model=linear_model,
loss=mse_criterion,
optimizer=Adam(),
feature_preprocessing=SeqToTensor([2]),
label_preprocessing=SeqToTensor([2]))
res0 = est.predict(df)
res0_c = res0.collect()
est.fit(df, 1, batch_size=4)
nn_model = NNModel(est.get_model(),
feature_preprocessing=SeqToTensor([2]))
res1 = nn_model.transform(df)
res2 = est.predict(df)
res1_c = res1.collect()
res2_c = res2.collect()
assert type(res1).__name__ == 'DataFrame'
assert type(res2).__name__ == 'DataFrame'
assert len(res1_c) == len(res2_c)
for idx in range(len(res1_c)):
assert res1_c[idx]["prediction"] == res2_c[idx]["prediction"]
with tempfile.TemporaryDirectory() as tempdirname:
temp_path = os.path.join(tempdirname, "model")
est.save(temp_path)
est2 = Estimator.from_bigdl(model=linear_model, loss=mse_criterion)
est2.load(temp_path,
optimizer=Adam(),
loss=mse_criterion,
feature_preprocessing=SeqToTensor([2]),
label_preprocessing=SeqToTensor([2]))
est2.set_constant_gradient_clipping(0.1, 1.2)
est2.clear_gradient_clipping()
res3 = est2.predict(df)
res3_c = res3.collect()
assert type(res3).__name__ == 'DataFrame'
assert len(res1_c) == len(res3_c)
for idx in range(len(res1_c)):
assert res1_c[idx]["prediction"] == res3_c[idx]["prediction"]
est2.fit(df, 4, batch_size=4)
data = self.sc.parallelize([((2.0, 1.0), (1.0, 2.0)),
((1.0, 2.0), (2.0, 1.0)),
((2.0, 1.0), (1.0, 2.0)),
((1.0, 2.0), (2.0, 1.0))])
data_shard = SparkXShards(data)
data_shard = data_shard.transform_shard(
lambda feature_label_tuple: {
"x": [
np.expand_dims(np.array(feature_label_tuple[0][0]), axis=0
),
np.expand_dims(np.array(feature_label_tuple[0][1]), axis=0)
],
"y": [
np.expand_dims(np.array(feature_label_tuple[1][0]), axis=0
),
np.expand_dims(np.array(feature_label_tuple[1][1]), axis=0)
]
})
res4 = est.predict(data_shard)
res4_c = res4.collect()
assert type(res4).__name__ == 'SparkXShards'
for idx in range(len(res4_c)):
assert abs(res4_c[idx]["prediction"][0][0] -
res3_c[idx]["prediction"][0]) == 0
assert abs(res4_c[idx]["prediction"][0][1] -
res3_c[idx]["prediction"][1]) == 0
est.fit(data_shard, 1, batch_size=4)
res5 = est.predict(data_shard)
res5_c = res5.collect()
res6 = est.predict(df)
res6_c = res6.collect()
for idx in range(len(res5_c)):
assert abs(res5_c[idx]["prediction"][0][0] -
res6_c[idx]["prediction"][0]) == 0
assert abs(res5_c[idx]["prediction"][0][1] -
res6_c[idx]["prediction"][1]) == 0
def read_file_spark(file_path, file_type, **kwargs):
sc = init_nncontext()
node_num, core_num = get_node_and_core_number()
backend = OrcaContext.pandas_read_backend
if backend == "pandas":
file_url_splits = file_path.split("://")
prefix = file_url_splits[0]
file_paths = []
if isinstance(file_path, list):
[
file_paths.extend(extract_one_path(path, os.environ))
for path in file_path
]
else:
file_paths = extract_one_path(file_path, os.environ)
if not file_paths:
raise Exception(
"The file path is invalid or empty, please check your data")
num_files = len(file_paths)
total_cores = node_num * core_num
num_partitions = num_files if num_files < total_cores else total_cores
rdd = sc.parallelize(file_paths, num_partitions)
if prefix == "hdfs":
pd_rdd = rdd.mapPartitions(
lambda iter: read_pd_hdfs_file_list(iter, file_type, **kwargs))
elif prefix == "s3":
pd_rdd = rdd.mapPartitions(
lambda iter: read_pd_s3_file_list(iter, file_type, **kwargs))
else:
def loadFile(iterator):
for x in iterator:
df = read_pd_file(x, file_type, **kwargs)
yield df
pd_rdd = rdd.mapPartitions(loadFile)
else: # Spark backend; spark.read.csv/json accepts a folder path as input
assert file_type == "json" or file_type == "csv", \
"Unsupported file type: %s. Only csv and json files are supported for now" % file_type
from pyspark.sql import SQLContext
sqlContext = SQLContext.getOrCreate(sc)
spark = sqlContext.sparkSession
# TODO: add S3 confidentials
# The following implementation is adapted from
# https://github.com/databricks/koalas/blob/master/databricks/koalas/namespace.py
# with some modifications.
if "mangle_dupe_cols" in kwargs:
assert kwargs[
"mangle_dupe_cols"], "mangle_dupe_cols can only be True"
kwargs.pop("mangle_dupe_cols")
if "parse_dates" in kwargs:
assert not kwargs["parse_dates"], "parse_dates can only be False"
kwargs.pop("parse_dates")
names = kwargs.get("names", None)
if "names" in kwargs:
kwargs.pop("names")
usecols = kwargs.get("usecols", None)
if "usecols" in kwargs:
kwargs.pop("usecols")
dtype = kwargs.get("dtype", None)
if "dtype" in kwargs:
kwargs.pop("dtype")
squeeze = kwargs.get("squeeze", False)
if "squeeze" in kwargs:
kwargs.pop("squeeze")
index_col = kwargs.get("index_col", None)
if "index_col" in kwargs:
kwargs.pop("index_col")
if file_type == "csv":
# Handle pandas-compatible keyword arguments
kwargs["inferSchema"] = True
header = kwargs.get("header", "infer")
if isinstance(names, str):
kwargs["schema"] = names
if header == "infer":
header = 0 if names is None else None
if header == 0:
kwargs["header"] = True
elif header is None:
kwargs["header"] = False
else:
raise ValueError("Unknown header argument {}".format(header))
if "quotechar" in kwargs:
quotechar = kwargs["quotechar"]
kwargs.pop("quotechar")
kwargs["quote"] = quotechar
if "escapechar" in kwargs:
escapechar = kwargs["escapechar"]
kwargs.pop("escapechar")
kwargs["escape"] = escapechar
# sep and comment are the same as pandas
if "comment" in kwargs:
comment = kwargs["comment"]
if not isinstance(comment, str) or len(comment) != 1:
raise ValueError(
"Only length-1 comment characters supported")
df = spark.read.csv(file_path, **kwargs)
if header is None:
df = df.selectExpr(*[
"`%s` as `%s`" % (field.name, i)
for i, field in enumerate(df.schema)
])
else:
df = spark.read.json(file_path, **kwargs)
# Handle pandas-compatible postprocessing arguments
if usecols is not None and not callable(usecols):
usecols = list(usecols)
renamed = False
if isinstance(names, list):
if len(set(names)) != len(names):
raise ValueError(
"Found duplicate names, please check your names input")
if usecols is not None:
if not callable(usecols):
# usecols is list
if len(names) != len(usecols) and len(names) != len(
df.schema):
raise ValueError("Passed names did not match usecols")
if len(names) == len(df.schema):
df = df.selectExpr(*[
"`%s` as `%s`" % (field.name, name)
for field, name in zip(df.schema, names)
])
renamed = True
else:
if len(names) != len(df.schema):
raise ValueError(
"The number of names [%s] does not match the number "
"of columns [%d]. Try names by a Spark SQL DDL-formatted "
"string." % (len(names), len(df.schema)))
df = df.selectExpr(*[
"`%s` as `%s`" % (field.name, name)
for field, name in zip(df.schema, names)
])
renamed = True
index_map = dict([(i, field.name)
for i, field in enumerate(df.schema)])
if usecols is not None:
if callable(usecols):
cols = [
field.name for field in df.schema if usecols(field.name)
]
missing = []
elif all(isinstance(col, int) for col in usecols):
cols = [
field.name for i, field in enumerate(df.schema)
if i in usecols
]
missing = [
col for col in usecols
if col >= len(df.schema) or df.schema[col].name not in cols
]
elif all(isinstance(col, str) for col in usecols):
cols = [
field.name for field in df.schema if field.name in usecols
]
if isinstance(names, list):
missing = [c for c in usecols if c not in names]
else:
missing = [col for col in usecols if col not in cols]
else:
raise ValueError(
"usecols must only be list-like of all strings, "
"all unicode, all integers or a callable.")
if len(missing) > 0:
raise ValueError(
"usecols do not match columns, columns expected but not found: %s"
% missing)
if len(cols) > 0:
df = df.select(cols)
if isinstance(names, list):
if not renamed:
df = df.selectExpr(*[
"`%s` as `%s`" % (col, name)
for col, name in zip(cols, names)
])
# update index map after rename
for index, col in index_map.items():
if col in cols:
index_map[index] = names[cols.index(col)]
if df.rdd.getNumPartitions() < node_num:
df = df.repartition(node_num)
def to_pandas(columns, squeeze=False, index_col=None):
def f(iter):
import pandas as pd
data = list(iter)
pd_df = pd.DataFrame(data, columns=columns)
if dtype is not None:
if isinstance(dtype, dict):
for col, type in dtype.items():
if isinstance(col, str):
if col not in pd_df.columns:
raise ValueError(
"column to be set type is not"
" in current dataframe")
pd_df[col] = pd_df[col].astype(type)
elif isinstance(col, int):
if index_map[col] not in pd_df.columns:
raise ValueError(
"column index to be set type is not"
" in current dataframe")
pd_df[index_map[col]] = pd_df[
index_map[col]].astype(type)
else:
pd_df = pd_df.astype(dtype)
if squeeze and len(pd_df.columns) == 1:
pd_df = pd_df.iloc[:, 0]
if index_col:
pd_df = pd_df.set_index(index_col)
return [pd_df]
return f
pd_rdd = df.rdd.mapPartitions(to_pandas(df.columns, squeeze,
index_col))
try:
data_shards = SparkXShards(pd_rdd)
except Exception as e:
alternative_backend = "pandas" if backend == "spark" else "spark"
print(
"An error occurred when reading files with '%s' backend, you may switch to '%s' "
"backend for another try. You can set the backend using "
"OrcaContext.pandas_read_backend" % (backend, alternative_backend))
raise e
return data_shards
def predict(self, data):
"""
Predict input data
:param data: data to be predicted. XShards, numpy array and list of numpy arrays are
supported. If data is XShards, each partition is a dictionary of {'x': feature},
where feature(label) is a numpy array or a list of numpy arrays.
:return: predicted result.
If the input data is XShards, the predict result is a XShards, each partition of the
XShards is a dictionary of {'prediction': result}, where the result is a numpy array or
a list of numpy arrays.
If the input data is numpy arrays or list of numpy arrays, the predict result is a numpy
array or a list of numpy arrays.
"""
def predict_transform(dict_data, batch_size):
assert isinstance(dict_data, dict), "each shard should be an dict"
assert "x" in dict_data, "key x should in each shard"
feature_data = dict_data["x"]
if isinstance(feature_data, np.ndarray):
assert feature_data.shape[1] <= batch_size, \
"The batch size of input data (the second dim) should be less than the model " \
"batch size, otherwise some inputs will be ignored."
elif isinstance(feature_data, list):
for elem in feature_data:
assert isinstance(elem, np.ndarray), "Each element in the x list should be " \
"a ndarray, but get " + \
elem.__class__.__name__
assert elem.shape[1] <= batch_size, "The batch size of each input data (the " \
"second dim) should be less than the " \
"model batch size, otherwise some inputs " \
"will be ignored."
else:
raise ValueError(
"x in each shard should be a ndarray or a list of ndarray."
)
return dict_data["x"]
sc = init_nncontext()
if isinstance(data, SparkXShards):
assert sc is not None, "You should pass sc(spark context) if data is a XShards."
from zoo.orca.learn.utils import convert_predict_rdd_to_xshard
transformed_data = data.transform_shard(predict_transform,
self.batch_size)
result_rdd = self.model.distributed_predict(
transformed_data.rdd, sc)
def update_shard(data):
shard, y = data
shard["prediction"] = y
return shard
return SparkXShards(data.rdd.zip(result_rdd).map(update_shard))
elif isinstance(data, (np.ndarray, list)):
total_core_num = self.core_num * self.node_num
if isinstance(data, np.ndarray):
assert data.shape[1] <= self.batch_size, "The batch size of input data (the " \
"second dim) should be less than the " \
"model batch size, otherwise some " \
"inputs will be ignored."
split_num = min(total_core_num, data.shape[0])
arrays = np.array_split(data, split_num)
data_rdd = sc.parallelize(arrays, numSlices=split_num)
elif isinstance(data, list):
flattened = nest.flatten(data)
data_length = len(flattened[0])
data_to_be_rdd = []
split_num = min(total_core_num, flattened[0].shape[0])
for i in range(split_num):
data_to_be_rdd.append([])
for x in flattened:
assert isinstance(x, np.ndarray), "the data in the data list should be " \
"ndarrays, but get " + \
x.__class__.__name__
assert len(x) == data_length, \
"the ndarrays in data must all have the same size in first dimension" \
", got first ndarray of size {} and another {}".format(data_length, len(x))
assert x.shape[1] <= self.batch_size, "The batch size of each input data (" \
"the second dim) should be less than " \
"the model batch size, otherwise some " \
"inputs will be ignored."
x_parts = np.array_split(x, split_num)
for idx, x_part in enumerate(x_parts):
data_to_be_rdd[idx].append(x_part)
data_to_be_rdd = [
nest.pack_sequence_as(data, shard)
for shard in data_to_be_rdd
]
data_rdd = sc.parallelize(data_to_be_rdd, numSlices=split_num)
result_rdd = self.model.distributed_predict(data_rdd, sc)
result_arr_list = result_rdd.collect()
result_arr = np.concatenate(result_arr_list, axis=0)
return result_arr
else:
raise ValueError(
"Only XShards, a numpy array and a list of numpy arrays are supported "
"as input data, but get " + data.__class__.__name__)
def predict(self, data, feature_cols=None):
"""
Predict input data
:param data: data to be predicted. XShards, Spark DataFrame, numpy array and list of numpy
arrays are supported. If data is XShards, each partition is a dictionary of {'x':
feature}, where feature(label) is a numpy array or a list of numpy arrays.
:param feature_cols: Feature column name(s) of data. Only used when data is a Spark
DataFrame. Default: None.
:return: predicted result.
If the input data is XShards, the predict result is a XShards, each partition
of the XShards is a dictionary of {'prediction': result}, where the result is a
numpy array or a list of numpy arrays.
If the input data is numpy arrays or list of numpy arrays, the predict result is
a numpy array or a list of numpy arrays.
"""
from pyspark.sql import DataFrame
def predict_transform(dict_data, batch_size):
assert isinstance(dict_data, dict), "each shard should be an dict"
assert "x" in dict_data, "key x should in each shard"
feature_data = dict_data["x"]
if isinstance(feature_data, np.ndarray):
assert feature_data.shape[0] <= batch_size, \
"The batch size of input data (the second dim) should be less than the model " \
"batch size, otherwise some inputs will be ignored."
elif isinstance(feature_data, list):
for elem in feature_data:
assert isinstance(elem, np.ndarray), "Each element in the x list should be " \
"a ndarray, but get " + \
elem.__class__.__name__
assert elem.shape[0] <= batch_size, "The batch size of each input data (the " \
"second dim) should be less than the " \
"model batch size, otherwise some inputs " \
"will be ignored."
else:
raise ValueError(
"x in each shard should be a ndarray or a list of ndarray."
)
return feature_data
sc = init_nncontext()
if isinstance(data, DataFrame):
from zoo.orca.learn.utils import dataframe_to_xshards, convert_predict_rdd_to_dataframe
xshards, _ = dataframe_to_xshards(data,
validation_data=None,
feature_cols=feature_cols,
label_cols=None,
mode="predict")
transformed_data = xshards.transform_shard(predict_transform,
self.batch_size)
result_rdd = self.model.distributed_predict(
transformed_data.rdd, sc)
def delete_useless_result(data):
shard, y = data
data_length = len(shard["x"])
return y[:data_length]
result_rdd = xshards.rdd.zip(result_rdd).map(delete_useless_result)
return convert_predict_rdd_to_dataframe(
data, result_rdd.flatMap(lambda data: data))
elif isinstance(data, SparkXShards):
transformed_data = data.transform_shard(predict_transform,
self.batch_size)
result_rdd = self.model.distributed_predict(
transformed_data.rdd, sc)
def update_shard(data):
shard, y = data
data_length = len(shard["x"])
shard["prediction"] = y[:data_length]
return shard
return SparkXShards(data.rdd.zip(result_rdd).map(update_shard))
elif isinstance(data, (np.ndarray, list)):
if isinstance(data, np.ndarray):
split_num = math.ceil(len(data) / self.batch_size)
arrays = np.array_split(data, split_num)
data_length_list = list(map(lambda arr: len(arr), arrays))
data_rdd = sc.parallelize(arrays, numSlices=split_num)
elif isinstance(data, list):
flattened = nest.flatten(data)
data_length = len(flattened[0])
data_to_be_rdd = []
split_num = math.ceil(flattened[0].shape[0] / self.batch_size)
for i in range(split_num):
data_to_be_rdd.append([])
for x in flattened:
assert isinstance(x, np.ndarray), "the data in the data list should be " \
"ndarrays, but get " + \
x.__class__.__name__
assert len(x) == data_length, \
"the ndarrays in data must all have the same size in first dimension" \
", got first ndarray of size {} and another {}".format(data_length, len(x))
x_parts = np.array_split(x, split_num)
for idx, x_part in enumerate(x_parts):
data_to_be_rdd[idx].append(x_part)
data_length_list = list(
map(lambda arr: len(arr), x_part))
data_to_be_rdd = [
nest.pack_sequence_as(data, shard)
for shard in data_to_be_rdd
]
data_rdd = sc.parallelize(data_to_be_rdd, numSlices=split_num)
result_rdd = self.model.distributed_predict(data_rdd, sc)
result_arr_list = result_rdd.collect()
for i in range(0, len(result_arr_list)):
result_arr_list[i] = result_arr_list[i][:data_length_list[i]]
result_arr = np.concatenate(result_arr_list, axis=0)
return result_arr
else:
raise ValueError(
"Only XShards, Spark DataFrame, a numpy array and a list of numpy arr"
"ays are supported as input data, but get " +
data.__class__.__name__)
