forked from huawei/mindspore2022
519 lines
16 KiB
Python
519 lines
16 KiB
Python
# Copyright 2020 Huawei Technologies Co., Ltd
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# ==============================================================================
|
|
"""
|
|
Testing profiling support in DE
|
|
"""
|
|
import json
|
|
import os
|
|
import numpy as np
|
|
import mindspore.common.dtype as mstype
|
|
import mindspore.dataset as ds
|
|
import mindspore.dataset.transforms.c_transforms as C
|
|
import mindspore.dataset.vision.c_transforms as vision
|
|
|
|
FILES = ["../data/dataset/testTFTestAllTypes/test.data"]
|
|
DATASET_ROOT = "../data/dataset/testTFTestAllTypes/"
|
|
SCHEMA_FILE = "../data/dataset/testTFTestAllTypes/datasetSchema.json"
|
|
|
|
PIPELINE_FILE = "./pipeline_profiling_1.json"
|
|
CPU_UTIL_FILE = "./minddata_cpu_utilization_1.json"
|
|
DATASET_ITERATOR_FILE = "./dataset_iterator_profiling_1.txt"
|
|
|
|
|
|
def set_profiling_env_var():
|
|
"""
|
|
Set the MindData Profiling environment variables
|
|
"""
|
|
os.environ['PROFILING_MODE'] = 'true'
|
|
os.environ['MINDDATA_PROFILING_DIR'] = '.'
|
|
os.environ['DEVICE_ID'] = '1'
|
|
|
|
|
|
def delete_profiling_files():
|
|
"""
|
|
Delete the MindData profiling files generated from the test.
|
|
Also disable the MindData Profiling environment variables.
|
|
"""
|
|
# Delete MindData profiling files
|
|
os.remove(PIPELINE_FILE)
|
|
os.remove(CPU_UTIL_FILE)
|
|
os.remove(DATASET_ITERATOR_FILE)
|
|
|
|
# Disable MindData Profiling environment variables
|
|
del os.environ['PROFILING_MODE']
|
|
del os.environ['MINDDATA_PROFILING_DIR']
|
|
del os.environ['DEVICE_ID']
|
|
|
|
|
|
def confirm_cpuutil(num_pipeline_ops):
|
|
"""
|
|
Confirm CPU utilization JSON file when <num_pipeline_ops> are in the pipeline
|
|
"""
|
|
with open(CPU_UTIL_FILE) as file1:
|
|
data = json.load(file1)
|
|
op_info = data["op_info"]
|
|
# Confirm <num_pipeline_ops>+1 ops in CPU util file (including op_id=-1 for monitor thread)
|
|
assert len(op_info) == num_pipeline_ops + 1
|
|
|
|
|
|
def test_profiling_simple_pipeline():
|
|
"""
|
|
Generator -> Shuffle -> Batch
|
|
"""
|
|
set_profiling_env_var()
|
|
|
|
source = [(np.array([x]),) for x in range(1024)]
|
|
data1 = ds.GeneratorDataset(source, ["data"])
|
|
data1 = data1.shuffle(64)
|
|
data1 = data1.batch(32)
|
|
# try output shape type and dataset size and make sure no profiling file is generated
|
|
assert data1.output_shapes() == [[32, 1]]
|
|
assert [str(tp) for tp in data1.output_types()] == ["int64"]
|
|
assert data1.get_dataset_size() == 32
|
|
|
|
# Confirm profiling files do not (yet) exist
|
|
assert os.path.exists(PIPELINE_FILE) is False
|
|
assert os.path.exists(CPU_UTIL_FILE) is False
|
|
assert os.path.exists(DATASET_ITERATOR_FILE) is False
|
|
|
|
try:
|
|
for _ in data1:
|
|
pass
|
|
|
|
# Confirm profiling files now exist
|
|
assert os.path.exists(PIPELINE_FILE) is True
|
|
assert os.path.exists(CPU_UTIL_FILE) is True
|
|
assert os.path.exists(DATASET_ITERATOR_FILE) is True
|
|
|
|
except Exception as error:
|
|
delete_profiling_files()
|
|
raise error
|
|
|
|
else:
|
|
delete_profiling_files()
|
|
|
|
|
|
def test_profiling_complex_pipeline():
|
|
"""
|
|
Generator -> Map ->
|
|
-> Zip
|
|
TFReader -> Shuffle ->
|
|
"""
|
|
set_profiling_env_var()
|
|
|
|
source = [(np.array([x]),) for x in range(1024)]
|
|
data1 = ds.GeneratorDataset(source, ["gen"])
|
|
data1 = data1.map(operations=[(lambda x: x + 1)], input_columns=["gen"])
|
|
|
|
pattern = DATASET_ROOT + "/test.data"
|
|
data2 = ds.TFRecordDataset(pattern, SCHEMA_FILE, shuffle=ds.Shuffle.FILES)
|
|
data2 = data2.shuffle(4)
|
|
|
|
data3 = ds.zip((data1, data2))
|
|
|
|
try:
|
|
for _ in data3:
|
|
pass
|
|
|
|
with open(PIPELINE_FILE) as f:
|
|
data = json.load(f)
|
|
op_info = data["op_info"]
|
|
assert len(op_info) == 5
|
|
for i in range(5):
|
|
if op_info[i]["op_type"] != "ZipOp":
|
|
assert "size" in op_info[i]["metrics"]["output_queue"]
|
|
assert "length" in op_info[i]["metrics"]["output_queue"]
|
|
assert "throughput" in op_info[i]["metrics"]["output_queue"]
|
|
else:
|
|
# Note: Zip is an inline op and hence does not have metrics information
|
|
assert op_info[i]["metrics"] is None
|
|
|
|
# Confirm CPU util JSON file content, when 5 ops are in the pipeline JSON file
|
|
confirm_cpuutil(5)
|
|
|
|
except Exception as error:
|
|
delete_profiling_files()
|
|
raise error
|
|
|
|
else:
|
|
delete_profiling_files()
|
|
|
|
|
|
def test_profiling_inline_ops_pipeline1():
|
|
"""
|
|
Test pipeline with inline ops: Concat and EpochCtrl
|
|
Generator ->
|
|
Concat -> EpochCtrl
|
|
Generator ->
|
|
"""
|
|
set_profiling_env_var()
|
|
|
|
# In source1 dataset: Number of rows is 3; its values are 0, 1, 2
|
|
def source1():
|
|
for i in range(3):
|
|
yield (np.array([i]),)
|
|
|
|
# In source2 dataset: Number of rows is 7; its values are 3, 4, 5 ... 9
|
|
def source2():
|
|
for i in range(3, 10):
|
|
yield (np.array([i]),)
|
|
|
|
data1 = ds.GeneratorDataset(source1, ["col1"])
|
|
data2 = ds.GeneratorDataset(source2, ["col1"])
|
|
data3 = data1.concat(data2)
|
|
|
|
try:
|
|
num_iter = 0
|
|
# Note: Do not explicitly set num_epochs argument in create_tuple_iterator() call
|
|
# Here i refers to index, d refers to data element
|
|
for i, d in enumerate(data3.create_tuple_iterator(output_numpy=True)):
|
|
num_iter += 1
|
|
t = d
|
|
assert i == t[0][0]
|
|
|
|
assert num_iter == 10
|
|
|
|
# Confirm pipeline is created with EpochCtrl op
|
|
with open(PIPELINE_FILE) as f:
|
|
data = json.load(f)
|
|
op_info = data["op_info"]
|
|
assert len(op_info) == 4
|
|
for i in range(4):
|
|
# Note: The following ops are inline ops: Concat, EpochCtrl
|
|
if op_info[i]["op_type"] in ("ConcatOp", "EpochCtrlOp"):
|
|
# Confirm these inline ops do not have metrics information
|
|
assert op_info[i]["metrics"] is None
|
|
else:
|
|
assert "size" in op_info[i]["metrics"]["output_queue"]
|
|
assert "length" in op_info[i]["metrics"]["output_queue"]
|
|
assert "throughput" in op_info[i]["metrics"]["output_queue"]
|
|
|
|
# Confirm CPU util JSON file content, when 4 ops are in the pipeline JSON file
|
|
confirm_cpuutil(4)
|
|
|
|
except Exception as error:
|
|
delete_profiling_files()
|
|
raise error
|
|
|
|
else:
|
|
delete_profiling_files()
|
|
|
|
|
|
def test_profiling_inline_ops_pipeline2():
|
|
"""
|
|
Test pipeline with many inline ops
|
|
Generator -> Rename -> Skip -> Repeat -> Take
|
|
"""
|
|
set_profiling_env_var()
|
|
|
|
# In source1 dataset: Number of rows is 10; its values are 0, 1, 2, 3, 4, 5 ... 9
|
|
def source1():
|
|
for i in range(10):
|
|
yield (np.array([i]),)
|
|
|
|
data1 = ds.GeneratorDataset(source1, ["col1"])
|
|
data1 = data1.rename(input_columns=["col1"], output_columns=["newcol1"])
|
|
data1 = data1.skip(2)
|
|
data1 = data1.repeat(2)
|
|
data1 = data1.take(12)
|
|
|
|
try:
|
|
for _ in data1:
|
|
pass
|
|
|
|
with open(PIPELINE_FILE) as f:
|
|
data = json.load(f)
|
|
op_info = data["op_info"]
|
|
assert len(op_info) == 5
|
|
for i in range(5):
|
|
# Check for these inline ops
|
|
if op_info[i]["op_type"] in ("RenameOp", "RepeatOp", "SkipOp", "TakeOp"):
|
|
# Confirm these inline ops do not have metrics information
|
|
assert op_info[i]["metrics"] is None
|
|
else:
|
|
assert "size" in op_info[i]["metrics"]["output_queue"]
|
|
assert "length" in op_info[i]["metrics"]["output_queue"]
|
|
assert "throughput" in op_info[i]["metrics"]["output_queue"]
|
|
|
|
# Confirm CPU util JSON file content, when 5 ops are in the pipeline JSON file
|
|
confirm_cpuutil(5)
|
|
|
|
except Exception as error:
|
|
delete_profiling_files()
|
|
raise error
|
|
|
|
else:
|
|
delete_profiling_files()
|
|
|
|
|
|
def test_profiling_sampling_interval():
|
|
"""
|
|
Test non-default monitor sampling interval
|
|
"""
|
|
set_profiling_env_var()
|
|
|
|
interval_origin = ds.config.get_monitor_sampling_interval()
|
|
|
|
ds.config.set_monitor_sampling_interval(30)
|
|
interval = ds.config.get_monitor_sampling_interval()
|
|
assert interval == 30
|
|
|
|
source = [(np.array([x]),) for x in range(1024)]
|
|
data1 = ds.GeneratorDataset(source, ["data"])
|
|
data1 = data1.shuffle(64)
|
|
data1 = data1.batch(32)
|
|
|
|
try:
|
|
for _ in data1:
|
|
pass
|
|
|
|
except Exception as error:
|
|
ds.config.set_monitor_sampling_interval(interval_origin)
|
|
delete_profiling_files()
|
|
raise error
|
|
|
|
else:
|
|
ds.config.set_monitor_sampling_interval(interval_origin)
|
|
delete_profiling_files()
|
|
|
|
|
|
def test_profiling_basic_pipeline():
|
|
"""
|
|
Test with this basic pipeline
|
|
Generator -> Map -> Batch -> Repeat -> EpochCtrl
|
|
"""
|
|
set_profiling_env_var()
|
|
|
|
def source1():
|
|
for i in range(8000):
|
|
yield (np.array([i]),)
|
|
|
|
# Create this basic and common pipeline
|
|
# Leaf/Source-Op -> Map -> Batch -> Repeat
|
|
data1 = ds.GeneratorDataset(source1, ["col1"])
|
|
|
|
type_cast_op = C.TypeCast(mstype.int32)
|
|
data1 = data1.map(operations=type_cast_op, input_columns="col1")
|
|
data1 = data1.batch(16)
|
|
data1 = data1.repeat(2)
|
|
|
|
try:
|
|
num_iter = 0
|
|
# Note: If create_tuple_iterator() is called with num_epochs>1, then EpochCtrlOp is added to the pipeline
|
|
for _ in data1.create_dict_iterator(num_epochs=2):
|
|
num_iter += 1
|
|
|
|
assert num_iter == 1000
|
|
|
|
with open(PIPELINE_FILE) as f:
|
|
data = json.load(f)
|
|
op_info = data["op_info"]
|
|
assert len(op_info) == 5
|
|
for i in range(5):
|
|
# Check for inline ops
|
|
if op_info[i]["op_type"] in ("EpochCtrlOp", "RepeatOp"):
|
|
# Confirm these inline ops do not have metrics information
|
|
assert op_info[i]["metrics"] is None
|
|
else:
|
|
assert "size" in op_info[i]["metrics"]["output_queue"]
|
|
assert "length" in op_info[i]["metrics"]["output_queue"]
|
|
assert "throughput" in op_info[i]["metrics"]["output_queue"]
|
|
|
|
# Confirm CPU util JSON file content, when 5 ops are in the pipeline JSON file
|
|
confirm_cpuutil(5)
|
|
|
|
except Exception as error:
|
|
delete_profiling_files()
|
|
raise error
|
|
|
|
else:
|
|
delete_profiling_files()
|
|
|
|
|
|
def test_profiling_cifar10_pipeline():
|
|
"""
|
|
Test with this common pipeline with Cifar10
|
|
Cifar10 -> Map -> Map -> Batch -> Repeat
|
|
"""
|
|
set_profiling_env_var()
|
|
|
|
# Create this common pipeline
|
|
# Cifar10 -> Map -> Map -> Batch -> Repeat
|
|
DATA_DIR_10 = "../data/dataset/testCifar10Data"
|
|
data1 = ds.Cifar10Dataset(DATA_DIR_10, num_samples=8000)
|
|
|
|
type_cast_op = C.TypeCast(mstype.int32)
|
|
data1 = data1.map(operations=type_cast_op, input_columns="label")
|
|
random_horizontal_op = vision.RandomHorizontalFlip()
|
|
data1 = data1.map(operations=random_horizontal_op, input_columns="image")
|
|
|
|
data1 = data1.batch(32)
|
|
data1 = data1.repeat(3)
|
|
|
|
try:
|
|
num_iter = 0
|
|
# Note: If create_tuple_iterator() is called with num_epochs=1, then EpochCtrlOp is NOT added to the pipeline
|
|
for _ in data1.create_dict_iterator(num_epochs=1):
|
|
num_iter += 1
|
|
|
|
assert num_iter == 750
|
|
|
|
with open(PIPELINE_FILE) as f:
|
|
data = json.load(f)
|
|
op_info = data["op_info"]
|
|
assert len(op_info) == 5
|
|
for i in range(5):
|
|
# Check for inline ops
|
|
if op_info[i]["op_type"] == "RepeatOp":
|
|
# Confirm these inline ops do not have metrics information
|
|
assert op_info[i]["metrics"] is None
|
|
else:
|
|
assert "size" in op_info[i]["metrics"]["output_queue"]
|
|
assert "length" in op_info[i]["metrics"]["output_queue"]
|
|
assert "throughput" in op_info[i]["metrics"]["output_queue"]
|
|
|
|
# Confirm CPU util JSON file content, when 5 ops are in the pipeline JSON file
|
|
confirm_cpuutil(5)
|
|
|
|
except Exception as error:
|
|
delete_profiling_files()
|
|
raise error
|
|
|
|
else:
|
|
delete_profiling_files()
|
|
|
|
|
|
def confirm_3ops_in_pipeline():
|
|
with open(PIPELINE_FILE) as file1:
|
|
data = json.load(file1)
|
|
op_info = data["op_info"]
|
|
# Confirm 3 ops in pipeline file
|
|
assert len(op_info) == 3
|
|
for i in range(3):
|
|
assert op_info[i]["op_type"] in ("GeneratorOp", "BatchOp", "EpochCtrlOp")
|
|
|
|
|
|
def confirm_2ops_in_pipeline():
|
|
with open(PIPELINE_FILE) as file1:
|
|
data = json.load(file1)
|
|
op_info = data["op_info"]
|
|
# Confirm 2 ops in pipeline file
|
|
assert len(op_info) == 2
|
|
for i in range(2):
|
|
assert op_info[i]["op_type"] in ("GeneratorOp", "BatchOp")
|
|
|
|
|
|
def test_profiling_epochctrl3():
|
|
"""
|
|
Test with these 2 sequential pipelines:
|
|
1) Generator -> Batch -> EpochCtrl
|
|
2) Generator -> Batch
|
|
Note: This is a simplification of the user scenario to use the same pipeline for training and then evaluation.
|
|
"""
|
|
set_profiling_env_var()
|
|
|
|
source = [(np.array([x]),) for x in range(64)]
|
|
data1 = ds.GeneratorDataset(source, ["data"])
|
|
data1 = data1.batch(32)
|
|
|
|
try:
|
|
# Test A - Call create_dict_iterator with num_epochs>1
|
|
num_iter = 0
|
|
# Note: If create_tuple_iterator() is called with num_epochs>1, then EpochCtrlOp is added to the pipeline
|
|
for _ in data1.create_dict_iterator(num_epochs=2):
|
|
num_iter += 1
|
|
assert num_iter == 2
|
|
|
|
confirm_3ops_in_pipeline()
|
|
confirm_cpuutil(3)
|
|
|
|
# Test B - Call create_dict_iterator with num_epochs=1
|
|
num_iter = 0
|
|
# Note: If create_tuple_iterator() is called with num_epochs=1,
|
|
# then EpochCtrlOp should not be NOT added to the pipeline
|
|
for _ in data1.create_dict_iterator(num_epochs=1):
|
|
num_iter += 1
|
|
assert num_iter == 2
|
|
|
|
# confirm_2ops_in_pipeline()
|
|
# MD BUG: Confirm pipeline file is not changed and wrongly still has 3 ops
|
|
confirm_3ops_in_pipeline()
|
|
# Confirm CPU util file has correct number of ops
|
|
confirm_cpuutil(2)
|
|
|
|
except Exception as error:
|
|
delete_profiling_files()
|
|
raise error
|
|
|
|
else:
|
|
delete_profiling_files()
|
|
|
|
|
|
def test_profiling_epochctrl2():
|
|
"""
|
|
Test with these 2 sequential pipelines:
|
|
1) Generator -> Batch
|
|
2) Generator -> Batch -> EpochCtrl
|
|
"""
|
|
set_profiling_env_var()
|
|
|
|
source = [(np.array([x]),) for x in range(64)]
|
|
data2 = ds.GeneratorDataset(source, ["data"])
|
|
data2 = data2.batch(16)
|
|
|
|
try:
|
|
# Test A - Call create_dict_iterator with num_epochs=1
|
|
num_iter = 0
|
|
# Note: If create_tuple_iterator() is called with num_epochs=1, then EpochCtrlOp is NOT added to the pipeline
|
|
for _ in data2.create_dict_iterator(num_epochs=1):
|
|
num_iter += 1
|
|
assert num_iter == 4
|
|
|
|
confirm_2ops_in_pipeline()
|
|
confirm_cpuutil(2)
|
|
|
|
# Test B - Call create_dict_iterator with num_epochs>1
|
|
num_iter = 0
|
|
# Note: If create_tuple_iterator() is called with num_epochs>1,
|
|
# then EpochCtrlOp should be added to the pipeline
|
|
for _ in data2.create_dict_iterator(num_epochs=2):
|
|
num_iter += 1
|
|
assert num_iter == 4
|
|
|
|
# confirm_3ops_in_pipeline()
|
|
# MD BUG: Confirm pipeline file is not changed and wrongly still has 2 ops
|
|
confirm_2ops_in_pipeline()
|
|
# Confirm CPU util file has correct number of ops
|
|
confirm_cpuutil(3)
|
|
|
|
except Exception as error:
|
|
delete_profiling_files()
|
|
raise error
|
|
|
|
else:
|
|
delete_profiling_files()
|
|
|
|
|
|
if __name__ == "__main__":
|
|
test_profiling_simple_pipeline()
|
|
test_profiling_complex_pipeline()
|
|
test_profiling_inline_ops_pipeline1()
|
|
test_profiling_inline_ops_pipeline2()
|
|
test_profiling_sampling_interval()
|
|
test_profiling_basic_pipeline()
|
|
test_profiling_cifar10_pipeline()
|
|
test_profiling_epochctrl3()
|
|
test_profiling_epochctrl2()
|