Skip to content
This repository was archived by the owner on May 11, 2024. It is now read-only.

Commit 0479fab

Browse files
chuanqi129chuanqi129
committed
Merge branch 'r1.0_rc' into 'master'
R1.0 rc MR See merge request intelai/tools!36
2 parents b901a54 + fccbe9d commit 0479fab

26 files changed

+3361
-2233
lines changed

‎api/examples/__init__.py

+17
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,17 @@
1+
#
2+
# -*- coding: utf-8 -*-
3+
#
4+
# Copyright (c) 2019 Intel Corporation
5+
#
6+
# Licensed under the Apache License, Version 2.0 (the "License");
7+
# you may not use this file except in compliance with the License.
8+
# You may obtain a copy of the License at
9+
#
10+
# http://www.apache.org/licenses/LICENSE-2.0
11+
#
12+
# Unless required by applicable law or agreed to in writing, software
13+
# distributed under the License is distributed on an "AS IS" BASIS,
14+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15+
# See the License for the specific language governing permissions and
16+
# limitations under the License.
17+
#

‎api/examples/include/__init__.py

+17
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,17 @@
1+
#
2+
# -*- coding: utf-8 -*-
3+
#
4+
# Copyright (c) 2019 Intel Corporation
5+
#
6+
# Licensed under the Apache License, Version 2.0 (the "License");
7+
# you may not use this file except in compliance with the License.
8+
# You may obtain a copy of the License at
9+
#
10+
# http://www.apache.org/licenses/LICENSE-2.0
11+
#
12+
# Unless required by applicable law or agreed to in writing, software
13+
# distributed under the License is distributed on an "AS IS" BASIS,
14+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15+
# See the License for the specific language governing permissions and
16+
# limitations under the License.
17+
#

‎api/examples/include/datasets2.py

+107
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,107 @@
1+
#!/usr/bin/env python
2+
# -*- coding: utf-8 -*-
3+
#
4+
# Copyright (c) 2018 Intel Corporation
5+
#
6+
# Licensed under the Apache License, Version 2.0 (the "License");
7+
# you may not use this file except in compliance with the License.
8+
# You may obtain a copy of the License at
9+
#
10+
# http://www.apache.org/licenses/LICENSE-2.0
11+
#
12+
# Unless required by applicable law or agreed to in writing, software
13+
# distributed under the License is distributed on an "AS IS" BASIS,
14+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15+
# See the License for the specific language governing permissions and
16+
# limitations under the License.
17+
#
18+
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
19+
#
20+
# Licensed under the Apache License, Version 2.0 (the "License");
21+
# you may not use this file except in compliance with the License.
22+
# You may obtain a copy of the License at
23+
#
24+
# http://www.apache.org/licenses/LICENSE-2.0
25+
#
26+
# Unless required by applicable law or agreed to in writing, software
27+
# distributed under the License is distributed on an "AS IS" BASIS,
28+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
29+
# See the License for the specific language governing permissions and
30+
# limitations under the License.
31+
# ==============================================================================
32+
33+
"""Benchmark dataset utilities.
34+
"""
35+
36+
import os
37+
from abc import abstractmethod
38+
39+
import tensorflow as tf
40+
41+
import preprocessing2
42+
43+
IMAGENET_NUM_TRAIN_IMAGES = 1281167
44+
IMAGENET_NUM_VAL_IMAGES = 50000
45+
46+
class Dataset2(object):
47+
"""Abstract class for cnn benchmarks dataset."""
48+
49+
def __init__(self, name, height=None, width=None, depth=None, data_dir=None,
50+
queue_runner_required=False, num_classes=1000):
51+
self.name = name
52+
self.height = height
53+
self.width = width
54+
self.depth = depth or 3
55+
56+
self.data_dir = data_dir
57+
self._queue_runner_required = queue_runner_required
58+
self._num_classes = num_classes
59+
60+
def tf_record_pattern(self, subset):
61+
return os.path.join(self.data_dir, '%s-*-of-*' % subset)
62+
63+
def reader(self):
64+
return tf.compat.v1.TFRecordReader()
65+
66+
@property
67+
def num_classes(self):
68+
return self._num_classes
69+
70+
@num_classes.setter
71+
def num_classes(self, val):
72+
self._num_classes = val
73+
74+
@abstractmethod
75+
def num_examples_per_epoch(self, subset):
76+
pass
77+
78+
def __str__(self):
79+
return self.name
80+
81+
def get_image_preprocessor(self):
82+
return None
83+
84+
def queue_runner_required(self):
85+
return self._queue_runner_required
86+
87+
def use_synthetic_gpu_images(self):
88+
return not self.data_dir
89+
90+
91+
class ImagenetData(Dataset2):
92+
"""Configuration for Imagenet dataset."""
93+
94+
def __init__(self, data_dir=None):
95+
super(ImagenetData, self).__init__('imagenet', 300, 300, data_dir=data_dir)
96+
97+
def num_examples_per_epoch(self, subset='train'):
98+
if subset == 'train':
99+
return IMAGENET_NUM_TRAIN_IMAGES
100+
elif subset == 'validation':
101+
return IMAGENET_NUM_VAL_IMAGES
102+
else:
103+
raise ValueError('Invalid data subset "%s"' % subset)
104+
105+
def get_image_preprocessor(self):
106+
return preprocessing2.RecordInputImagePreprocessor
107+

‎api/examples/include/eval_image_classifier_optimize.py

+247
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,247 @@
1+
#
2+
# -*- coding: utf-8 -*-
3+
#
4+
# Copyright (c) 2018 Intel Corporation
5+
#
6+
# Licensed under the Apache License, Version 2.0 (the "License");
7+
# you may not use this file except in compliance with the License.
8+
# You may obtain a copy of the License at
9+
#
10+
# http://www.apache.org/licenses/LICENSE-2.0
11+
#
12+
# Unless required by applicable law or agreed to in writing, software
13+
# distributed under the License is distributed on an "AS IS" BASIS,
14+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15+
# See the License for the specific language governing permissions and
16+
# limitations under the License.
17+
#
18+
# SPDX-License-Identifier: EPL-2.0
19+
#
20+
21+
import time
22+
from argparse import ArgumentParser
23+
24+
import tensorflow as tf
25+
from tensorflow.python.tools.optimize_for_inference_lib import optimize_for_inference
26+
from tensorflow.python.framework import dtypes
27+
28+
import datasets2 as datasets
29+
30+
31+
# override by args
32+
INPUTS = "input"
33+
OUTPUTS = "predict"
34+
35+
INCEPTION_V3_IMAGE_SIZE = 224
36+
37+
38+
class eval_classifier_optimized_graph:
39+
"""Evaluate image classifier with optimized TensorFlow graph"""
40+
41+
def __init__(self):
42+
43+
arg_parser = ArgumentParser(description='Parse args')
44+
45+
arg_parser.add_argument('-b', "--batch-size",
46+
help="Specify the batch size. If this " \
47+
"parameter is not specified or is -1, the " \
48+
"largest ideal batch size for the model will " \
49+
"be used.",
50+
dest="batch_size", type=int, default=-1)
51+
52+
arg_parser.add_argument('-e', "--num-inter-threads",
53+
help='The number of inter-thread.',
54+
dest='num_inter_threads', type=int, default=0)
55+
56+
arg_parser.add_argument('-a', "--num-intra-threads",
57+
help='The number of intra-thread.',
58+
dest='num_intra_threads', type=int, default=0)
59+
60+
arg_parser.add_argument('-g', "--input-graph",
61+
help='Specify the input graph for the transform tool',
62+
dest='input_graph')
63+
64+
arg_parser.add_argument('-i', "--input",
65+
help='Specify the input of the model',
66+
dest='input')
67+
arg_parser.add_argument('-o', "--output",
68+
help='Specify the output of the model',
69+
dest='output')
70+
arg_parser.add_argument('--image_size', dest='image_size',
71+
help='image size',
72+
type=int, default=224)
73+
74+
arg_parser.add_argument('-d', "--data-location",
75+
help='Specify the location of the data. '
76+
'If this parameter is not specified, '
77+
'the benchmark will use random/dummy data.',
78+
dest="data_location", default=None)
79+
80+
arg_parser.add_argument('-r', "--accuracy-only",
81+
help='For accuracy measurement only.',
82+
dest='accuracy_only', action='store_true')
83+
84+
arg_parser.add_argument("--warmup-steps", type=int, default=10,
85+
help="number of warmup steps")
86+
arg_parser.add_argument("--steps", type=int, default=50,
87+
help="number of steps")
88+
89+
arg_parser.add_argument(
90+
'--data-num-inter-threads', dest='data_num_inter_threads',
91+
help='number threads across operators',
92+
type=int, default=16)
93+
arg_parser.add_argument(
94+
'--data-num-intra-threads', dest='data_num_intra_threads',
95+
help='number threads for data layer operator',
96+
type=int, default=14)
97+
arg_parser.add_argument(
98+
'--num-cores', dest='num_cores',
99+
help='number of cores',
100+
type=int, default=28)
101+
arg_parser.add_argument(
102+
'--env', dest='env', help='specific Tensorflow env',
103+
default='mkl'
104+
)
105+
106+
self.args = arg_parser.parse_args()
107+
108+
# validate the arguments specific for InceptionV3
109+
self.validate_args()
110+
111+
def run(self):
112+
"""run benchmark with optimized graph"""
113+
114+
print("Run inference")
115+
116+
data_config = tf.compat.v1.ConfigProto()
117+
data_config.intra_op_parallelism_threads = self.args.data_num_intra_threads
118+
data_config.inter_op_parallelism_threads = self.args.data_num_inter_threads
119+
data_config.use_per_session_threads = 1
120+
121+
infer_config = tf.compat.v1.ConfigProto()
122+
if self.args.env == 'mkl':
123+
print("Set inter and intra for mkl")
124+
infer_config.intra_op_parallelism_threads = self.args.num_intra_threads
125+
infer_config.inter_op_parallelism_threads = self.args.num_inter_threads
126+
infer_config.use_per_session_threads = 1
127+
128+
data_graph = tf.Graph()
129+
with data_graph.as_default():
130+
if (self.args.data_location):
131+
print("Inference with real data.")
132+
dataset = datasets.ImagenetData(self.args.data_location)
133+
preprocessor = dataset.get_image_preprocessor()(
134+
self.args.image_size, self.args.image_size, self.args.batch_size,
135+
num_cores=self.args.num_cores,
136+
resize_method='bilinear')
137+
images, labels = preprocessor.minibatch(dataset, subset='validation')
138+
else:
139+
print("Inference with dummy data.")
140+
input_shape = [self.args.batch_size, self.args.image_size, self.args.image_size, 3]
141+
images = tf.random.uniform(input_shape, 0.0, 255.0, dtype=tf.float32, name='synthetic_images')
142+
143+
infer_graph = tf.Graph()
144+
with infer_graph.as_default():
145+
graph_def = tf.compat.v1.GraphDef()
146+
with tf.compat.v1.gfile.FastGFile(self.args.input_graph, 'rb') as input_file:
147+
input_graph_content = input_file.read()
148+
graph_def.ParseFromString(input_graph_content)
149+
150+
output_graph = optimize_for_inference(graph_def, [self.args.input],
151+
[self.args.output], dtypes.float32.as_datatype_enum, False)
152+
tf.import_graph_def(output_graph, name='')
153+
154+
# Definite input and output Tensors for detection_graph
155+
input_tensor = infer_graph.get_tensor_by_name(self.args.input + ':0')
156+
output_tensor = infer_graph.get_tensor_by_name(self.args.output + ':0')
157+
158+
data_sess = tf.compat.v1.Session(graph=data_graph, config=data_config)
159+
infer_sess = tf.compat.v1.Session(graph=infer_graph, config=infer_config)
160+
161+
num_processed_images = 0
162+
#num_remaining_images = datasets.IMAGENET_NUM_VAL_IMAGES
163+
num_remaining_images = 50000
164+
165+
if (not self.args.accuracy_only):
166+
iteration = 0
167+
warm_up_iteration = self.args.warmup_steps
168+
total_run = self.args.steps
169+
total_time = 0
170+
171+
while num_remaining_images >= self.args.batch_size and iteration < total_run:
172+
iteration += 1
173+
174+
data_load_start = time.time()
175+
image_np = data_sess.run(images)
176+
data_load_time = time.time() - data_load_start
177+
178+
num_processed_images += self.args.batch_size
179+
num_remaining_images -= self.args.batch_size
180+
181+
start_time = time.time()
182+
infer_sess.run([output_tensor], feed_dict={input_tensor: image_np})
183+
time_consume = time.time() - start_time
184+
185+
# only add data loading time for real data, not for dummy data
186+
if self.args.data_location:
187+
time_consume += data_load_time
188+
189+
print('Iteration %d: %.6f sec' % (iteration, time_consume))
190+
if iteration > warm_up_iteration:
191+
total_time += time_consume
192+
193+
time_average = total_time / (iteration - warm_up_iteration)
194+
print('Average time: %.6f sec' % (time_average))
195+
196+
print('Batch size = %d' % self.args.batch_size)
197+
if (self.args.batch_size == 1):
198+
print('Latency: %.3f ms' % (time_average * 1000))
199+
200+
print('Throughput: %.3f images/sec' % (self.args.batch_size / time_average))
201+
202+
else: # accuracy check
203+
total_accuracy1, total_accuracy5 = (0.0, 0.0)
204+
205+
while num_remaining_images >= self.args.batch_size:
206+
# Reads and preprocess data
207+
np_images, np_labels = data_sess.run([images, labels])
208+
num_processed_images += self.args.batch_size
209+
num_remaining_images -= self.args.batch_size
210+
211+
start_time = time.time()
212+
# Compute inference on the preprocessed data
213+
predictions = infer_sess.run(output_tensor,
214+
{input_tensor: np_images})
215+
elapsed_time = time.time() - start_time
216+
217+
with tf.Graph().as_default() as accu_graph:
218+
accuracy1 = tf.reduce_sum(
219+
input_tensor=tf.cast(tf.nn.in_top_k(predictions=tf.constant(predictions),
220+
targets=tf.constant(np_labels), k=1), tf.float32))
221+
222+
accuracy5 = tf.reduce_sum(
223+
input_tensor=tf.cast(tf.nn.in_top_k(predictions=tf.constant(predictions),
224+
targets=tf.constant(np_labels), k=5), tf.float32))
225+
with tf.compat.v1.Session() as accu_sess:
226+
np_accuracy1, np_accuracy5 = accu_sess.run([accuracy1, accuracy5])
227+
228+
total_accuracy1 += np_accuracy1
229+
total_accuracy5 += np_accuracy5
230+
231+
print("Iteration time: %0.4f ms" % elapsed_time)
232+
print("Processed %d images. (Top1 accuracy, Top5 accuracy) = (%0.4f, %0.4f)" \
233+
% (num_processed_images, total_accuracy1 / num_processed_images,
234+
total_accuracy5 / num_processed_images))
235+
236+
def validate_args(self):
237+
"""validate the arguments"""
238+
239+
if not self.args.data_location:
240+
if self.args.accuracy_only:
241+
raise ValueError("You must use real data for accuracy measurement.")
242+
243+
244+
if __name__ == "__main__":
245+
246+
evaluate_opt_graph = eval_classifier_optimized_graph()
247+
evaluate_opt_graph.run()

0 commit comments

Comments
 (0)