-
Notifications
You must be signed in to change notification settings - Fork 2
/
benchmark.py
195 lines (156 loc) · 8.37 KB
/
benchmark.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
from __future__ import absolute_import, division, print_function, unicode_literals
from utils import *
import tensorflow as tf
from models import resnet_cifar_model
import time
import numpy as np
'''tensorflow/benchmarks/blob/2696206fc01860d7b06fd02a01626f56abbab40a/scripts/tf_cnn_benchmarks/benchmark_cnn.py#L934'''
def get_perf_timing(batch_size, step_train_times, scale=1):
"""Calculate benchmark processing speed."""
times = np.array(step_train_times)
speeds = batch_size / times
time_mean = np.mean(times)
speed_mean = scale * batch_size / time_mean
speed_uncertainty = np.std(speeds) / np.sqrt(float(len(speeds)))
speed_jitter = 1.4826 * np.median(np.abs(speeds - np.median(speeds)))
return speed_mean, speed_uncertainty, speed_jitter
class Callbacks(tf.keras.callbacks.Callback):
def __init__(self, batch_size, display_every=20, num_gpu=1):
self.batch_size = batch_size
self.display_every = display_every
self.num_gpu = num_gpu
self.start_time = 0
self.train_time = 0
self.step_train_times = []
self.speeds = []
self._chief_worker_only = True
def on_train_batch_begin(self, batch, logs=None):
if (batch > 0):
self.start_time = time.time()
def on_train_batch_end(self, batch, logs=None):
if (batch > 0):
self.train_time = time.time() - self.start_time
self.step_train_times.append(self.train_time)
if ((batch % self.display_every) == 0):
speed_mean, speed_uncertainty, speed_jitter = get_perf_timing(self.batch_size, self.step_train_times, self.num_gpu)
self.speeds.append(speed_mean)
log_str = "{d0:d}\t{f1:0.1f}\t\t{f2:0.4f}\t\t{f3:0.4f}".format(d0=batch, f1=speed_mean, f2=logs['loss'], f3=logs['accuracy']*100)
print_msg(log_str, 'info')
def on_train_end(self, logs=None):
if (self.speeds):
speeds = np.array(self.speeds)
speed_mean = np.mean(speeds)
log_str = "total images/sec: {f0:0.2f}".format(f0=speed_mean)
print_msg(log_str, 'step')
class Benchmark(object):
def __init__(self, epochs, steps_per_epoch, batch_size=128, display_every=20, num_gpu=1, model='resnet56', strategy=None):
self.epochs = epochs
self.steps_per_epoch = steps_per_epoch
self.batch_size = batch_size
self.display_every = display_every
self.num_gpu = num_gpu
self.strategy = strategy
with self.strategy.scope():
self.model = self.create_model(model_name=model)
self.loss_fit = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
self.loss_loop = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE, from_logits=True)
self.optimizer = tf.keras.optimizers.SGD(learning_rate=0.1,momentum=0.9, nesterov=True)
checkpoint = tf.train.Checkpoint(optimizer=self.optimizer, model=self.model)
self.train_loss = tf.keras.metrics.Mean(name='train_loss')
self.test_loss = tf.keras.metrics.Mean(name='test_loss')
self.train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')
self.test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='test_accuracy')
self.callbacks = [Callbacks(self.batch_size, self.display_every, self.num_gpu), tf.keras.callbacks.LearningRateScheduler(self.decay)]
def create_model(self, model_name):
if model_name == 'resnet56':
self.model = resnet_cifar_model.resnet56(classes=10)
else:
err_msg = "Model name \"{}\" cannot be found!".format(model_name)
print_msg(err_msg, 'err')
sys.exit('Error!')
return self.model
def compile_model(self):
self.model.compile(optimizer=self.optimizer, loss=self.loss_fit, metrics=['accuracy'])
def decay(self, epoch):
if epoch < 3:
return 1e-3
elif epoch >= 3 and epoch < 7:
return 1e-4
else:
return 1e-5
def compute_loss(self, labels, predictions, batch_size=128):
per_example_loss = self.loss_loop(labels, predictions)
return tf.nn.compute_average_loss(per_example_loss, global_batch_size=batch_size)
def train_step(self, inputs):
image, label = inputs
with tf.GradientTape() as tape:
predictions = self.model(image, training=True)
loss = self.compute_loss(label, predictions)
gradients = tape.gradient(loss, self.model.trainable_variables)
self.optimizer.apply_gradients(zip(gradients, self.model.trainable_variables))
self.train_loss(loss)
self.train_accuracy(label, predictions)
return loss
@tf.function
def distributed_train_step(self, dataset_inputs):
per_replica_losses = self.strategy.experimental_run_v2(self.train_step, args=(dataset_inputs,))
return self.strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_losses, axis=None)
def test_step(inputs):
images, labels = inputs
predictions = self.model(images, training=False)
t_loss = self.loss_loop(labels, predictions)
test_loss.update_state(t_loss)
test_accuracy.update_state(labels, predictions)
def run(self, train_dataset, test_dataset, train_mode):
if train_mode == 'loop':
return self.loop_train(train_dataset, test_dataset)
elif train_mode == 'fit':
return self.fit_train(train_dataset.repeat(), test_dataset.repeat())
def loop_train(self, train_dataset, test_dataset):
#convert a tf.data.Dataset to "per-replica" values
if self.strategy._tf_api_names[0] != "distribute.OneDeviceStrategy":
train_dist_dataset = self.strategy.experimental_distribute_dataset(train_dataset)
test_dist_dataset = self.strategy.experimental_distribute_dataset(test_dataset)
else:
train_dist_dataset = train_dataset
test_dist_dataset = test_dataset
with self.strategy.scope():
print_msg("Warming Up...", 'info')
for x in train_dataset.take(self.num_gpu):
self.distributed_train_step(x)
header_str = "{s0:s}\t{s1:s}\t\t{s2:s}\t{s3:s}".format(s0='Step', s1='Img/sec', s2='total_loss', s3='accuracy(%)')
print_msg(header_str, 'step')
batch=0
step_train_times = []
speeds = []
for epoch in range(self.epochs):
for inputs in train_dist_dataset:
# store start time
start_time = time.time()
self.distributed_train_step(inputs)
template = ('Epoch: {}, Train Loss: {}, Train Accuracy: {}, '
'Test Loss: {}, Test Accuracy: {}')
batch += 1
if ((batch % self.display_every) == 0):
#Measuring elapsed time
train_time = time.time() - start_time
step_train_times.append(train_time)
speed_mean, speed_uncertainty, speed_jitter = get_perf_timing(self.batch_size, step_train_times, self.num_gpu)
speeds.append(speed_mean)
log_str = "{d0:d}\t{f1:0.1f}\t\t{f2:0.4f}\t\t{f3:0.4f}".format(d0=batch, f1=speed_mean, f2=self.train_loss.result(), f3=self.train_accuracy.result()*100)
print_msg(log_str, 'info')
speeds = np.array(speeds)
speed_mean = np.mean(speeds)
log_str = "total images/sec: {f0:0.2f}".format(f0=speed_mean)
print_msg(log_str, 'step')
def fit_train(self, train_dataset, test_dataset):
# Compile the model in the context of strategy.scope
with self.strategy.scope():
self.compile_model()
print_msg("Warming Up...", 'info')
self.model.fit(train_dataset.take(self.num_gpu), epochs=1, steps_per_epoch=1, verbose=0)
header_str = "{s0:s}\t{s1:s}\t\t{s2:s}\t{s3:s}".format(s0='Step', s1='Img/sec', s2='total_loss', s3='accuracy(%)')
print_msg(header_str, 'step')
history = self.model.fit(train_dataset, epochs=self.epochs, steps_per_epoch=self.steps_per_epoch, verbose=0, callbacks=self.callbacks)
return (history.history['loss'][-1],
history.history['accuracy'][-1])