-
Notifications
You must be signed in to change notification settings - Fork 341
/
eval.py
194 lines (151 loc) · 7.83 KB
/
eval.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
"""
At each tick, evaluate the latest checkpoint against some validation data.
Or, you can run once by passing --run_once. OR, you can pass a --requested_step_seq of comma separated checkpoint #s that already exist that it can run in a row.
This program expects a training base directory with the data, and md.json file
There will be sub-directories for each run underneath with the name run-<PID>
where <PID> is the training program's process ID. To run this program, you
will need to pass --train_dir <DIR> which is the base path name, --run_id <PID>
and if you are using a custom name for your checkpoint, you should
pass that as well (most times you probably wont). This will yield a model path:
<DIR>/run-<PID>/checkpoint
Note: If you are training to use the same GPU you can supposedly
suspend the process. I have not found this works reliably on my Linux machine.
Instead, I have found that, often times, the GPU will not reclaim the resources
and in that case, your eval may run out of GPU memory.
You can alternately run trainining for a number of steps, break the program
and run this, then restarting training from the old checkpoint. I also
found this inconvenient. In order to control this better, the program
requires that you explict placement of inference. It defaults to the CPU
so that it can easily run side by side with training. This does make it
much slower than if it was on the GPU, but for evaluation this may not be
a major problem. To place on the gpu, just pass --device_id /gpu:<ID> where
<ID> is the GPU ID
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from datetime import datetime
import math
import time
from data import inputs
import numpy as np
import tensorflow as tf
from model import select_model, get_checkpoint
import os
import json
tf.app.flags.DEFINE_string('train_dir', '/home/dpressel/dev/work/AgeGenderDeepLearning/Folds/tf/test_fold_is_0',
'Training directory (where training data lives)')
tf.app.flags.DEFINE_integer('run_id', 0,
'This is the run number (pid) for training proc')
tf.app.flags.DEFINE_string('device_id', '/cpu:0',
'What processing unit to execute inference on')
tf.app.flags.DEFINE_string('eval_dir', '/home/dpressel/dev/work/AgeGenderDeepLearning/Folds/tf/eval_test_fold_is_0',
'Directory to put output to')
tf.app.flags.DEFINE_string('eval_data', 'valid',
'Data type (valid|train)')
tf.app.flags.DEFINE_integer('num_preprocess_threads', 4,
'Number of preprocessing threads')
tf.app.flags.DEFINE_integer('eval_interval_secs', 60 * 5,
"""How often to run the eval.""")
tf.app.flags.DEFINE_integer('num_examples', 10000,
"""Number of examples to run.""")
tf.app.flags.DEFINE_boolean('run_once', False,
"""Whether to run eval only once.""")
tf.app.flags.DEFINE_integer('image_size', 227,
'Image size')
tf.app.flags.DEFINE_integer('batch_size', 128,
'Batch size')
tf.app.flags.DEFINE_string('checkpoint', 'checkpoint',
'Checkpoint basename')
tf.app.flags.DEFINE_string('model_type', 'default',
'Type of convnet')
tf.app.flags.DEFINE_string('requested_step_seq', '', 'Requested step to restore')
FLAGS = tf.app.flags.FLAGS
def eval_once(saver, summary_writer, summary_op, logits, labels, num_eval, requested_step=None):
"""Run Eval once.
Args:
saver: Saver.
summary_writer: Summary writer.
top_k_op: Top K op.
summary_op: Summary op.
"""
top1 = tf.nn.in_top_k(logits, labels, 1)
top2 = tf.nn.in_top_k(logits, labels, 2)
with tf.Session() as sess:
checkpoint_path = '%s/run-%d' % (FLAGS.train_dir, FLAGS.run_id)
model_checkpoint_path, global_step = get_checkpoint(checkpoint_path, requested_step, FLAGS.checkpoint)
saver.restore(sess, model_checkpoint_path)
# Start the queue runners.
coord = tf.train.Coordinator()
try:
threads = []
for qr in tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS):
threads.extend(qr.create_threads(sess, coord=coord, daemon=True,
start=True))
num_steps = int(math.ceil(num_eval / FLAGS.batch_size))
true_count1 = true_count2 = 0
total_sample_count = num_steps * FLAGS.batch_size
step = 0
print(FLAGS.batch_size, num_steps)
while step < num_steps and not coord.should_stop():
start_time = time.time()
v, predictions1, predictions2 = sess.run([logits, top1, top2])
duration = time.time() - start_time
sec_per_batch = float(duration)
examples_per_sec = FLAGS.batch_size / sec_per_batch
true_count1 += np.sum(predictions1)
true_count2 += np.sum(predictions2)
format_str = ('%s (%.1f examples/sec; %.3f sec/batch)')
print(format_str % (datetime.now(),
examples_per_sec, sec_per_batch))
step += 1
# Compute precision @ 1.
at1 = true_count1 / total_sample_count
at2 = true_count2 / total_sample_count
print('%s: precision @ 1 = %.3f (%d/%d)' % (datetime.now(), at1, true_count1, total_sample_count))
print('%s: recall @ 2 = %.3f (%d/%d)' % (datetime.now(), at2, true_count2, total_sample_count))
summary = tf.Summary()
summary.ParseFromString(sess.run(summary_op))
summary.value.add(tag='Precision @ 1', simple_value=at1)
summary.value.add(tag=' Recall @ 2', simple_value=at2)
summary_writer.add_summary(summary, global_step)
except Exception as e: # pylint: disable=broad-except
coord.request_stop(e)
coord.request_stop()
coord.join(threads, stop_grace_period_secs=10)
def evaluate(run_dir):
with tf.Graph().as_default() as g:
input_file = os.path.join(FLAGS.train_dir, 'md.json')
print(input_file)
with open(input_file, 'r') as f:
md = json.load(f)
eval_data = FLAGS.eval_data == 'valid'
num_eval = md['%s_counts' % FLAGS.eval_data]
model_fn = select_model(FLAGS.model_type)
with tf.device(FLAGS.device_id):
print('Executing on %s' % FLAGS.device_id)
images, labels, _ = inputs(FLAGS.train_dir, FLAGS.batch_size, FLAGS.image_size, train=not eval_data, num_preprocess_threads=FLAGS.num_preprocess_threads)
logits = model_fn(md['nlabels'], images, 1, False)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(run_dir, g)
saver = tf.train.Saver()
if FLAGS.requested_step_seq:
sequence = FLAGS.requested_step_seq.split(',')
for requested_step in sequence:
print('Running %s' % sequence)
eval_once(saver, summary_writer, summary_op, logits, labels, num_eval, requested_step)
else:
while True:
print('Running loop')
eval_once(saver, summary_writer, summary_op, logits, labels, num_eval)
if FLAGS.run_once:
break
time.sleep(FLAGS.eval_interval_secs)
def main(argv=None): # pylint: disable=unused-argument
run_dir = '%s/run-%d' % (FLAGS.eval_dir, FLAGS.run_id)
if tf.gfile.Exists(run_dir):
tf.gfile.DeleteRecursively(run_dir)
tf.gfile.MakeDirs(run_dir)
evaluate(run_dir)
if __name__ == '__main__':
tf.app.run()