-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainloop.py
147 lines (112 loc) · 5.98 KB
/
trainloop.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
import logging
import os
import tensorflow as tf
import numpy as np
import networks.cls_lstm
import util.global_config as global_config
import util.helper
import util.network_io_utils
import valloop
# Train the lstm with input data given as feed_dict.
def train(sess, tensors, input_pipeline_out, frcnn_out, target_cls, ordered_last_region_proposals,
cls_weight=0.5, reg_weight=0.5):
input_tensors = {
'train_step': tensors['combined']['train_step'],
'total_loss': tensors['combined']['loss'],
'cls_inputs': tensors['cls']['inputs'],
'cls_targets': tensors['cls']['targets'],
# This is a list.
'cls_predictions': tensors['cls']['predictions'],
'reg_targets': tensors['lstm']['targets'],
'reg_predictions': tensors['lstm']['predictions'],
'summary': tensors['summary']
}
with util.helper.timeit() as train_time:
out = sess.run(input_tensors, feed_dict={
tensors['placeholders']['groundtruth_bbs']: input_pipeline_out['groundtruth_bbs'],
tensors['placeholders']['target_bbs']: input_pipeline_out['target_bbs'],
tensors['placeholders']['images']: input_pipeline_out['images'],
tensors['placeholders']['region_proposals']: frcnn_out,
tensors['placeholders']['target_cls']: target_cls,
tensors['placeholders']['ordered_last_region_proposals']: ordered_last_region_proposals,
tensors['combined']['cls_weight']: cls_weight,
tensors['combined']['reg_weight']: reg_weight
})
return out, train_time.time()
def interval_actions(epoch, step, globalstep,
input_time, frcnn_time, train_time,
loss, train_writer, summary, saver, sess):
# sequence_images, out_input_pipeline, out_frcnn, out_lstm):
if globalstep % 1 == 0:
logging.info(
'Epoch %d, step %d, global step %d (%.3f/%.3f/%.3f sec input/frcnn_predict/train). Loss %.3f.' % (
epoch, step, globalstep, input_time, frcnn_time, train_time, loss))
if globalstep % global_config.cfg['summary_interval'] == 0:
train_writer.add_summary(summary, globalstep)
if globalstep % global_config.cfg['result_interval'] == 0:
pass
if globalstep % global_config.cfg['save_interval'] == 0:
saver.save(sess, os.path.join(global_config.cfg['checkpoints'], 'checkpoint'),
global_step=globalstep)
def run(sess, input_pipeline_tensors, input_handles, network_tensors,
train_writer, epoch, saver, globalstep, frcnn,
cls_weight, reg_weight, frcnn_saved, validate=True):
sess.run(input_handles['training_initializer'])
step = 0
# Go through one epoch.
while True:
try:
input_pipeline_out, input_time = util.network_io_utils.read_input(sess, input_pipeline_tensors,
input_handles, 'training')
sequence_id = ''.join([i.decode() for i in input_pipeline_out['sequence_id'][0]])
if sequence_id in frcnn_saved:
frcnn_out = frcnn_saved[sequence_id]
frcnn_time = 0
else:
frcnn_out, frcnn_time = util.network_io_utils.predict_frcnn(input_pipeline_out['images'], frcnn)
frcnn_saved[sequence_id] = frcnn_out
util.helper.savedict(frcnn_saved, global_config.cfg['frcnn_saved_file'])
# ============================================================================================
# DO THE STUFF HERE.
# print(input_pipeline_out['groundtruth_bbs'].shape)
# Shape (batch_size, sequence_size, 10).
target_cls = networks.cls_lstm.generate_cls_gt(input_pipeline_out, frcnn_out)
# TODO: Does this even make sense???
# Generate ordered_last_region_proposals with shape (batch_size, sequence_length, 10, 4).
batch_size = global_config.cfg['batch_size']
sequence_length = global_config.cfg['backprop_step_size']
ordered_last_region_proposals = np.zeros((batch_size, sequence_length, 10, 4))
for t in range(sequence_length - 1):
for j in range(input_pipeline_out['groundtruth_bbs'].shape[0]):
gtbb = np.reshape(input_pipeline_out['groundtruth_bbs'][j, t], (10, 4))
if t == 0:
ordered_last_region_proposals[j, t + 1] = frcnn_out[j, t]
else:
for i in range(10):
# i = player id.
predid = int(target_cls[j, t, i])
if predid == 0:
# Take lstm.
ordered_last_region_proposals[j, t + 1, i] = gtbb[i]
else:
# Take rp.
ordered_last_region_proposals[j, t + 1, i] = frcnn_out[j, t, predid - 1]
out, train_time = train(sess, network_tensors, input_pipeline_out, frcnn_out, target_cls,
ordered_last_region_proposals,
cls_weight, reg_weight)
interval_actions(epoch, step, globalstep,
input_time, frcnn_time, train_time,
out['total_loss'], train_writer, out['summary'],
saver, sess)
if validate:
# TODO
target_cls_init = target_cls[:, 0, :]
if globalstep % global_config.cfg['validation_interval'] == 0:
logging.info('**Start validation.**')
valloop.run(sess, input_pipeline_tensors, input_handles, network_tensors, frcnn, target_cls_init)
logging.info('**Finished validation.**')
step += 1
globalstep += 1
except tf.errors.OutOfRangeError:
break
return globalstep