-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
180 lines (136 loc) · 7.49 KB
/
train.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
#%%
from pathlib import Path
import os, sys, argparse
import yaml
from model.yolo3 import Yolo3
from utils.data_provider import DataGenerator
from utils.ade20k_util import Ade20kUtil
from utils.coco_util import CocoUtil
import numpy as np
import tensorflow as tf
from tqdm import tqdm
from model.model_utils import *
import time
from datetime import datetime
import cv2
#%%
if __name__ == "__main__" :
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=str, required=True)
args = parser.parse_args()
config = yaml.load("".join(Path(args.config).open("r").readlines()), Loader=yaml.FullLoader)
# config = yaml.load("".join(Path("configs/train/train_test.yaml").open("r").readlines()), Loader=yaml.FullLoader)
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = ",".join([str(i) for i in config["gpu_indices"]])
model = Yolo3(config=config)
# print(model.bbox_out)
if config["dataset"] == "coco" :
dg = DataGenerator(CocoUtil(Path(config["dataset_path"]), config=config, datatype="train"), config=config)
dg_v = DataGenerator(CocoUtil(Path(config["dataset_path"]), config=config, datatype="test"), config=config)
elif config["dataset"] == "ade20k" :
dg = DataGenerator(Ade20kUtil(Path(config["dataset_path"]), config=config), config=config)
tfconfig = tf.ConfigProto(allow_soft_placement=True)
tfconfig.gpu_options.allow_growth = True
model.sess = tf.Session(config=tfconfig)
saver = tf.train.Saver(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES), max_to_keep=20)
global_vars = tf.global_variables()
is_not_initialized = model.sess.run([tf.is_variable_initialized(var) for var in global_vars])
not_initialized_vars = [v for (v, f) in zip(global_vars, is_not_initialized) if not f]
if not_initialized_vars is not None:
model.sess.run(tf.variables_initializer(not_initialized_vars))
epoch = config["epoch"]
present_epoch = config["present_epoch"]
print(f"training start")
print(f"total epoch : {str(epoch)}")
if not Path(config["log_dir"]).exists() : Path(config["log_dir"]).mkdir(parents=True)
log_file = Path(config["log_dir"])/(datetime.now().strftime('%Y_%m_%d_%H_%M_%S')+".txt")
log_file.open("a+").write("epoch;iteration;total_loss;bbox_xy_loss;bbox_wh_loss;objness_loss;class_loss\n")
if present_epoch > 1 :
save_path = (Path(config["weight_store_path"])/(config["dataset"]+"_"+str(present_epoch)))
saver.restore(model.sess, str(save_path/"model.ckpt"))
print("model restore")
for e in range(present_epoch, epoch+1) :
lr = config["learning_rate"] * (config["lr_decaying_rate"]**(e/config["lr_decaying_interval"]))
loss_epoch_total = []
tqdm_iterator = tqdm(range(dg.len()))
for i in tqdm_iterator :
# print(i)
images, gts = dg.get_item(i)
if len(images) == 0 or len(gts) == 0 :
continue
image_inputs = np.array(images)
bbox_input = np.array([v[:, :4] for v in gts])
class_input = np.squeeze([v[:, 4:5] for v in gts])
if len(class_input) < 1 :
continue
vv = False
for ii in range(len(bbox_input)) :
vv = vv + np.any(bbox_input[ii][:, 2:4] - bbox_input[ii][:, :2] <= 0)
if vv : continue
b_bbox_inputs = []
m_bbox_inputs = []
s_bbox_inputs = []
b_class_inputs = []
m_class_inputs = []
s_class_inputs = []
b_objness_inputs = []
m_objness_inputs = []
s_objness_inputs = []
b_noobjness_inputs = []
m_noobjness_inputs = []
s_noobjness_inputs = []
for ii in range(len(gts)) :
bbox_input_tmp, class_input_tmp, objness_input_tmp, noobj_input_tmp = data_to_gt(model.priors_set, bbox_input[ii], class_input[ii], config["class_number"])
b_bbox_inputs.append(bbox_input_tmp[0])
m_bbox_inputs.append(bbox_input_tmp[1])
s_bbox_inputs.append(bbox_input_tmp[2])
b_class_inputs.append(class_input_tmp[0])
m_class_inputs.append(class_input_tmp[1])
s_class_inputs.append(class_input_tmp[2])
b_objness_inputs.append(objness_input_tmp[0])
m_objness_inputs.append(objness_input_tmp[1])
s_objness_inputs.append(objness_input_tmp[2])
b_noobjness_inputs.append(noobj_input_tmp[0])
m_noobjness_inputs.append(noobj_input_tmp[1])
s_noobjness_inputs.append(noobj_input_tmp[2])
# save single image and its objness scoremap
# save_files_from_final_output(image_inputs[ii], objness_input_tmp, noobj_input_tmp)
feed = {
model.image_inputs : image_inputs,
model.gt_placeholders[0]["bbox_computed_gt"] : b_bbox_inputs,
model.gt_placeholders[0]["objectness_gt"] : b_objness_inputs,
model.gt_placeholders[0]["noobjectness_gt"] : b_noobjness_inputs,
model.gt_placeholders[0]["class_computed_gt"] : b_class_inputs,
model.gt_placeholders[1]["bbox_computed_gt"] : m_bbox_inputs,
model.gt_placeholders[1]["objectness_gt"] : m_objness_inputs,
model.gt_placeholders[1]["noobjectness_gt"] : m_noobjness_inputs,
model.gt_placeholders[1]["class_computed_gt"] : m_class_inputs,
model.gt_placeholders[2]["bbox_computed_gt"] : s_bbox_inputs,
model.gt_placeholders[2]["objectness_gt"] : s_objness_inputs,
model.gt_placeholders[2]["noobjectness_gt"] : s_noobjness_inputs,
model.gt_placeholders[2]["class_computed_gt"] : s_class_inputs,
model.decaying_lr : lr,
model.is_training : True
}
tensors = [model.train_op, model.final_loss, model.losses["bbox_xy_loss"], model.losses["bbox_wh_loss"], model.losses["objness_loss"], model.losses["class_loss"]]
_, loss_tmp, bbox_xy_loss, bbox_wh_loss, objness_loss, class_loss = model.sess.run(tensors, feed_dict=feed)
loss_epoch_total.append(loss_tmp)
tqdm_iterator.set_description(f"Loss : {str(np.mean(loss_epoch_total))}, bbox_xy_loss : {str(bbox_xy_loss)}, bbox_wh_loss : {str(bbox_wh_loss)}, obj_loss : {str(objness_loss)}, class_loss : {str(class_loss)}")
log_file.open("a+").write(f"{str(e)};{str(i)};{str(loss_tmp)};{str(bbox_xy_loss)};{str(bbox_wh_loss)};{str(objness_loss)};{str(class_loss)}\n")
print(f"{str(e)} epoch is over!! avg loss {str(np.mean(loss_epoch_total))}")
log_file.open("a+").write(f"{str(e)} epoch is over!! avg loss {str(np.mean(loss_epoch_total))}\n")
loss_epoch_total = []
if e%config["validation_interval"] == 0 :
valid_count = 1
for i in range(valid_count) :
images_v, gts_v = dg_v.get_item(i)
validate_data(model, images_v, gts_v)
if e%config["saving_interval"] == 0 and e > 1 and e != config["present_epoch"] :
save_path = (Path(config["weight_store_path"])/(config["dataset"]+"_"+str(e)))
if not save_path.exists() : save_path.mkdir(parents=True)
print(save_path/"model.ckpt")
saver.save(model.sess, str(save_path/"model.ckpt"))
print("SAVE!")
pass
dg.on_epoch_end()
#%%