def train():
total_timer = Timer()
train_timer = Timer()
load_timer = Timer()
max_epoch = 30
epoch_step = int(cfg.train_num//cfg.batch_size)
t = 1
for epoch in range(1, max_epoch + 1):
print('-'*25, 'epoch', epoch,'/',str(max_epoch), '-'*25)
t_loss = 0
ll_loss = 0
r_loss = 0
c_loss = 0
for step in range(1, epoch_step + 1):
t = t + 1
total_timer.tic()
load_timer.tic()
images, labels, imnm, num_boxes, imsize = data.get()
# load_timer.toc()
feed_dict = {input_: images,
get_boxes: labels[..., 1:5][:, ::-1, :],
get_classes: labels[..., 0].reshape((cfg.batch_size, -1))[:, ::-1]
}
if cfg.cnt_branch:
_, g_step_, tt_loss, cl_loss, cn_loss, re_loss, lr_ = sess.run(
[train_op,
global_step,
total_loss,
cls_loss,
cnt_loss,
reg_loss,
lr], feed_dict = feed_dict)
else:
_, g_step_, tt_loss, cl_loss, re_loss, lr_ = sess.run(
[train_op,
global_step,
total_loss,
cls_loss,
reg_loss,
lr], feed_dict = feed_dict)
total_timer.toc()
if g_step_%50 ==0:
sys.stdout.write('\r>> ' + 'iters '+str(g_step_)+str('/')+str(epoch_step*max_epoch)+' loss '+str(tt_loss) + ' ')
sys.stdout.flush()
summary_str = sess.run(summary_op, feed_dict = feed_dict)
train_total_summary = tf.Summary(value=[
tf.Summary.Value(tag="config/learning rate", simple_value=lr_),
tf.Summary.Value(tag="train/classification/focal_loss", simple_value=cfg.class_weight*cl_loss),
tf.Summary.Value(tag="train/classification/cnt_loss", simple_value=cfg.cnt_weight*cn_loss),
# tf.Summary.Value(tag="train/p_nm", simple_value=p_nm_),
tf.Summary.Value(tag="train/regress_loss", simple_value=cfg.regress_weight*re_loss),
# tf.Summary.Value(tag="train/clone_loss", simple_value=cfg.class_weight*cl_loss + cfg.regress_weight*re_loss + cfg.cnt_weight*cn_loss),
# tf.Summary.Value(tag="train/l2_loss", simple_value=l2_loss),
tf.Summary.Value(tag="train/total_loss", simple_value=tt_loss)
])
print('curent speed: ', total_timer.diff, 'remain time: ', total_timer.remain(g_step_, epoch_step*max_epoch))
summary_writer.add_summary(summary_str, g_step_)
summary_writer.add_summary(train_total_summary, g_step_)
if g_step_%10000 == 0:
print('saving checkpoint')
saver.save(sess, cfg.ckecpoint_file + '/model.ckpt', g_step_)
total_timer.toc()
sys.stdout.write('\n')
print('>> mean loss', t_loss)
print('curent speed: ', total_timer.average_time, 'remain time: ', total_timer.remain(g_step_, epoch_step*max_epoch))
print('saving checkpoint')
saver.save(sess, cfg.ckecpoint_file + '/model.ckpt', g_step_)
restorer = tf.train.Saver(g_list)
restorer.restore(sess, restore_path)
if __name__ == '__main__':
val_timer = Timer()
data = pascal_voc('test', False, cfg.test_img_path, cfg.test_label_path,
cfg.test_img_txt, False)
val_pred = []
# saved_pred = {}
gt_dict = {}
val_rloss = 0
val_closs = 0
val_cnt_loss = 0
for val_step in range(1, cfg.test_num + 1):
val_timer.tic()
images, labels, imnm, num_boxes, imsize = data.get()
feed_dict = {
input_: images,
get_boxes: labels[..., 1:5][:, ::-1, :],
get_classes: labels[..., 0].reshape((cfg.batch_size, -1))[:, ::-1]
}
if cfg.cnt_branch:
b, s, l, valcloss_, valcntloss_, valrloss_ = sess.run(
[nms_box, nms_score, nms_label, cls_loss, cnt_loss, reg_loss],
feed_dict=feed_dict)
else:
b, s, l, valcloss_, valrloss_ = sess.run(
[nms_box, nms_score, nms_label, cls_loss, reg_loss],
feed_dict=feed_dict)
val_rloss += valrloss_ / cfg.test_num
