def __init__(self, model_path,name,version,export_path='./model'):
'''
:param model_path: ckpt模型保存地址
:param name: 转换后模型标识
:param version: 版本
:param export_path: 转换后存放地址
'''
self.classes = data_pre.get_classes(cfg.Main.classes) # 类别
self.anchors = data_pre.get_anchors(cfg.Main.anchors) # anchors
self.model_path = tf.train.get_checkpoint_state(model_path).model_checkpoint_path
self.output_size = 416
self.batch_size = 1
self.export_path = export_path
self.name = name
self.version = version
self.ses = tf.Session()
self.is_training = cfg.Test.is_training
np.random.seed(21)
with tf.name_scope('inputs'):
self.input_images = tf.placeholder(dtype=tf.float32,
shape=[self.batch_size, self.output_size, self.output_size, 3],
name='input_images')
with tf.name_scope('model_create'):
self.model = Yolo3(input_value=self.input_images, is_training=self.is_training)
with tf.name_scope('saver_log'):
self.saver = tf.train.Saver(tf.global_variables())
def __init__(self,images_path,model_path,output_size,is_save=False):
self.classes = data_pre.get_classes(cfg.Main.classes) # 类别
self.anchors = data_pre.get_anchors(cfg.Main.anchors) # anchors
self.images_path = images_path
self.batch_size = len(self.images_path)
self.model_path = tf.train.get_checkpoint_state(model_path).model_checkpoint_path
# self.model_path = './checkpoint/darknet/yolov3.ckpt'
self.output_size = output_size
self.ses = tf.Session()
self.is_training = cfg.Test.is_training
self.is_save = is_save
self.nms_score = cfg.Test.nms_score
self.bboxes_score = cfg.Test.bboxes_score
np.random.seed(21)
with tf.name_scope('inputs'):
self.input_images = tf.placeholder(dtype=tf.float32,shape=[self.batch_size,self.output_size,self.output_size,3], name='input_images')
with tf.name_scope('model_create'):
self.model = Yolo3(input_value=self.input_images,is_training=self.is_training)
with tf.name_scope('saver_log'):
# global_variables = tf.global_variables()
# ckpt_variables = self.get_all_variables_name_from_ckpt(self.model_path)
# print(ckpt_variables)
# for ids,value in enumerate(global_variables):
# print('global:',value)
# print('ckpt:',ckpt_variables[ids])
# print('\n')
#
# os._exit(0)
self.saver = tf.train.Saver(tf.global_variables())
def __init__(self, input_value, is_training):
self.inputs = input_value
self.anchors = data_pre.get_anchors(cfg.Main.anchors)
self.classes = data_pre.get_classes(cfg.Main.classes)
self.max_iou_ = 0.5
self.strides = [8, 16, 32]
self.backbone = cfg.Main.backbone
self.is_training = is_training
with tf.variable_scope('yolo3_model'):
# 52,26,13
self.conv_one, self.conv_two, self.conv_three = self.__network()
# 52,26,13
self.pre_one = self.__yolo_layer(self.conv_one, self.strides[0],
self.anchors[0:3])
self.pre_two = self.__yolo_layer(self.conv_two, self.strides[1],
self.anchors[3:6])
# error
self.pre_three = self.__yolo_layer(self.conv_three,
self.strides[2],
self.anchors[6:9])
def __init__(self):
self.classes = data_pre.get_classes(cfg.Main.classes)
self.output_size = 416
self.bboxes_score = cfg.Test.bboxes_score
self.nms_score = cfg.Test.nms_score
def __init__(self):
# data
self.classes = data_pre.get_classes(cfg.Main.classes) # 类别
self.anchors = data_pre.get_anchors(cfg.Main.anchors) # anchors
self.train_set = data_pre.Dataset()
self.test_set = data_pre.Dataset('test')
# Train
self.model_savefile = cfg.Train.darknet_savefile if cfg.Main.backbone == 0 else cfg.Train.mobilenet_savefile
self.epoch = cfg.Train.epoch
self.is_twostep = cfg.Train.is_twostep
self.one_step = cfg.Train.one_step
self.two_step = cfg.Train.two_step
self.batch_size = cfg.Train.batch_size
self.is_training = cfg.Train.is_training
self.input_size = cfg.Train.input_size
# learning_rate
self.first_learn_rate = 1e-4 # cos衰减学习率 ,
self.decay_steps = 200 # 衰减步长
self.alpha = 0.3 # alpha
# other
self.darknet_savefile = './checkpoint/darknet_yolo3/yolov3.ckpt'
self.ses = tf.Session(config=gpu_config)
np.random.seed(21)
self.train_epoch_loss, self.test_epoch_loss = [], []
with tf.name_scope('inputs'):
self.input_images = tf.placeholder(dtype=tf.float32,
shape=[self.batch_size, self.input_size, self.input_size, 3],
name='input_images')
self.one_bbox = tf.placeholder(dtype=tf.float32, name='one_bbox')
self.two_bbox = tf.placeholder(dtype=tf.float32, name='two_bbox')
self.three_bbox = tf.placeholder(dtype=tf.float32, name='three_bbox')
self.one_recover = tf.placeholder(dtype=tf.float32, name='one_recover')
self.two_recover = tf.placeholder(dtype=tf.float32, name='two_recover')
self.three_recover = tf.placeholder(dtype=tf.float32, name='three_recover')
with tf.name_scope('model_create'):
self.model = Yolo3(input_value=self.input_images, is_training=self.is_training)
self.load_variables = tf.global_variables()
with tf.name_scope('loss'):
self.giou_loss, self.conf_loss, self.pro_loss = self.model.loss_total(self.one_bbox, self.two_bbox,
self.three_bbox, self.one_recover,
self.two_recover, self.three_recover)
self.total_loss = self.giou_loss + self.conf_loss + self.pro_loss
with tf.variable_scope('learn_rate'):
self.global_step = tf.Variable(0.0, dtype=tf.float32, name='global_step', trainable=False)
self.save_step = tf.Variable(0.0, dtype=tf.float32, name='save_step', trainable=False)
self.learn_rate = tf.train.exponential_decay(learning_rate=1e-4, global_step=self.global_step,
decay_steps=2000, decay_rate=0.9, staircase=True,
name='learning_rate')
# self.learn_rate = 1e-4
self.global_step = tf.assign_add(self.global_step, 1.0)
self.save_step = tf.assign_add(self.save_step, 1.0)
# with tf.name_scope("define_weight_decay"):
# moving_ave = tf.train.ExponentialMovingAverage(0.9995).apply(tf.trainable_variables())
with tf.name_scope('optimizer'):
# 第一部分: 主要训练分类与回归
self.first_trainable_var_list = []
for var in tf.trainable_variables():
var_name = var.op.name
var_name_mess = str(var_name).split('/')
if var_name_mess[1] in ['conv_three', 'conv_two', 'conv_one']:
self.first_trainable_var_list.append(var)
self.first_op = tf.train.AdamOptimizer(self.learn_rate).minimize(self.total_loss,
var_list=self.first_trainable_var_list)
# 第二部分: 整体训练
self.second_trainable_var_list = tf.trainable_variables()
self.second_op = tf.train.GradientDescentOptimizer(self.learn_rate).minimize(self.total_loss,
var_list=self.second_trainable_var_list)
# tf.keras.callbacks.ReduceLROnPlateau()
# 梯度截断
# optimizer = tf.train.AdamOptimizer(self.learn_rate,beta1=0.5)
# grads, variable = zip(*optimizer.compute_gradients(self.total_loss))
# grads, global_norm = tf.clip_by_global_norm(grads,5)
# self.second_op = optimizer.apply_gradients(zip(grads,variable))
# grad_op = tf.ConditionalAccumulator(dtype=tf.float32)
# grad_op.set_global_step(self.global_step)
# grad_op.apply_grad()
with tf.name_scope('saver_log'):
self.load = tf.train.Saver(self.load_variables)
self.saver = tf.train.Saver(tf.global_variables())
giou_loss = tf.summary.scalar("giou_loss", self.giou_loss)
conf_loss = tf.summary.scalar("conf_loss", self.conf_loss)
prob_loss = tf.summary.scalar("prob_loss", self.pro_loss)
total_loss = tf.summary.scalar("total_loss", self.total_loss)
learn_rate = tf.summary.scalar("learn_rate", self.learn_rate)
train_epoch_loss = tf.summary.scalar('train_epoch_loss', np.mean(self.train_epoch_loss))
test_epoch_loss = tf.summary.scalar('test_epoch_loss', np.mean(self.test_epoch_loss))
logdir = './log/'
# if os.path.exists(logdir): shutil.rmtree(logdir)
# os.mkdir(logdir)
self.batch_summary_op = tf.summary.merge([giou_loss, conf_loss, prob_loss, total_loss, learn_rate])
self.epoch_summary_op = tf.summary.merge([train_epoch_loss, test_epoch_loss])
self.batch_summary_writer = tf.summary.FileWriter(logdir + '/batch/', graph=self.ses.graph)
self.epoch_summary_writer = tf.summary.FileWriter(logdir + '/epoch/', graph=self.ses.graph)
self.batch_summary_writer.add_graph(self.ses.graph)
self.epoch_summary_writer.add_graph(self.ses.graph)
def __init__(self):
# data
self.classes = data_pre.get_classes(cfg.Main.classes) # 类别
self.anchors = data_pre.get_anchors(cfg.Main.anchors) # anchors
# Dataset数据集
self.train_set = dataset_TFRecord.Dataset('train')
self.test_set = dataset_TFRecord.Dataset('test')
self.train_itera = self.train_set.bulid_data().make_one_shot_iterator()
self.test_itera = self.test_set.bulid_data().make_one_shot_iterator()
# Train
self.model_savefile = cfg.Train.darknet_savefile if cfg.Main.backbone == 0 else cfg.Train.mobilenet_savefile
self.epoch = cfg.Train.epoch
self.one_step = cfg.Train.one_step
self.two_step = cfg.Train.two_step
self.batch_size = cfg.Train.batch_size
self.is_training = cfg.Train.is_training
self.input_size = cfg.Train.input_size
# learning_rate
self.first_learn_rate = 1e-4 # cos衰减学习率 ,
self.decay_steps = 200 # 衰减步长
self.alpha = 0.3 # alpha
# other
self.darknet_savefile = './checkpoint/darknet/yolov3.ckpt'
self.ses = tf.Session()
self.is_train_flag = True
np.random.seed(21)
with tf.name_scope('inputs'):
if self.is_train_flag: # 判断是train, 还是test, 切换不同数据集
inputs = self.train_itera.get_next()
else:
inputs = self.test_itera.get_next()
# reshape维度
self.input_images = tf.reshape(inputs[3],[self.batch_size,416,416,3])
self.one_bbox = tf.reshape(inputs[0],[self.batch_size,52, 52, 3, 65])
self.two_bbox = tf.reshape(inputs[1],[self.batch_size,26, 26, 3, 65])
self.three_bbox = tf.reshape(inputs[2],[self.batch_size,13, 13, 3, 65])
self.one_recover = tf.reshape(inputs[4],[self.batch_size,150,4])
self.two_recover = tf.reshape(inputs[5],[self.batch_size,150,4])
self.three_recover = tf.reshape(inputs[6],[self.batch_size,150,4])
with tf.name_scope('model_create'):
self.model = Yolo3(input_value=self.input_images,is_training=self.is_training)
self.load_variables = tf.global_variables()
with tf.name_scope('loss'):
self.giou_loss, self.conf_loss, self.pro_loss = self.model.loss_total(self.one_bbox, self.two_bbox,
self.three_bbox, self.one_recover,
self.two_recover, self.three_recover)
self.total_loss = self.giou_loss + self.conf_loss + self.pro_loss
with tf.variable_scope('learn_rate'):
self.global_step = tf.Variable(1.0, dtype=tf.float32, name='global_step', trainable=False)
# self.learn_rate = tf.train.cosine_decay(learning_rate=self.first_learn_rate, global_step=self.global_step,
# decay_steps=self.decay_steps, alpha=self.alpha)
self.learn_rate = tf.train.exponential_decay(learning_rate=1e-4,global_step=self.global_step,decay_steps=300,decay_rate=0.9,name='learning_rate')
self.global_step = tf.assign_add(self.global_step, 1.0)
# with tf.name_scope("define_weight_decay"):
# moving_ave = tf.train.ExponentialMovingAverage(0.9995).apply(tf.trainable_variables())
with tf.name_scope('optimizer'):
# 第一部分: 主要训练分类与回归
self.first_trainable_var_list = []
for var in tf.trainable_variables():
var_name = var.op.name
var_name_mess = str(var_name).split('/')
if var_name_mess[1] in ['conv_three', 'conv_two', 'conv_one']:
self.first_trainable_var_list.append(var)
self.first_op = tf.train.AdamOptimizer(self.learn_rate).minimize(self.total_loss,
var_list=self.first_trainable_var_list)
# 第二部分: 整体训练
second_trainable_var_list = tf.trainable_variables()
# self.second_op = tf.train.AdamOptimizer(self.learn_rate).minimize(self.total_loss,
# var_list=second_trainable_var_list)
self.second_op = tf.train.AdamOptimizer(self.learn_rate).minimize(self.total_loss,var_list=second_trainable_var_list)
# 梯度截断
# optimizer = tf.train.AdamOptimizer(self.learn_rate,beta1=0.5)
# grads, variable = zip(*optimizer.compute_gradients(self.total_loss))
# grads, global_norm = tf.clip_by_global_norm(grads,5)
# self.second_op = optimizer.apply_gradients(zip(grads,variable))
with tf.name_scope('saver_log'):
self.load = tf.train.Saver(self.load_variables)
self.saver = tf.train.Saver(tf.global_variables())
tf.summary.scalar("giou_loss", self.giou_loss)
tf.summary.scalar("conf_loss", self.conf_loss)
tf.summary.scalar("prob_loss", self.pro_loss)
tf.summary.scalar("total_loss", self.total_loss)
tf.summary.scalar("learn_rate", self.learn_rate)
logdir = './log/'
if os.path.exists(logdir): shutil.rmtree(logdir)
os.mkdir(logdir)
self.summary_op = tf.summary.merge_all()
self.summary_writer = tf.summary.FileWriter(logdir, graph=self.ses.graph)