def network_input( choice):
parser = Parser(int(choice))
trainset = dataset(parser, Config.Train_tfrecord, Config.BATCH_SIZE, shuffle=Config.SHUFFLE_SIZE)
validset = dataset(parser, Config.Valid_tfrecord, 1, shuffle=None)
example = tf.cond(is_training, lambda: trainset.get_next(), lambda: validset.get_next())
images, y_true = example
return images, y_true
def load_train_val_data(train_path, val_path):
parser_train = Parser(ANCHORS, NUM_CLASSES, image_size=(416, 416))
parser_test = Parser(ANCHORS, NUM_CLASSES)
trainset = dataset(parser_train,
train_path,
BATCH_SIZE,
shuffle=SHUFFLE_SIZE,
multi_image_size=False)
testset = dataset(parser_test, val_path, BATCH_SIZE, shuffle=None)
return trainset, testset
def show_example(self):
sess = tf.Session()
classes = os.listdir(self.source_dir[:-len(".tfrecords") or None])
train_tfrecord = self.source_dir
anchors = utils.get_anchors(self.anchor_dir, self.img_h, self.img_w)
parser = Parser(image_h=self.img_h,
image_w=self.img_w,
anchors=anchors,
num_classes=self.num_classes,
cell_size=self.cell_size,
debug=True)
trainset = dataset(parser, train_tfrecord, 1, shuffle=1)
is_training = tf.placeholder(tf.bool)
example = trainset.get_next()
image, boxes = sess.run(example)
image, boxes = image[0], boxes[0]
n_box = len(boxes)
print(boxes)
image = np.repeat(image, 3, axis=2)
for i in range(n_box):
image = cv2.rectangle(
image, (int(float(boxes[i][0])), int(float(boxes[i][1]))),
(int(float(boxes[i][2])), int(float(boxes[i][3]))),
(255, 0, 0), 1)
label = str(int(float(boxes[i][4])))
image = cv2.putText(
image, label,
(int(float(boxes[i][0])), int(float(boxes[i][1]))),
cv2.FONT_HERSHEY_SIMPLEX, .6, (0, 255, 0), 1, 2)
image = Image.fromarray(np.uint8(image * 255))
image.show()
BATCH_SIZE = 8
STEPS = 25000
LR = 0.001 # if Nan, set 0.0005, 0.0001
DECAY_STEPS = 100
DECAY_RATE = 0.9
SHUFFLE_SIZE = 200
CLASSES = utils.read_coco_names(class_name_path)
ANCHORS = utils.get_anchors('data/voc_anchors.txt', IMAGE_H, IMAGE_W)
NUM_CLASSES = len(CLASSES)
EVAL_INTERNAL = 100
SAVE_INTERNAL = 500
parser = Parser(IMAGE_H, IMAGE_W, ANCHORS, NUM_CLASSES)
trainset = dataset(parser, train_tfrecord, BATCH_SIZE, shuffle=SHUFFLE_SIZE)
testset = dataset(parser, test_tfrecord , BATCH_SIZE, shuffle=None)
is_training = tf.placeholder(tf.bool)
example = tf.cond(is_training, lambda: trainset.get_next(), lambda: testset.get_next())
images, *y_true = example
model = yolov3.yolov3(NUM_CLASSES, ANCHORS)
with tf.variable_scope('yolov3'):
pred_feature_map = model.forward(images, is_training=is_training)
loss = model.compute_loss(pred_feature_map, y_true)
y_pred = model.predict(pred_feature_map)
tf.summary.scalar("loss/coord_loss", loss[1])
tf.summary.scalar("loss/sizes_loss", loss[2])
import tensorflow as tf
from core import utils
from PIL import Image
from core.dataset import Parser, dataset
sess = tf.Session()
BATCH_SIZE = 1
SHUFFLE_SIZE = 1
CLASSES = utils.read_coco_names('./data/class.names')
ANCHORS = utils.get_anchors('./data/anchors.txt')
NUM_CLASSES = len(CLASSES)
TRAIN_TFRECORD = "./data/train_data/val.tfrecords"
TEST_TFRECORD = "./data/val_data/val.tfrecords"
parser_train = Parser(ANCHORS, NUM_CLASSES)
parser_test = Parser(ANCHORS, NUM_CLASSES)
trainset = dataset(parser_train, TRAIN_TFRECORD, BATCH_SIZE, shuffle=SHUFFLE_SIZE, multi_image_size=False)
testset = dataset(parser_test, TEST_TFRECORD, BATCH_SIZE, shuffle=None)
is_training = tf.placeholder(tf.bool)
example = testset.get_next()
for l in range(1):
res = sess.run(example)
image = res[0][0] * 255
y_true = res[1:]
boxes = utils.decode_gtbox(y_true)
n_box = len(boxes)
for i in range(n_box):
image = cv2.rectangle(image,(int(float(boxes[i][0])),
int(float(boxes[i][1]))),
(int(float(boxes[i][2])),
int(float(boxes[i][3]))), (255,0,0), 1)
CLASSES = utils.read_coco_names('./data/raccoon.names')
NUM_CLASSES = len(CLASSES)
ANCHORS = utils.get_anchors('./data/raccoon_anchors.txt', IMAGE_H, IMAGE_W)
CKPT_FILE = "./checkpoint/yolov3.ckpt-2500"
IOU_THRESH = 0.5
SCORE_THRESH = 0.3
all_detections = []
all_annotations = []
all_aver_precs = {CLASSES[i]: 0. for i in range(NUM_CLASSES)}
test_tfrecord = "./raccoon_dataset/raccoon_*.tfrecords"
parser = Parser(IMAGE_H, IMAGE_W, ANCHORS, NUM_CLASSES)
testset = dataset(parser,
test_tfrecord,
batch_size=1,
shuffle=None,
repeat=False)
images_tensor, *y_true_tensor = testset.get_next()
model = yolov3.yolov3(NUM_CLASSES, ANCHORS)
with tf.variable_scope('yolov3'):
pred_feature_map = model.forward(images_tensor, is_training=False)
y_pred_tensor = model.predict(pred_feature_map)
saver = tf.train.Saver()
saver.restore(sess, CKPT_FILE)
try:
image_idx = 0
while True:
LR = 0.0001 # if Nan, set 0.0005, 0.0001
DECAY_STEPS = 100
DECAY_RATE = 0.9
SHUFFLE_SIZE = 45 # 这个要根据输入的训练图片数量进行修改！！！
CLASSES = utils.read_coco_names('../data/object.names')
ANCHORS = utils.get_anchors('../data/object_anchors.txt', IMAGE_H, IMAGE_W)
NUM_CLASSES = len(CLASSES)
EVAL_INTERNAL = 1
SAVE_INTERNAL = 1
# 在一个计算图开始前，将文件读入到queue中，TFRecord可以格式统一管理存储数据
train_tfrecord = "../data/images_train.tfrecords"
test_tfrecord = "../data/images_test.tfrecords"
parser = Parser(IMAGE_H, IMAGE_W, ANCHORS, NUM_CLASSES)
trainset = dataset(parser, train_tfrecord, batch_size=BATCH_SIZE, shuffle=SHUFFLE_SIZE)
testset = dataset(parser, test_tfrecord, batch_size=BATCH_SIZE, shuffle=None) # 这里我用全部测试集来训练，方便看效果
is_training = tf.placeholder(tf.bool) # 占位符，运行时必须传入值
# 根据是is_training判断是不是在训练，然后调用trainset或者testset的get_next
example = tf.cond(is_training, lambda: trainset.get_next(), lambda: testset.get_next())
# y_true = [feature_map_1 , feature_map_2 , feature_map_3]
images, *y_true = example # a,*c = 1,2,3,4 a=1, c = [2,3,4]
model = YOLOv3.yolov3(NUM_CLASSES, ANCHORS)
with tf.variable_scope('yolov3'):
NUM_CLASSES = len(CLASSES)
CLASS_WEIGHTS = np.ones(NUM_CLASSES, dtype='float32')
TRUE_BOX_BUFFER = 20
tfrecord = "./test_data/train0003.tfrecords"
sess = tf.Session()
parser = parser(IMAGE_H,
IMAGE_W,
GRID_H,
GRID_W,
ANCHORS,
NUM_CLASSES,
DEBUG=True)
trainset = dataset(parser, tfrecord, BATCH_SIZE, shuffle=1)
example = trainset.get_next()
for l in range(2):
image, boxes = sess.run(example)
image, boxes = image[0], boxes[0]
n_box = len(boxes)
for i in range(n_box):
image = cv2.rectangle(
image, (int(float(boxes[i][0])), int(float(boxes[i][1]))),
(int(float(boxes[i][2])), int(float(boxes[i][3]))), (255, 0, 0), 1)
image = Image.fromarray(np.uint8(image))
image.show()
import tensorflow as tf
from core import utils, yolov3
from core.dataset import dataset, Parser
from PIL import Image
sess = tf.Session()
IMAGE_H, IMAGE_W = 416, 416
BATCH_SIZE = 1
SHUFFLE_SIZE = 200
CLASSES = utils.read_coco_names("data/SVHN/SVHN.names")
ANCHORS = utils.get_anchors("data/SVHN/SVHN_anchors.txt", IMAGE_H, IMAGE_W)
NUM_CLASSES = len(CLASSES)
test_tfrecord = "data/SVHN/tfrecords/quick_test_data.tfrecords"
parser = Parser(IMAGE_H, IMAGE_W, ANCHORS, NUM_CLASSES)
testset = dataset(parser, test_tfrecord, BATCH_SIZE, shuffle=None)
is_training = tf.placeholder(tf.bool)
example = testset.get_next()
images, *y_true = example
model = yolov3.yolov3(NUM_CLASSES, ANCHORS)
with tf.variable_scope('yolov3'):
pred_feature_map = model.forward(images, is_training=is_training)
loss = model.compute_loss(pred_feature_map, y_true)
y_pred = model.predict(pred_feature_map)
saver = tf.train.Saver()
saver.restore(sess, "data/SVHN/checkpoint5/yolov3.ckpt-4000")
def train(self):
ANCHORS = utils.get_anchors(self.anchors_path, self.img_h, self.img_w)
parser = Parser(image_h=self.img_h,
image_w=self.img_w,
anchors=ANCHORS,
num_classes=self.num_classes)
trainset = dataset(parser,
self.train_records,
self.batch_size,
shuffle=self.shuffle_size)
testset = dataset(parser,
self.test_records,
self.batch_size,
shuffle=None)
is_training = tf.placeholder(tf.bool)
example = tf.cond(is_training, lambda: trainset.get_next(),
lambda: testset.get_next())
images, y_true = example
model = yolov3.yolov3(self.num_classes, ANCHORS)
with tf.variable_scope('yolov3'):
# Give the images to the network, and receive a prediction
# feature map
pred_feature_map = model.forward(images,
is_training=is_training,
n_filters_dn=self.n_filters_dn,
n_strides_dn=self.n_strides_dn,
n_ksizes_dn=self.n_ksizes_dn)
loss = model.compute_loss(pred_feature_map, y_true,
self.iou_threshold)
y_pred = model.predict(pred_feature_map)
tf.summary.scalar("loss/coord_loss", loss[1])
tf.summary.scalar("loss/sizes_loss", loss[2])
tf.summary.scalar("loss/confs_loss", loss[3])
tf.summary.scalar("loss/class_loss", loss[4])
global_step = tf.Variable(0,
trainable=True,
collections=[tf.GraphKeys.LOCAL_VARIABLES])
write_op = tf.summary.merge_all()
writer_train = tf.summary.FileWriter("../../data/train_summary",
sess.graph)
writer_test = tf.summary.FileWriter("../../data/test_summary")
update_vars = tf.contrib.framework.get_variables_to_restore(
include=["yolov3/yolo-v3"])
lr = tf.train.exponential_decay(self.learning_rate,
global_step,
decay_steps=self.decay_steps,
decay_rate=self.decay_rate,
staircase=True)
optimizer = tf.train.AdamOptimizer(lr)
# set dependencies for BN ops
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss[0],
var_list=update_vars,
global_step=global_step)
sess.run([
tf.global_variables_initializer(),
tf.local_variables_initializer()
])
saver = tf.train.Saver(max_to_keep=2)
for step in range(self.steps):
run_items = sess.run([train_op, write_op, y_pred, y_true] + loss,
feed_dict={is_training: True})
if (step + 1) % self.eval_internal == 0:
train_rec_value, train_prec_value = utils.evaluate(
run_items[2], run_items[3])
writer_train.add_summary(run_items[1], global_step=step)
writer_train.flush() # Flushes the event file to disk
if (step + 1) % self.save_internal == 0:
saver.save(sess,
save_path=self.checkpoint_path,
global_step=step + 1)
if (step + 1) % self.print_every_n == 0:
print(f"=> STEP {step+1} [TRAIN]:\tloss_xy: " +
f"{run_items[5]:.4f} \tloss_wh:{run_items[6]:.4f} \t" +
f"loss_conf:{run_items[7]:.4f} \tloss_class:" +
f"{run_items[8]:.4f}")
run_items = sess.run([write_op, y_pred, y_true] + loss,
feed_dict={is_training: False})
if (step + 1) % self.eval_internal == 0:
test_rec_value, test_prec_value = utils.evaluate(
run_items[1], run_items[2])
print(f"\n{20*'='}> evaluation result <{20*'='}\n")
print(f"=> STEP {step+1} [TRAIN]:\trecall:" +
f"{train_rec_value:.2f} \tprecision:" +
f"{train_prec_value:.4f}")
print(f"=> STEP {step+1} [VALID]:\trecall:" +
f"{test_rec_value:.2f} \tprecision:" +
f"{test_prec_value:.4f}")
print(f"\n{20*'='}> evaluation result <{20*'='}\n")
writer_test.add_summary(run_items[0], global_step=step)
writer_test.flush() # Flushes the event file to disk
TRUE_BOX_BUFFER = 80
# test_tfrecord = "./test_data/train0003.tfrecords"
train_tfrecord = "../voc/train*.tfrecords"
val_tfrecord = "../voc/val*.tfrecords"
sess = tf.Session()
parser = parser(IMAGE_H,
IMAGE_W,
GRID_H,
GRID_W,
ANCHORS,
NUM_CLASSES,
TRUE_BOX_BUFFER,
DEBUG=False)
trainset = dataset(parser, train_tfrecord, BATCH_SIZE, shuffle=1)
valset = dataset(parser, val_tfrecord, BATCH_SIZE, shuffle=None)
is_training = tf.placeholder(tf.bool)
example = tf.cond(is_training, lambda: trainset.get_next(),
lambda: valset.get_next())
input_image, y_true, true_boxes = example
feature_extractor = backbone.FullYoloFeature(input_image, is_training)
features = feature_extractor.feature
# ========================> method 1 <================================
# output = tf.keras.layers.Conv2D(NUM_ANCHORS * (5 + NUM_CLASSES),
# (1,1), strides=(1,1),
# padding='same',
def split_dataset_for_oldScene_KCV(args):
data_dir = args.cur_data_dir
num_negatives = args.num_negatives
test_candidates_nums = args.test_candidates_nums
kcv = args.kcv
name = 'oldScene_neg{}_testCandi{}'.format(num_negatives,
test_candidates_nums)
result_root = os.path.join(data_dir, 'split_data_oldScene', name)
mashup_api_list = meta_data(args).mashup_api_list
mashup_api_dict = list2dict(mashup_api_list)
# 返回K个数据对象，每个可以使用相关属性
if os.path.exists(dataset(args, result_root, name,
kcv - 1).train_df_path): # 已经划分过
for i in range(kcv):
print('has splited data in kcv mode before,read them!')
data = dataset(args, result_root, name, i)
data.initialize() # 从文件中读取对象
set_ds(data) # 设置唯一实例
yield data
else: # 还未划分过
mashup_ids = list(mashup_api_dict.keys())
all_apis = set(meta_data(args).api_df.index.tolist()) # 所有api的id
# 首先为每个mashup指定确定的正负例，候选api等
# {mid:api_instances}
mid2true_instances, mid2false_instances, mid2candidate_instances = {}, {}, {}
for mashup_id, api_ids in mashup_api_dict.items(): # api_ids是set
unobserved_apis_list = list(all_apis - api_ids)
random.shuffle(unobserved_apis_list)
mid2true_instances[mashup_id] = {}
mid2false_instances[mashup_id] = {}
mid2candidate_instances[mashup_id] = {}
api_ids_list = list(api_ids) # 已选择的apis，做正例
mid2true_instances[mashup_id] = api_ids_list
all_neg_num = min(
meta_data(args).api_num, num_negatives * len(api_ids_list))
mid2false_instances[
mashup_id] = unobserved_apis_list[:all_neg_num] # 负例
if test_candidates_nums == 'all': # 选取全部api做测试
mid2candidate_instances[mashup_id] = list(all_apis)
else: # 选取部分作为测试，实际组件api和部分unobserved
test_candidates_nums = int(test_candidates_nums)
mid2candidate_instances[
mashup_id] = api_ids_list + unobserved_apis_list[:
test_candidates_nums]
random.shuffle(mashup_ids)
batch = len(mashup_ids) // kcv
for i in range(kcv): # 每个kcv
start_index = i * batch
batch_stopindex = len(mashup_ids) if i == kcv - 1 else (i +
1) * batch
test_mashups = mashup_ids[start_index:batch_stopindex]
train_mashups = mashup_ids[:start_index] + mashup_ids[
batch_stopindex:-1]
train_df = pd.DataFrame(columns=['mashup', 'api', 'label'])
for mashup_id in train_mashups:
for true_api_id in mid2true_instances[mashup_id]:
train_df.append(
{
'mashup': mashup_id,
'api': true_api_id,
'label': 1
},
ignore_index=True)
for false_api_id in mid2false_instances[mashup_id]:
train_df.append(
{
'mashup': mashup_id,
'api': false_api_id,
'label': 0
},
ignore_index=True)
# test和train格式不同
# test mashup和api的一行list是多个测试样本,而all_ground_api_ids,test_slt_ids的一行对应前者的一行
test_df = pd.DataFrame(columns=[
'mashup', 'slt_apis', 'candidate_apis', 'all_ground_api_ids'
])
for mashup_id in test_mashups:
test_df.append(
{
'mashup': mashup_id,
'candidate_apis': mid2candidate_instances[mashup_id],
'all_ground_api_ids': mid2true_instances[mashup_id]
},
ignore_index=True)
data = dataset(args, result_root, name, i)
data.initialize(train_df, test_df)
set_ds(data) # 设置唯一实例
print('{}/{} dataset, build done!'.format(i, kcv))
yield data
def split_dataset_for_newScene_KCV(args):
"""
新场景划分数据
:param data_dir: 要划分数据的路径
:param num_negatives: 负采样比例
:param slt_num: 指定的最大已选择服务的数目
:param slt_combination_num: 真实组件服务中,只选取一部分组合作为已选服务,缓解数据不平衡问题: eg: C10/3 C50/3
# :param train_positive_samples: 每个训练用的mashup,除了已选服务,剩下的保留多少个服务作为训练正例，防止组件太多的mashup所占的比例太大
:param test_candidates_nums: 每个mashup要评价多少个待测负例item: 为all时全部评价
:param kcv:
:return: 某折的dataset对象
"""
data_dir = args.cur_data_dir
num_negatives = args.num_negatives
slt_num = args.slt_item_num
slt_combination_num = args.combination_num
test_candidates_nums = args.test_candidates_nums
kcv = args.kcv
name = 'newScene_neg{}_sltNum{}_com{}_testCandi{}'.format(
num_negatives, slt_num, slt_combination_num, test_candidates_nums)
result_root = os.path.join(data_dir, 'split_data_newScene', name)
# 返回K个dataset对象，每个可以使用相关属性
if os.path.exists(dataset(args, result_root, name,
kcv - 1).train_df_path): # 已经划分过
for i in range(kcv):
print('data has been splited in kcv mode before,read them!')
data = dataset(args, result_root, name, i)
data.initialize() # 从文件中读取对象
set_ds(data) # 设置唯一实例
yield data
else:
mashup_api_list = meta_data(args).mashup_api_list
mashup_api_dict = list2dict(mashup_api_list)
mashup_ids = meta_data(args).mashup_df.index.tolist()
mashup_ids.remove(0) # 占位
all_apis = set(meta_data(args).api_df.index.tolist()) # 所有api的id
all_apis.remove(0)
# 1.首先为每个mashup指定已选服务和对应的正负例(训练)/待测api(测试)
# {mid:{slt_aid_list:api_instances}
mid2true_instances, mid2false_instances, mid2candidate_instances = {}, {}, {}
for mashup_id, api_ids in mashup_api_dict.items(): # api_ids是set
unobserved_apis_list = list(all_apis - api_ids)
random.shuffle(unobserved_apis_list)
mid2true_instances[mashup_id] = {}
mid2false_instances[mashup_id] = {}
mid2candidate_instances[mashup_id] = {}
api_ids_list = list(api_ids)
max_slt_num = min(slt_num,
len(api_ids_list) -
1) # eg:最大需要三个已选服务，但是只有2个services构成
for act_slt_num in range(max_slt_num): # 选择1个时，两个时...
act_slt_num += 1
combinations = list(
itertools.combinations(api_ids_list, act_slt_num))
if slt_combination_num != 'all': # 只选取一部分组合,缓解数据不平衡问题
slt_combination_num = min(len(combinations),
slt_combination_num)
combinations = combinations[:slt_combination_num]
for slt_api_ids in combinations: # 随机组合已选择的api，扩大数据量 # 组合产生,当做已选中的apis
train_api_ids = list(
api_ids - set(slt_api_ids)
) # masked observed interaction 用于训练或测试的
# if train_positive_samples != 'all': # 选择一部分正例 做训练或测试
# train_positive_samples_num = min(len(train_api_ids), train_positive_samples) # 最多50个，一般没有那么多
# train_api_ids = train_api_ids[:train_positive_samples_num]
mid2true_instances[mashup_id][
slt_api_ids] = train_api_ids # 训练用正例 slt_api_ids是tuple
num_negative_instances = min(
num_negatives * len(train_api_ids),
len(unobserved_apis_list))
mid2false_instances[mashup_id][
slt_api_ids] = unobserved_apis_list[:
num_negative_instances] # 随机选择的负例
if test_candidates_nums == 'all': # 待预测
test_candidates_list = list(all_apis -
set(slt_api_ids))
else:
test_candidates_nums = int(test_candidates_nums)
test_candidates_list = unobserved_apis_list[:
test_candidates_nums] + train_api_ids
mid2candidate_instances[mashup_id][
slt_api_ids] = test_candidates_list
random.shuffle(mashup_ids)
batch = len(mashup_ids) // kcv
# 2.然后，根据上面的结果划分为各个KCV，训练和测试
for i in range(kcv): # 每个kcv
start_index = i * batch
batch_stopindex = len(mashup_ids) if i == kcv - 1 else (i +
1) * batch
test_mashups = mashup_ids[start_index:batch_stopindex]
train_mashups = mashup_ids[:start_index] + mashup_ids[
batch_stopindex:-1]
print(train_mashups)
print(test_mashups)
train_df = pd.DataFrame(
columns=['mashup', 'slt_apis', 'api', 'label'])
for mashup_id in train_mashups:
for slt_api_ids, true_api_instances in mid2true_instances[
mashup_id].items():
for true_api_id in true_api_instances:
train_df = train_df.append(
{
'mashup': mashup_id,
'slt_apis': slt_api_ids,
'api': true_api_id,
'label': 1
},
ignore_index=True)
for slt_api_ids, false_api_instances in mid2false_instances[
mashup_id].items():
for false_api_id in false_api_instances:
train_df = train_df.append(
{
'mashup': mashup_id,
'slt_apis': slt_api_ids,
'api': false_api_id,
'label': 0
},
ignore_index=True)
# test和train格式不同: train是一行一个样本; test的待测api太多，节省空间，一行:
# 一个mashup，多个待测api,一份all_ground_api_ids,一份test_slt_ids
test_df = pd.DataFrame(columns=[
'mashup', 'slt_apis', 'candidate_apis', 'all_ground_api_ids'
])
for mashup_id in test_mashups:
for slt_api_ids, candidate_api_instances in mid2candidate_instances[
mashup_id].items():
test_df = test_df.append(
{
'mashup':
mashup_id,
'slt_apis':
slt_api_ids,
'candidate_apis':
candidate_api_instances,
'all_ground_api_ids':
mid2true_instances[mashup_id][slt_api_ids]
},
ignore_index=True)
# 3.根据训练集和测试集的划分，初始化dataset对象!!!
data = dataset(args, result_root, name, i)
data.initialize(train_df, test_df)
set_ds(data) # 设置唯一实例
print('{}/{} dataset, build done!'.format(i, kcv))
yield data
print('you have splited and saved them!')