def main(argv):
tf_flags = tf.app.flags.FLAGS
# gpu config.
# config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
# config.gpu_options.allow_growth = True
# config = config
if tf_flags.phase == "train":
with tf.Session() as sess:
# when use queue to load data, not use with to define sess
train_model = unet.UNet(sess, tf_flags)
train_model.train(tf_flags.batch_size, tf_flags.training_steps,
tf_flags.summary_steps,
tf_flags.checkpoint_steps, tf_flags.save_steps)
else:
with tf.Session() as sess:
# test on a image pair.
test_model = unet.UNet(sess, tf_flags)
test_model.load(tf_flags.checkpoint)
image, output_masks = test_model.test()
# return numpy ndarray.
# save two images.
filename_A = "input.png"
filename_B = "output_masks.png"
cv2.imwrite(filename_A, np.uint8(image[0].clip(0., 1.) * 255.))
cv2.imwrite(filename_B,
np.uint8(output_masks[0].clip(0., 1.) * 255.))
# Utilize cv2.imwrite() to save images.
print("Saved files: {}, {}".format(filename_A, filename_B))
def main(argv):
# tf.app.flags.FLAGS接受命令行传递参数或者tf.app.flags定义的默认参数
tf_flags = tf.flags.FLAGS
# gpu config.
# tf.ConfigProto()函数用在创建session的时候，用来对session进行参数配置
config = tf.ConfigProto()
# tf提供了两种控制GPU资源使用的方法，第一种方式就是限制GPU的使用率:
# config.gpu_options.per_process_gpu_memory_fraction = 0.5 # 占用50%显存
# 第二种是让TensorFlow在运行过程中动态申请显存，需要多少就申请多少:
config.gpu_options.allow_growth = True
if tf_flags.phase == "train":
# 使用上面定义的config设置session
with tf.Session(config=config) as sess:
# when use queue to load data, not use with to define sess
# 定义Unet模型
train_model = unet.UNet(sess, tf_flags)
# 训练Unet网络，参数：batch_size,训练迭代步......
train_model.train(tf_flags.batch_size, tf_flags.training_steps,
tf_flags.summary_steps, tf_flags.checkpoint_steps, tf_flags.save_steps)
else:
with tf.Session(config=config) as sess:
# test on a image pair.
test_model = unet.UNet(sess, tf_flags)
# test阶段:加载checkpoint文件的数据给模型参数初始化
test_model.load(tf_flags.checkpoint)
test_model.test(os.path.join(tf_flags.testing_set,"test"))
print("Saved test files successfully !")
def main(argv):
tf_flags = tf.app.flags.FLAGS
# 参数名称、参数默认值、参数描述
tf.app.flags.DEFINE_string("output_dir", "model_output",
"checkpoint and summary directory.")
tf.app.flags.DEFINE_string("phase", "train", "model phase: train/test.")
tf.app.flags.DEFINE_string("training_set", "./datasets",
"dataset path for training.")
tf.app.flags.DEFINE_string("testing_set", "./datasets/test",
"dataset path for testing one image pair.")
tf.app.flags.DEFINE_integer("batch_size", 64, "batch size for training.")
tf.app.flags.DEFINE_integer("training_steps", 100000,
"total training steps.")
tf.app.flags.DEFINE_integer("summary_steps", 100, "summary period.")
tf.app.flags.DEFINE_integer("checkpoint_steps", 1000, "checkpoint period.")
tf.app.flags.DEFINE_integer("save_steps", 500, "checkpoint period.")
tf.app.flags.DEFINE_string("checkpoint", None,
"checkpoint name for restoring.")
# gpu config
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.5
if tf_flags.phase == "train":
with tf.Session(config=config) as sess:
# with tf.Session() as sess:
train_model = unet.UNet(sess, tf_flags)
train_model.train(tf_flags.batch_size, tf_flags.training_steps,
tf_flags.summary_steps,
tf_flags.checkpoint_steps, tf_flags.save_steps)
else:
with tf.Session(config=config) as sess:
# with tf.Session() as sess:
test_model = unet.UNet(sess, tf_flags)
test_model.load(tf_flags.checkpint)
image, output_masks = test_model.test()
filename_A = "input.png"
filename_B = "output_masks.png"
cv2.imwrite(filename_A, np.uint8(image[0].clip(0., 1.) * 255.))
cv2.imwrite(filename_B,
np.uint8(output_masks[0].clip(0., 1.) * 255.))
print("Saved files : {}, {}".format(filename_A, filename_B))
def main():
global best_prec1
batch_size = args.batch_size
split = args.split
num_workers = args.num_workers
if args.model == 'resnet34unet':
model = unet.UNet(3, 1, activation='sigmoid', weight=None)
preprocess = get_preprocessing_fn('resnet34')
else:
raise ValueError(args.model, ' is not in our model')
valid_aug = augmentation.get_augmentations('valid', 1.0, args.image_size)
valid_dataset = SIIM.SIIM_ACR(mode='valid',
split=split,
preprocess=preprocess,
augmentation=valid_aug)
val_dataloader = DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
pin_memory=True)
if args.use_flip:
valid_dataset_tta = SIIM.SIIM_ACR(mode='valid',
split=split,
preprocess=preprocess,
augmentation=valid_aug,
is_tta=True)
global val_dataloader_tta
val_dataloader_tta = DataLoader(valid_dataset_tta,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
pin_memory=True)
model = nn.DataParallel(model, device_ids=args.gpus).cuda()
global metric
metric = metrics.FscoreMetric(threshold=None)
for i in range(args.snapshot):
if os.path.exists(os.path.join('ckpt', args.model, str(args.split))):
print('=> loading checkpoint {}'.format(
os.path.join('ckpt', args.model, str(args.split))))
ckpt = torch.load(
os.path.join('ckpt', args.model, str(args.split),
str(i) + '_size%d' % args.image_size + '.pth'))
model.load_state_dict(ckpt)
else:
raise ValueError('=> no checkpoint found at {}'.format(
os.path.join('ckpt', args.model)))
prec1 = validate(val_dataloader, model, i)
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
# gpu config.
config = tf.ConfigProto()
#config.gpu_options.per_process_gpu_memory_fraction = 0.5
# config.gpu_options.allow_growth = True
if FLAGS.phase == 'train':
with tf.Session(config=config) as sess:
# when use queue to load data, not use with to define sess
train_model = unet.UNet(sess, FLAGS, is_training=True)
train_model.train(
batch_size=FLAGS.train_batch_size,
training_number_of_steps=FLAGS.training_number_of_steps,
summary_steps=FLAGS.summary_steps,
checkpoint_steps=FLAGS.checkpoint_steps,
save_steps=FLAGS.save_steps,
dataset_dir=FLAGS.dataset_dir,
dataset=FLAGS.dataset,
tf_initial_checkpoint_dir=FLAGS.tf_initial_checkpoint_dir)
else:
with tf.Session(config=config) as sess:
# test on a image pair.
test_model = unet.UNet(sess, FLAGS)
test_model.load(FLAGS.checkpoint)
image, output_masks = test_model.test()
# return numpy ndarray.
# save two images.
print('\n\n\n\n???\n\n\n\n')
filename_A = 'input.png'
filename_B = 'output_masks.png'
cv2.imwrite(filename_A, np.uint8(image[0].clip(0., 1.) * 255.))
cv2.imwrite(filename_B,
np.uint8(output_masks[0].clip(0., 1.) * 255.))
# Utilize cv2.imwrite() to save images.
print('Saved files: {}, {}'.format(filename_A, filename_B))
def main():
global best_prec1
batch_size = args.batch_size
split = args.split
num_workers = args.num_workers
scheduler_step = args.epochs
print('using split %d' % split)
if args.model == 'resnet34unet':
model = unet.UNet(3, 1, activation=None, dr=args.dr)
preprocess = get_preprocessing_fn('resnet34')
else:
raise ValueError(args.model, ' is not in our model')
if os.path.exists(os.path.join('ckpt', args.model, str(
args.split))) is False:
os.makedirs(os.path.join('ckpt', args.model, str(args.split)))
if os.path.exists(os.path.join('logdir', args.model, str(
args.split))) is False:
os.makedirs(os.path.join('logdir', args.model, str(args.split)))
train_aug = augmentation.get_augmentations('train', 1.0, args.image_size)
valid_aug = augmentation.get_augmentations('valid', 1.0, args.image_size)
train_dataset = SIIM.SIIM_ACR(
mode='train',
split=split,
preprocess=preprocess,
augmentation=train_aug,
)
valid_dataset = SIIM.SIIM_ACR(
mode='valid',
split=split,
preprocess=preprocess,
augmentation=valid_aug,
)
print('train valid dataset init successfully')
if args.use_total:
train_dataset = SIIM.SIIM_ACR(mode='train',
split=split,
preprocess=preprocess,
augmentation=train_aug,
use_total=True)
# weights = train_dataset.weight + 1
# train_sampler = WeightedRandomSampler(weights, num_samples=len(train_dataset))
train_dataloader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
pin_memory=True)
val_dataloader = DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
pin_memory=True)
model = nn.DataParallel(model, device_ids=args.gpus).cuda()
global metric
metric = metrics.FscoreMetric(threshold=0.5)
cudnn.benchmark = True
criterion_1 = loss.BCEDiceLoss().cuda()
# criterion_1 = nn.BCEWithLogitsLoss()
# criterion_1 = loss.WeightedBCELoss().cuda()
# criterion_1 = loss.DiceLoss().cuda()
criterion_2 = lovasz_losses.lovasz_hinge
if args.optim == 'sgd':
optimizer = torch.optim.SGD(model.parameters(),
args.learning_rate,
momentum=0.9,
weight_decay=1e-4)
elif args.optim == 'adam':
optimizer = torch.optim.Adam(model.parameters(), args.learning_rate)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, T_max=args.epochs, eta_min=args.min_lr)
if args.pretrained_model:
ckpt = torch.load(args.pretrained_model)
model.load_state_dict(ckpt)
print('use pretrained model on', args.pretrained_model)
i = 0
if args.use_total and args.resume:
if os.path.exists(os.path.join('ckpt', args.model)):
print('=> loading checkpoint {}'.format(
os.path.join('ckpt', args.model)))
ckpt = torch.load(
os.path.join('ckpt', args.model,
str(i) + '_size%d' % args.image_size + '.pth'))
model.load_state_dict(ckpt)
else:
print('=> no checkpoint found at {}'.format(
os.path.join('ckpt', args.model,
str(i) + '_size%d' % args.image_size + '.pth')))
elif args.resume:
if os.path.exists(os.path.join('ckpt', args.model, str(args.split))):
print('=> loading checkpoint {}'.format(
os.path.join('ckpt', args.model, str(args.split))))
ckpt = torch.load(
os.path.join('ckpt', args.model, str(args.split),
str(i) + '_size%d' % args.image_size + '.pth'))
model.load_state_dict(ckpt)
else:
print('=> no checkpoint found at {}'.format(
os.path.join('ckpt', args.model,
str(i) + '_size%d' % args.image_size + '.pth')))
for i in range(args.snapshot):
for epoch in range(args.epochs):
if args.lovasz is False:
train(train_dataloader, model, criterion_1, optimizer, epoch,
i)
else:
train(train_dataloader, model, criterion_2, optimizer, epoch,
i)
lr_scheduler.step()
if (epoch + 1) % args.val_freq == 0 or epoch == args.epochs - 1:
if args.lovasz is False:
prec1 = validate(val_dataloader, model, criterion_1, epoch,
i)
else:
prec1 = validate(val_dataloader, model, criterion_2, epoch,
i)
if prec1 > best_prec1:
best_prec1 = prec1
best_param = model.state_dict()
if args.use_total:
torch.save(
best_param,
os.path.join(
'ckpt', args.model,
str(i) + '_size%d' % args.image_size + '.pth'))
else:
torch.save(
best_param,
os.path.join(
'ckpt', args.model, str(args.split),
str(i) + '_size%d' % args.image_size + '.pth'))
optimizer = torch.optim.SGD(model.parameters(),
lr=args.learning_rate,
momentum=0.9,
weight_decay=1e-4)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, scheduler_step, args.min_lr)
best_prec1 = 0
def main():
global best_prec1
batch_size = args.batch_size
lr = args.learning_rate
epochs = args.epochs
val_freq = args.val_freq
split = args.split
num_workers = args.num_workers
if os.path.exists(os.path.join('logdir', args.model, str(
args.split))) is False:
os.makedirs(os.path.join('logdir', args.model, str(args.split)))
if args.model == 'resnet34unet':
model = unet.UNet(3, 1, activation='sigmoid', weight=None)
preprocess = get_preprocessing_fn('resnet34')
else:
raise ValueError(args.model, ' is not in our model')
valid_aug = augmentation.get_augmentations('valid', 1.0, args.image_size)
valid_dataset = SIIM_test.SIIM_ACR(mode='test',
preprocess=preprocess,
augmentation=valid_aug)
val_dataloader = DataLoader(valid_dataset,
batch_size=1,
shuffle=False,
num_workers=num_workers,
pin_memory=True)
model = nn.DataParallel(model, device_ids=args.gpus).cuda()
global metric
metric = metrics.FscoreMetric(threshold=None)
# cudnn.benchmark = True
for i in range(args.snapshot):
if args.pretrained_model:
ckpt = torch.load(args.pretrained_model)
model.load_state_dict(ckpt)
print('use pretrained model on', args.pretrained_model)
elif args.use_total:
print('=> loading checkpoint {}'.format(
os.path.join('ckpt', args.model)))
ckpt = torch.load(
os.path.join('ckpt', args.model,
str(i) + '_size%d' % args.image_size + '.pth'))
model.load_state_dict(ckpt)
elif os.path.exists(os.path.join('ckpt', args.model, str(args.split))):
print('=> loading checkpoint {}'.format(
os.path.join('ckpt', args.model, str(args.split))))
ckpt = torch.load(
os.path.join('ckpt', args.model, str(args.split),
str(i) + '_size%d' % args.image_size + '.pth'))
model.load_state_dict(ckpt)
else:
print('=> no checkpoint found at {}'.format(
os.path.join('ckpt', args.model)))
prec1 = test(val_dataloader, model, i)
import numpy as np
X_train = np.load('../data/DRIVE/584_584/X_train.npy')
Y_train = np.load('../data/DRIVE/584_584/Y_train.npy')
Y_test = np.load('../data/DRIVE/Y_test.npy')
data_gen_args = dict(rotation_range=0.2,
width_shift_range=0.05,
height_shift_range=0.05,
shear_range=0.05,
zoom_range=0.05,
horizontal_flip=True,
fill_mode='nearest')
myGene = trainGenerator(1, '../data/DRIVE/training', 'images', '1st_manual', data_gen_args, save_to_dir=None)
model = unet.UNet(input_shape=(584, 565, 3))
model.summary()
model_checkpoint = ModelCheckpoint('unet_membrane.hdf5', monitor='loss', verbose=1, save_best_only=True)
model.fit_generator(myGene, steps_per_epoch=300, epochs=1, callbacks=[model_checkpoint])
testGene = testGenerator("../data/DRIVE/test/")
results = model.predict_generator(testGene, 20, verbose=1)
print(results[0])
np.save('results.npy', results)
saveResult("../data/DRIVE/test/", results)
Y_test = np.array(Y_test, dtype="int").flatten()
results = np.array(results > 0, dtype="int").flatten()
F1 = f1_score(Y_test, results, labels=None, average='binary', sample_weight=None)
print(">> F1-Score = {:.2f}%".format(np.mean(F1 * 100)))
import scipy.io
from skimage.transform import resize
from scipy.spatial import distance
cuda_device = 'cuda:2'
device = torch.device(cuda_device if torch.cuda.is_available() else "cpu")
torch.cuda.set_device(device)
num_class = 37
H = 600
W = 480
model_folder = '600_480_pow_5_binary_unet_setup_1'
mat_name = 'mat500'
setup = model_folder[-1]
model = unet.UNet(1, num_class, [64, 128, 256, 512]).to(device)
# 저장된 model 로드
model.load_state_dict(
torch.load('../../saved_models/Laplace/setup' + setup + '/' +
model_folder + '/loss_4.5879849302392e-05_E_759.pth',
map_location=cuda_device)['model'])
model.eval()
###################
def angle(v1, v2):
v1 = np.array(v1)
v2 = np.array(v2)
v1 = v1 - [H / 2, W / 2]
v2 = v2 - [H / 2, W / 2]