def eval(test_ds,
model=MSI_FCN(),
ckpt_dir='./work_dir/msi_fcn/ckpt-*'
):
# checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
checkpoint = tf.train.Checkpoint(model=model)
if 'ckpt' not in ckpt_dir:
path=tf.train.latest_checkpoint(ckpt_dir)
else:
path=ckpt_dir
status=checkpoint.restore(path)
print("resotre model from {}".format(path))
Metric = Metrics()
n = 1
start = time.time()
for img_file, lab_file in test_ds:
img, lab = get_test_data(img_file, lab_file)
print("start inference {}th image".format(n))
pred = model(img)
m = Metric.update_state(lab, pred, is_train=False)
# print("p: {}, r: {}, IUcrack: {}".format( m['tp'], m['fp'], m['fn']))
n += 1
metric = Metric.overall_metrics()
for key, val in metric.items():
print("{}: {}".format(key, val))
end = time.time()
print("total time: ", end -start)
def main():
args = parse_args()
test_dir = os.path.join(args.root, args.test)
testannot_dir = os.path.join(args.root, args.testannot)
test_ds = get_filename_list(test_dir, testannot_dir)
# MSI_FCN
if args.model == 'msi_fcn':
model_config = {
"input_scales": 4,
"dcu_gr": 16,
"dense_gr": 24,
"filters": 64,
"expansion": 2,
"msc_filters": [2, 2, 2, 2],
"k": (7, 5, 3, 1),
"up_filters": 2,
"num_layers": (4, 4, 4, 4),
"num_classes": 2,
"use_msc": True,
"use_up_block": False
}
model = MSI_FCN(**model_config)
# FCN-VGG
elif args.model == 'fcn':
model_config = {"filters": 64, "expansion": 2, "num_classes": 2}
model = FCN_vgg16(**model_config)
# FCD
elif args.model == 'fcd':
model_config = {
"growth_rate": 12,
"td_filters": [48, 112, 192, 304, 464, 656, 896],
"up_filters": [1088, 816, 578, 384, 256],
"down_layers": [4, 4, 4, 4, 4, 4],
"up_layers": [4, 4, 4, 4, 4],
"num_classes": 2
}
model = FCD(**model_config)
else:
raise ValueError("args.model should be 'msi_fcn', 'fcn' or 'fcd'.")
ckpt = args.ckpt
assert ckpt is not None
# print model params
# model.build(input_shape=(None,256,256,3))
# print(model.summary())
for k, v in model_config.items():
print("{}: {}".format(k, v))
eval(test_ds, model, ckpt_dir=ckpt)
def main():
root = '/home/yel/yel/data/Aerialgoaf/detail/'
# root = '/home/yel/yel/data/DeepCrack-master/dataset/Deepcrack/'
img_dir = root + 'test'
label_dir = root + 'testannot'
save_dir = './visualization/msi_fcn_17/'
test_ds = get_filename_list(img_dir, label_dir)
# MSI_FCN
model_config = {
"input_scales": 4,
"dcu_gr": 16,
"dense_gr": 24,
"filters": 64,
"expansion": 2,
"msc_filters": [2, 2, 2, 2],
"k": (7, 5, 3, 1),
"up_filters": 2,
"num_layers": (4, 4, 4, 4),
"num_classes": 2,
"use_msc": True,
"use_up_block": False
}
# FCN_VGG16
# model_config = {"filters":64,"expansion":2,"num_classes":2}
ckpt_dir = './work_dir/msi_fcn_4scales/'
ckpt_name = 'ckpt-17'
model = MSI_FCN(**model_config, display_stages=True)
# model = FCN_vgg16(**model_config)
checkpoint = tf.train.Checkpoint(model=model)
if ckpt_name:
path = os.path.join(ckpt_dir, ckpt_name)
else:
path = tf.train.latest_checkpoint(ckpt_dir)
status = checkpoint.restore(path)
if path is not None:
print("resotre model from {}".format(path))
n = 1
if not os.path.exists(save_dir):
os.makedirs(save_dir)
for img_file, lab_file in test_ds:
img, lab = get_test_data(img_file, lab_file)
filename = img_file.split('/')[-1]
print("inference {}th image:".format(n))
pred = model(img)
show_all_branch(pred, save_dir, filename, rgb=True)
n += 1
def main():
# root = 'D:\AerialGoaf\detail\\512x512\\'
root = "D:/data/detail/"
# root = '/home/yel/yel/data/DeepCrack-master/dataset/DeepCrack/'
# root = '/home/yel/yel/data/road_crack/'
img_dir = root + 'test'
label_dir = root + 'testannot'
test_ds = get_filename_list(img_dir, label_dir)
# MSI_FCN
model_config = {
"input_scales": 4,
"dcu_gr": 16,
"dense_gr": 24,
"filters": 64,
"expansion": 2,
"msc_filters": [2, 2, 2, 2],
"k": (7, 5, 3, 1),
"up_filters": 2,
"num_layers": (4, 4, 4, 4),
"num_classes": 2,
"use_msc": True,
"use_up_block": False
}
# FCN_VGG16
# model_config = {"filters": 64, "expansion": 2, "num_classes": 2}
# FCD
# model_config = {"growth_rate": 12, "td_filters": [48, 112, 192, 304, 464, 656, 896],
# "up_filters": [1088, 816, 578, 384, 256], "down_layers": [4, 4, 4, 4, 4, 4],
# "up_layers": [4, 4, 4, 4, 4], "num_classes": 2}
ckpt_dir = '../work_dir/msi_fcn_4scales'
# model = FCD(**model_config)
model = MSI_FCN(**model_config)
# model = FCN_vgg16(**model_config)
for k, v in model_config.items():
print("{}: {}".format(k, v))
eval(test_ds, model, ckpt_dir=ckpt_dir, ckpt_name='ckpt-11017')
def main():
root = '/home/yel/yel/data/Aerialgoaf/detail/'
# root = '/home/yel/yel/data/DeepCrack-master/dataset/Deepcrack/'
img_dir = root + 'train'
label_dir = root + 'trainannot'
val_dir = root + 'val'
vallabel_dir = root + 'valannot'
train_ds = get_dataset(img_dir, label_dir, batch_size=5)
val_ds = get_dataset(val_dir, vallabel_dir, batch_size=5)
model = MSI_FCN()
lr = tf.keras.optimizers.schedules.ExponentialDecay(2e-4, 10000, 0.1)
optimizer = tf.keras.optimizers.Adam(lr, beta_1=0.5)
fit(train_ds=train_ds,
val_ds=val_ds,
model=model,
optimizer=optimizer,
loss_func=WSCE,
work_dir='../work_dir/msi_fcn_2',
epochs=100,
fine_tune=True)
image = Image.open(img_path)
image = np.array(image)
h, w, c = image.shape
patch = image[:512, w - 512:, :]
img_in = tf.convert_to_tensor(patch, dtype=tf.float32)
img_in = tf.image.resize(img_in, (256, 256))
img_in = img_in / 255
img_in = tf.expand_dims(img_in, 0)
debug = 1
# MSI_FCN
model_config = {
"input_scales": 4,
"dcu_gr": 16,
"dense_gr": 24,
"filters": 64,
"expansion": 2,
"msc_filters": [2, 2, 2, 2],
"k": (7, 5, 3, 1),
"up_filters": 2,
"num_layers": (4, 4, 4, 4),
"num_classes": 2,
"use_msc": True,
"use_up_block": False
}
work_dir = './work_dir/msi_fcn_4scales/weights/ckpt-17'
model = MSI_FCN(**model_config)
checkpoint = tf.train.Checkpoint(model=model)
checkpoint.restore(work_dir)
pred = model(img_in)
writeImage(pred, "pred.png")
import tensorflow as tf
import time
import os
from model.msi_fcn import MSI_FCN
from core.data import get_dataset
from core.loss import WSCE
from core.metrics import show_metrics
import datetime
root = '/home/yel/yel/data/Aerialgoaf/detail/'
img_dir = root + 'train'
label_dir = root + 'trainannot'
train_ds = get_dataset(img_dir, label_dir, batch_size=3)
model = MSI_FCN()
optimizer = tf.keras.optimizers.Adam(2e-4, beta_1=0.5)
checkpoint_dir = './training_checkpoints'
# checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
checkpoint = tf.train.Checkpoint(optimizer=optimizer, model=model)
ckpt_manager = tf.train.CheckpointManager(checkpoint, checkpoint_dir, max_to_keep=5)
log_dir = './logs/'
summary_writer = tf.summary.create_file_writer(
log_dir + "fit/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
# @tf.function
def train_step(model, input, label, loss_object, optimizer, show_metrics, summary_writer, step):
with tf.GradientTape() as t:
output = model(input, training=True)
loss = loss_object(output, label)
def fit(train_ds,
val_ds,
model=MSI_FCN(),
loss_func=WSCE,
optimizer=tf.keras.optimizers.Adam(2e-4, beta_1=0.5),
Metricor=Metrics(),
work_dir='./work_dir/msi_fcn',
epochs=100,
fine_tune=False):
# checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
checkpoint = tf.train.Checkpoint(optimizer=optimizer, model=model)
ckpt_manager = tf.train.CheckpointManager(checkpoint,
work_dir,
max_to_keep=5)
summary_writer = tf.summary.create_file_writer(work_dir)
# n = 22000
n = 0
if fine_tune:
path = tf.train.latest_checkpoint(work_dir)
# path ="/home/yel/yel/Pyproject/MSI_FCN/work_dir/msi_fcn_5/ckpt-15000"
n = int(path.split('-')[1])
checkpoint.restore(path)
print("restore from: {}".format(path))
for epoch in range(epochs):
start = time.time()
# Train
for inputs, label in train_ds:
output, loss = train_step(inputs, label, model, loss_func,
optimizer)
metrics = Metricor.update_state(label, output, is_train=True)
# metrics = Metricor(label,output)
write_sumamry(summary_writer, loss, metrics, step=n, eval=False)
if (n + 1) % 10 == 0:
print_summary(metrics, loss, n, epoch, val=False)
if (n + 1) % 100 == 0:
if val_ds is not None:
for val_inputs, val_label in val_ds.take(1):
output, loss = train_step(val_inputs, val_label, model,
loss_func, optimizer)
metrics = Metricor.update_state(val_label,
output,
is_train=True)
# metrics = Metricor(label,output)
write_sumamry(summary_writer,
loss,
metrics,
step=n,
eval=True)
print_summary(metrics, loss, n, epoch, val=True)
n += 1
# saving (checkpoint) the model every 20 epochs
if (epoch + 1) % 10 == 0:
ckpt_manager.save(checkpoint_number=n)
consume_time = time.time() - start
print('Time taken for epoch {} is {:2f} sec, eta:{} \n'.format(
epoch + 1, consume_time,
datetime.timedelta(seconds=(epochs - epoch - 1) * consume_time)))
ckpt_manager.save(checkpoint_number=n)
def main():
args = parse_args()
train_dir = os.path.join(args.root, args.train)
trainannot_dir = os.path.join(args.root, args.trainannot)
val_dir = os.path.join(args.root, args.val)
valannot_dir = os.path.join(args.root, args.valannot)
train_ds = get_dataset(train_dir, trainannot_dir, batch_size=4)
val_ds = get_dataset(val_dir, valannot_dir, batch_size=4)
# val_ds = None
fine_tune = args.finetune
# fine_tune=False
# MSI_FCN
if args.model == 'msi_fcn':
model_config = {
"input_scales": 4,
"dcu_gr": 16,
"dense_gr": 24,
"filters": 64,
"expansion": 2,
"msc_filters": [2, 2, 2, 2],
"k": (7, 5, 3, 1),
"up_filters": 2,
"num_layers": (4, 4, 4, 4),
"num_classes": 2,
"use_msc": True,
"use_up_block": False
}
model = MSI_FCN(**model_config)
# FCN-VGG
elif args.model == 'fcn':
model_config = {"filters": 64, "expansion": 2, "num_classes": 2}
model = FCN_vgg16(**model_config)
# FCD
elif args.model == 'fcd':
model_config = {
"growth_rate": 12,
"td_filters": [48, 112, 192, 304, 464, 656, 896],
"up_filters": [1088, 816, 578, 384, 256],
"down_layers": [4, 4, 4, 4, 4, 4],
"up_layers": [4, 4, 4, 4, 4],
"num_classes": 2
}
model = FCD(**model_config)
else:
raise ValueError("args.model should be 'msi_fcn', 'fcn' or 'fcd'.")
work_dir = args.work_dir
# print model params
# model.build(input_shape=(None,256,256,3))
# print(model.summary())
lr = tf.keras.optimizers.schedules.ExponentialDecay(2e-4, 5000, 0.95)
optimizer = tf.keras.optimizers.Adam(lr)
for k, v in model_config.items():
print("{}: {}".format(k, v))
fit(train_ds=train_ds,
val_ds=val_ds,
model=model,
optimizer=optimizer,
loss_func=WSCE,
work_dir=work_dir,
epochs=60,
fine_tune=fine_tune)