def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-g', '--gpu', type=int, required=True, help='GPU id')
args = parser.parse_args()
gpu = args.gpu
# 0. config
cmd = 'git log -n1 --format="%h"'
vcs_version = subprocess.check_output(cmd, shell=True).strip()
timestamp = datetime.datetime.now().strftime('%Y%m%d-%H%M%S')
out = 'fcn8s_atonce_VCS-%s_TIME-%s' % (
vcs_version,
timestamp,
)
out = osp.join(here, 'logs', out)
# 1. dataset
dataset_train = datasets.BridgeSeg(split='train',
rcrop=[400, 400],
use_class_weight=False)
dataset_train_nocrop = datasets.BridgeSeg(split='train',
use_class_weight=False)
dataset_valid = datasets.BridgeSeg(split='validation',
use_class_weight=False)
if dataset_train.class_weight is not None:
print("Using class weigths: ", dataset_train.class_weight)
iter_train = chainer.iterators.MultiprocessIterator(dataset_train,
batch_size=1,
shared_mem=10**7)
iter_valid = chainer.iterators.MultiprocessIterator(dataset_valid,
batch_size=1,
shared_mem=10**7,
repeat=False,
shuffle=False)
# 2. model
n_class = len(dataset_train.class_names)
class_weight = dataset_train.class_weight
vgg = fcn.models.VGG16()
chainer.serializers.load_npz(vgg.pretrained_model, vgg)
model = fcn.models.FCN8sAtOnce(n_class=n_class, class_weight=class_weight)
model.init_from_vgg16(vgg)
if gpu >= 0:
cuda.get_device(gpu).use()
model.to_gpu()
# 3. optimizer
optimizer = chainer.optimizers.MomentumSGD(lr=1.0e-10, momentum=0.99)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(rate=0.0005))
for p in model.params():
if p.name == 'b':
p.update_rule = chainer.optimizers.momentum_sgd.MomentumSGDRule(
lr=optimizer.lr * 2, momentum=0)
model.upscore2.disable_update()
model.upscore8.disable_update()
model.upscore_pool4.disable_update()
# training loop
trainer = fcn.Trainer(
device=gpu,
model=model,
optimizer=optimizer,
iter_train=iter_train,
iter_valid=iter_valid,
out=out,
max_iter=100000,
)
trainer.train()
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-g', '--gpu', type=int, required=True, help='GPU id')
parser.add_argument('-da', '--data-augmentation', type=int, \
default=0, choices=(0,1),
help='Data augmentation flag. Default 0, 1 for data augmentation')
parser.add_argument('-d', '--deck-mask', type=int, default=1, choices=(0,1),\
help='Applying deck mask. Default 1, 0 for not masking deck')
parser.add_argument('-e', '--epochs', type=int, default=100, choices=range(1000), \
help='Number of epochs', metavar='range(0...1000)')
args = parser.parse_args()
gpu = args.gpu
# 0. config
cmd = 'git log -n1 --format="%h"'
vcs_version = subprocess.check_output(cmd, shell=True).strip()
timestamp = datetime.datetime.now().strftime('%Y%m%d-%H%M%S')
out = 'fcn32s_VCS-%s_TIME-%s' % (
vcs_version,
timestamp,
)
out = osp.join(here, 'logs', out)
# 1. dataset
deck_flag = bool(args.deck_mask)
data_augmentation = bool(args.data_augmentation)
class_weight_flag = False
train_dataset = datasets.BridgeSeg(split='train',
rcrop=[512, 512],
use_class_weight=class_weight_flag,
black_out_non_deck=deck_flag,
use_data_augmentation=data_augmentation)
train_dataset_nocrop = datasets.BridgeSeg(
split='train',
use_class_weight=class_weight_flag,
black_out_non_deck=deck_flag,
use_data_augmentation=False)
test_dataset = datasets.BridgeSeg(split='validation',
use_class_weight=class_weight_flag,
black_out_non_deck=deck_flag,
use_data_augmentation=False)
iter_train = chainer.iterators.MultiprocessIterator(train_dataset,
batch_size=1,
shared_mem=10**8)
iter_train_nocrop = chainer.iterators.MultiprocessIterator(
train_dataset_nocrop,
batch_size=1,
shared_mem=10**8,
repeat=False,
shuffle=False)
iter_valid = chainer.iterators.MultiprocessIterator(test_dataset,
batch_size=1,
shared_mem=10**8,
repeat=False,
shuffle=False)
train_samples = len(train_dataset)
nbepochs = args.epochs
# 2. model
n_class = len(train_dataset.class_names)
class_weight = train_dataset.class_weight
vgg = fcn.models.VGG16()
chainer.serializers.load_npz(vgg.pretrained_model, vgg)
model = fcn.models.FCN32s(n_class=n_class, class_weight=class_weight)
model.init_from_vgg16(vgg)
if gpu >= 0:
cuda.get_device(gpu).use()
model.to_gpu()
# 3. optimizer
optimizer = chainer.optimizers.MomentumSGD(lr=1.0e-10, momentum=0.99)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(rate=0.0005))
for p in model.params():
if p.name == 'b':
p.update_rule = chainer.optimizers.momentum_sgd.MomentumSGDRule(
lr=optimizer.lr * 2, momentum=0)
model.upscore.disable_update()
# training loop
# pdb.set_trace()
trainer = fcn.Trainer(device=gpu,
model=model,
optimizer=optimizer,
iter_train=iter_train,
iter_train_noncrop=iter_train_nocrop,
iter_valid=iter_valid,
out=out,
max_iter=train_samples * nbepochs,
interval_validate=train_samples)
trainer.train(fold=0)