def __init__(self, pretrained=True, num_classes=212, ckpt=None, **kwargs):
super(Dense201, self).__init__(num_init_features=64,
growth_rate=32,
block_config=(6, 12, 48, 32),
num_classes=num_classes,
**kwargs)
if pretrained:
pretrained_dict = dict(
model_zoo.load_url(densenet.model_urls['densenet201']))
print(pretrained_dict.keys())
del pretrained_dict['classifier.weight']
del pretrained_dict['classifier.bias']
model_dict = self.state_dict()
pattern = re.compile(
r'^(.*denselayer\d+\.(?:norm|relu|conv))\.((?:[12])\.(?:weight|bias|running_mean|running_var))$'
)
for key in list(pretrained_dict.keys()):
res = pattern.match(key)
if res:
new_key = res.group(1) + res.group(2)
pretrained_dict[new_key] = pretrained_dict[key]
del pretrained_dict[key]
model_dict.update(pretrained_dict)
self.load_state_dict(model_dict)
self.sigmoid = nn.Sigmoid()
if not ckpt is None:
saver = Saver(ckpt, 'model')
saver.load_last_ckpt(self)
#PATH = './ckpt'
a = AlignDataset(
'/data/icme/data/picture',
'/data/icme/data/landmark',
'/data/icme/data/pred_landmark',
'/data/icme/valid',
Align(
'cache/mean_landmarks.pkl',
(224, 224),
(0.15, 0.1),
), # idx=list(range(51, 66))),
phase='eval',
# ldmk_ids=list(range(51, 66))
)
#Saver.dir=PATH
saver = Saver('backup', 'model')
current = None
net.eval()
epoch_size = len(a)
metrics = Metrics().add_nme().add_auc()
model_name = 'align-jitter227-model-120000.pth'
saver.load(net, model_name)
all_pr = []
all_gt = []
save_dir = '/data/icme/data/pred_landmark_align'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
batch_iterator = iter(DataLoader(a, batch_size=1, shuffle=False,
num_workers=4))
# '/data/icme/data/landmark',
# '/data/icme/train',
# Align('./cache/mean_landmarks.pkl', (224, 224), (0.2, 0.1),
# ),# idx=list(range(51, 66))),
# flip=True,
# max_jitter=3,
# max_radian=0
# # ldmk_ids=list(range(51, 66))
# )
batch_iterator = iter(DataLoader(a, batch_size=cfg.batch_size, shuffle=True, num_workers=4))
criterion = loss.get_criterion(cfg.loss)
# criterion = WingLoss(10, 2)
optimizer = optim.Adam(net.parameters(), lr=0, weight_decay=cfg.weight_decay)
# optimizer = optim.SGD(net.parameters(), lr=0, weight_decay=cfg.weight_decay, momentum=0.9)
saver = Saver(os.path.join(cfg.root, 'ckpt'), 'model', 10)
last = saver.last_ckpt()
start_iter = 0 if last is None else int(last.split('.')[0].split('-')[-1])
if start_iter > 0:
saver.load(net, last)
lr_gen.set_global_step(start_iter)
running_loss = 0.0
batch_size = cfg.batch_size
epoch_size = len(a) // batch_size
epoch = start_iter // epoch_size + 1
for iteration in range(start_iter, cfg.max_iter + 1):
if iteration % epoch_size == 0:
# create batch iterator
batch_iterator = iter(DataLoader(a, batch_size,
shuffle=True, num_workers=4))
(0.15, 0.1),
), # idx=list(range(51, 66))),
shape=(256, 256),
flip=True,
# ldmk_ids=list(range(51, 66))
)
batch_iterator = iter(
DataLoader(a, batch_size=args.batch_size, shuffle=True, num_workers=4))
criterion = WingLoss(10, 2)
# optimizer = optim.Adam(net.parameters(), lr=args.lr, weight_decay=args.weight_decay)
optimizer = optim.SGD(net.parameters(),
lr=args.lr,
weight_decay=args.weight_decay,
momentum=0.9)
saver = Saver('../ckpt', 'model', 10)
last = saver.last_ckpt()
start_iter = 0 if last is None else int(last.split('.')[0].split('-')[-1])
if start_iter > 0:
saver.load(net, last)
running_loss = 0.0
batch_size = args.batch_size
epoch_size = len(a) // batch_size
epoch = start_iter // epoch_size + 1
for iteration in range(start_iter, 300001):
if iteration % epoch_size == 0:
# create batch iterator
batch_iterator = iter(
DataLoader(a, batch_size, shuffle=True, num_workers=4))
epoch += 1
'/data/icme/data/picture',
'/data/icme/data/landmark',
'/data/icme/data/pred_landmark',
'/data/icme/valid',
Align(
'../cache/mean_landmarks.pkl',
(256, 256),
(0.15, 0.1),
), # idx=list(range(51, 66))),
phase='eval',
shape=(256, 256),
# ldmk_ids=list(range(51, 66))
)
batch_iterator = iter(DataLoader(a, batch_size=4, shuffle=True, num_workers=4))
#Saver.dir=PATH
saver = Saver('../ckpt', 'model')
current = None
net.eval()
batch_size = 4
epoch_size = len(a) // batch_size
writer = SummaryWriter('logs/BA/valid')
metrics = Metrics().add_nme(0.9).add_auc(decay=0.9).add_loss(decay=0.9)
while True:
if current == saver.last_ckpt():
time.sleep(1)
else:
last = saver.last_ckpt()
current_iter = 0 if last is None else int(
last.split('.')[0].split('-')[-1])
while True:
try:
a = BBoxDataset('/data/icme/data/picture',
'/data/icme/data/landmark',
'/data/icme/bbox',
'/data/icme/valid', phase='eval')
# a = AlignDataset('/data/icme/data/picture',
# '/data/icme/data/landmark',
# '/data/icme/data/pred_landmark',
# '/data/icme/valid',
# Align('../cache/mean_landmarks.pkl', (224, 224), (0.2, 0.1),
# ), # idx=list(range(51, 66))),
# phase='eval',
# # ldmk_ids=list(range(51, 66))
# )
batch_iterator = iter(DataLoader(a, batch_size=4, shuffle=True, num_workers=4))
#Saver.dir=PATH
saver = Saver(os.path.join(cfg.root, 'ckpt'), 'model')
saver2 = Saver(os.path.join(cfg.root, 'snapshot'), 'model', max_keep=10)
current = None
net.eval()
batch_size = 4
epoch_size = len(a) // batch_size
writer = SummaryWriter(os.path.join(cfg.root, 'logs/valid'))
metrics = Metrics().add_nme(0.9).add_auc(decay=0.9).add_loss(decay=0.9)
while True:
if current == saver.last_ckpt():
time.sleep(1)
else:
last = saver.last_ckpt()
current_iter = 0 if last is None else int(last.split('.')[0].split('-')[-1])
while True:
try: