def main():
''' Main function '''
if cfg.SYSTEM.UPDATE_CFG:
cfg.merge_from_file(cfg.SYSTEM.CFG_FILE)
cfg.freeze() #冻结参数
print(cfg)
vis.log('configs:\n {}'.format(cfg.clone()))
#LOAD DATA
training_data, validation_data = prepare_stanford40(cfg.clone())
#training_data, validation_data = prepare_pascal(cfg.clone())
#LOAD MODEL
model = resnet34(pretrained=False)
model = model.cuda()
model.load_state_dict(t.load(cfg.MODEL.PRETRAINED))
#------------- optim -----------------#
optimizer = optim.Adam(
model.parameters(),
lr=cfg.OPTIM.LR,
)
scheduler = ReduceLROnPlateau(optimizer,
mode='min',
factor=cfg.OPTIM.LR_DECAY,
patience=1,
verbose=True,
threshold=1e-4,
threshold_mode='rel',
cooldown=1,
min_lr=0.00005,
eps=1e-8)
#------------- train ----------------#
train(model, training_data, validation_data, optimizer, scheduler,
cfg.clone())
def main():
if cfg.SYSTEM.UPDATE_CFG:
cfg.merge_from_file(cfg.SYSTEM.CFG_FILE)
cfg.freeze() #冻结参数
vis_env.log('configs:\n {}'.format(cfg.clone()))
#load data
traindata, valdata = stanford40(cfg)
#load model
net = model.multi_two(True)
net = net.cuda()
optimizer = t.optim.Adagrad(net.parameters(), lr=cfg.OPTIM.LR)
lr_strategy = t.optim.lr_scheduler.CosineAnnealingLR(
optimizer,
300,
)
lossfunc = t.nn.CrossEntropyLoss()
train(net, traindata, valdata, optimizer, lossfunc, lr_strategy)
def main():
if cfg.SYSTEM.UPDATE_CFG:
cfg.merge_from_file(cfg.SYSTEM.CFG_FILE)
cfg.freeze() #冻结参数
vis_env.log('configs:\n {}'.format(cfg.clone()))
#load data
traindata, valdata = stanford40(cfg)
#load model
net = model.twokey_copy(True) #getattr(model,cfg.MODEL.NAME)
#for param in net.layer1.parameters():
# param.requires_grad = False
net = net.cuda()
optimizer = t.optim.Adagrad(filter(lambda p: p.requires_grad,
net.parameters()),
lr=cfg.OPTIM.LR)
lr_strategy = t.optim.lr_scheduler.CosineAnnealingLR(
optimizer,
50,
)
lossfunc = t.nn.SmoothL1Loss()
train(net, traindata, valdata, optimizer, lossfunc, lr_strategy)
def main():
''' Main function '''
if cfg.SYSTEM.UPDATE_CFG:
cfg.merge_from_file(cfg.SYSTEM.CFG_FILE)
cfg.freeze() #冻结参数
print(cfg)
vis.log('configs:\n {}'.format(cfg.clone()))
#LOAD DATA
##training_data, validation_data = prepare_stanford40(cfg.clone())
#training_data, validation_data = prepare_pascal(cfg.clone())
#normalize = transforms.Normalize(mean=cfg.mean, std=cfg.std)
img_norm_cfg = dict(mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True)
val_dataset = StanfordDataset(
ann_file=
'/home/share/LabServer/DATASET/stanford40/ImageSplits/test.txt',
img_prefix='/home/share/LabServer/DATASET/stanford40/',
img_scale=(224, 224),
img_norm_cfg=img_norm_cfg,
size_divisor=32,
proposal_file='/home/share/LabServer/GLnet/stanford_test_bbox_new.pkl',
test_mode=False,
)
train_dataset = StanfordDataset(
ann_file=
'/home/share/LabServer/DATASET/stanford40/ImageSplits/train.txt',
img_prefix='/home/share/LabServer/DATASET/stanford40/',
img_scale=(224, 224),
img_norm_cfg=img_norm_cfg,
size_divisor=32,
proposal_file='/home/share/LabServer/GLnet/stanford_train_bbox_new.pkl',
test_mode=False,
)
train_loader = DataLoader(train_dataset,
batch_size=10,
shuffle=True,
sampler=None,
collate_fn=partial(collate, samples_per_gpu=2),
num_workers=4)
val_loader = DataLoader(val_dataset,
batch_size=10,
shuffle=False,
sampler=None,
collate_fn=partial(collate, samples_per_gpu=2),
num_workers=4)
#LOAD MODEL
model = proposalModel()
model = model.to("cuda:0")
#submodel = resnet50(True)
#submodel = submodel.to("cuda:1")
#------------- optim -----------------#
optimizer = optim.Adam(
model.parameters(),
lr=cfg.OPTIM.LR,
)
scheduler = ReduceLROnPlateau(optimizer,
mode='min',
factor=cfg.OPTIM.LR_DECAY,
patience=1,
verbose=True,
threshold=1e-4,
threshold_mode='rel',
cooldown=1,
min_lr=0.00005,
eps=1e-8)
#------------- train ----------------#
train(model, train_loader, val_loader, optimizer, scheduler, cfg.clone())