def train_sequence_si():
ks = [10, 20, 30]
fecs = [[0, 1, 2, 3, 4, 5], [0, 1, 2, 3, 4, 5], [0, 1, 2, 3, 4, 5]]
logdir = 'modeldir/stage_all/seq_si/'
from loss import FECLoss
for i, k in enumerate(ks):
for fec in fecs[i]:
cfg = _allrun_config_si(k)
cfg['lr'] = 0.00001
cfg['batch'] = 64
cfg['epochs'] = 200
cfg['scheduler'] = True
cfg['patience'] = 30
cfg['step'] = 0
if fec % 2 == 1:
continue
if fec == 0:
cfg['criterion'] = torch.nn.BCELoss()
cfg['model'] = 'resnet18b4_si_k%d' % (k)
else:
cfg['criterion'] = FECLoss(alpha=cfg['batch'] * fec)
cfg['model'] = 'resnet18b4_si_k%d_fec%d' % (k, fec)
cfg['model_dir'] = '%s/%s' % (logdir, cfg['model'])
model = nn.DataParallel(sinet.SiNet(nblock=4, k=k).cuda())
train.run_train(model, cfg)
def train_smallnet_si_boost_fec(s=2, k=10):
cfg = util.default_cfg()
train_dataset = dataset.StratifySIDataset(mode='train', stage=s, k=k)
val_dataset = dataset.StratifySIDataset(mode='val', stage=s, k=k)
test_dataset = dataset.StratifySIDataset(mode='test', stage=s, k=k)
cfg['train'] = train_dataset
cfg['val'] = val_dataset
cfg['test'] = test_dataset
cfg['batch'] = 64
cfg['epochs'] = 500
cfg['scheduler'] = True
cfg['decay'] = 0.01
cfg['lr'] = 0.0001
cfg['patience'] = 20
cfg['collate'] = default_collate
cfg['instance'] = _train_si
from loss import FECLoss
cfg['criterion'] = FECLoss(alpha=64, reduction='none')
cfg['model'] = 'smallnet_si_k%d_boost_fec1' % (k)
cfg['model_dir'] = 'modeldir/stage%d/smallnet_si_k%d_boost_fec1' % (s, k)
model_pth = os.path.join(cfg['model_dir'], 'model.pth')
model = nn.DataParallel(sinet.SmallNet(k=k).cuda())
if os.path.exists(model_pth):
ckp = torch.load(model_pth)
model.load_state_dict(ckp['model'])
cfg['step'] = ckp['epoch'] + 1
print("load pretrained model", model_pth, "start epoch:", cfg['step'])
run_train(model, cfg)
def train_smallnet_stratify_pj_fecq(s=2, k=10):
cfg = util.default_cfg()
cfg = train._config_stratify_pj_dataset(cfg, s, k)
model = nn.DataParallel(sinet.SmallNet(k=k).cuda())
from loss import FECLoss
cfg['criterion'] = FECLoss(alpha=8)
cfg['model'] = 'smallnet_pj_k%d_fec0.25' % (k)
cfg['model_dir'] = 'modeldir/stage%d/smallnet_pj_k%d_fec0.25' % (s, k)
cfg = train._train_config_pj(model, cfg)
cfg['scheduler'] = False
cfg['lr'] = 0.0001
cfg['epochs'] = 1000
train.run_train(model, cfg)
def train_tinynet_stratify_si(s=2, k=10):
cfg = util.default_cfg()
cfg = train._config_stratify_si_dataset(cfg, s, k)
from loss import FECLoss
cfg['criterion'] = FECLoss(alpha=32)
cfg['model'] = 'tinynet_si_k%d_fec0.5' % (k)
cfg['model_dir'] = 'modeldir/stage%d/tinynet_si_k%d_fec0.5' % (s, k)
cfg['collate'] = default_collate
cfg['instance'] = train._train_si
model = nn.DataParallel(sinet.TinyNet(k=k).cuda())
cfg = train._train_config_si(model, cfg)
train.run_train(model, cfg)
def train_resnet_si(k=10):
cfg = _allrun_config_si(k)
from loss import FECLoss
cfg['criterion'] = FECLoss(alpha=64)
model = nn.DataParallel(sinet.SiNet(nblock=4, k=k).cuda())
cfg['model'] = 'resnet18b4_si_k%d_fec1' % (k)
cfg['model_dir'] = 'modeldir/stage_all/resnet18b4_si_k%d_fec1' % (k)
cfg['lr'] = 0.0001
model_pth = os.path.join(cfg['model_dir'], 'model.pth')
if os.path.exists(model_pth):
ckp = torch.load(model_pth)
model.load_state_dict(ckp['model'])
cfg['step'] = ckp['epoch'] + 1
print("load pretrained model", model_pth, "start epoch:", cfg['step'])
train.run_train(model, cfg)