def train_sfpn_small_balance(depth=50, roiSize=224):
train_dataset = sdataset.SmallPatchDataset("train", roiSize, balance=True)
val_dataset = sdataset.SmallPatchDataset("val", roiSize)
test_dataset = sdataset.SmallPatchDataset("test", roiSize)
from sfpn import SFPN
model = nn.DataParallel(SFPN().cuda())
dcfg = {"bsize": 192, "nworker": 20, "collate": default_collate}
model_name = "sp%d_sfpn%d_small_balance" % (roiSize, depth)
train(model, model_name, train_dataset, val_dataset, test_dataset, dcfg)
def train_srdn_balance(roiSize=32):
train_dataset = sdataset.SmallPatchDataset("train", roiSize, balance=True)
val_dataset = sdataset.SmallPatchDataset("val", roiSize)
test_dataset = sdataset.SmallPatchDataset("test", roiSize)
from srdn import RDN
# model = nn.DataParallel(
# RDN(g0=32, d=4, c=6, k=16, roiSize=roiSize).cuda())
model = nn.DataParallel(RDN(g0=16, d=3, c=4, k=16, roiSize=roiSize).cuda())
dcfg = {"bsize": 4096, "nworker": 20, "collate": default_collate}
model_name = "sp%d_rdn_small_balance_d3c4k16" % roiSize
train(model, model_name, train_dataset, val_dataset, test_dataset, dcfg)
def train_sfpn_small_aug(depth=50, roiSize=224):
top = torchvision.transforms.ToPILImage()
hf = torchvision.transforms.RandomHorizontalFlip()
vf = torchvision.transforms.RandomVerticalFlip()
rot = torchvision.transforms.RandomRotation(30)
size = torchvision.transforms.Resize((roiSize, roiSize))
tot = torchvision.transforms.ToTensor()
trfm = torchvision.transforms.Compose([top, hf, vf, rot, size, tot])
train_dataset = sdataset.SmallPatchDataset("train",
roiSize=roiSize,
transform=trfm)
val_dataset = sdataset.SmallPatchDataset("val", roiSize=roiSize)
test_dataset = sdataset.SmallPatchDataset("test", roiSize=roiSize)
from sfpn import SFPN
model = nn.DataParallel(SFPN().cuda())
# dcfg = {"bsize": 192, "nworker": 20, "collate": default_collate}
dcfg = {"bsize": 128, "nworker": 20, "collate": default_collate}
model_name = "sp%d_sfpn%d_small_aug" % (roiSize, depth)
train(model, model_name, train_dataset, val_dataset, test_dataset, dcfg)