val_loader = Loader(val_path,
arg.batch_size,
transform=preprocess,
sampler=arg.sampler,
torch_type=arg.dtype,
cpus=arg.cpus,
shuffle=False,
drop_last=False)
test_loader = Loader(test_path,
1,
torch_type=arg.dtype,
cpus=arg.cpus,
shuffle=False,
drop_last=False)
norm_layer = nn.BatchNorm2d
act = nn.ReLU
if arg.model == "unet":
net = Unet2D(feature_scale=arg.feature_scale, act=act)
elif arg.model == "unetres":
net = UnetRes2D(1, nn.InstanceNorm2d, is_pool=arg.pool)
net = nn.DataParallel(net).to(torch_device)
recon_loss = nn.BCEWithLogitsLoss()
model = CNNTrainer(arg, net, torch_device, recon_loss=recon_loss)
if arg.test is False:
model.train(train_loader, val_loader)
model.test(test_loader, val_loader)
if arg.model == "fusion":
net = Fusionnet(arg.in_channel, arg.out_channel, arg.ngf, arg.clamp)
elif arg.model == "unet":
net = Unet2D(feature_scale=arg.feature_scale)
elif arg.model == "unet_sh":
net = UnetSH2D(arg.sh_size, feature_scale=arg.feature_scale)
else:
raise NotImplementedError("Not Implemented Model")
net = nn.DataParallel(net).to(torch_device)
if arg.loss == "l2":
recon_loss = nn.L2Loss()
elif arg.loss == "l1":
recon_loss = nn.L1Loss()
recon_loss = EdgeWeightedLoss(1, 10)
#recon_loss=nn.BCELoss()
model = CNNTrainer(arg,
net,
torch_device,
recon_loss=recon_loss,
logger=logger)
model.load(filename="epoch[0123]_losssum[0.005018].pth.tar")
if arg.test == 0:
model.train(train_loader, test_loader)
if arg.test == 1:
model.test(test_loader)
# utils.slack_alarm("zsef123", "Model %s Done"%(arg.save_dir))
shuffle=False,
drop_last=False)
test_loader = Loader(test_path,
1,
torch_type=arg.dtype,
cpus=arg.cpus,
shuffle=False,
drop_last=False)
norm_layer = nn.BatchNorm2d
act = nn.ReLU
if arg.model == "unet":
net = Unet2D(feature_scale=arg.feature_scale, act=act)
elif arg.model == "unetres":
net = UnetRes2D(1, nn.InstanceNorm2d, is_pool=arg.pool)
elif arg.model == "unetbr":
net = UnetBR2D(1, nn.InstanceNorm2d, is_pool=arg.pool)
elif arg.model == "vnet":
net = VNet(elu=False)
net = nn.DataParallel(net).to(torch_device)
recon_loss = nn.BCEWithLogitsLoss()
if arg.model == "vnet":
model = VNetTrainer(arg, net, torch_device, recon_loss=recon_loss)
else:
model = CNNTrainer(arg, net, torch_device, recon_loss=recon_loss)
if arg.test is False:
model.train(train_loader, val_loader)
model.test(test_loader, val_loader)
valid_path,
arg.batch_size,
image_type=arg.output,
cpus=arg.cpus,
)
train_loader = MARLoader(
train_path,
arg.batch_size,
image_type=arg.output,
cpus=arg.cpus,
)
if arg.model == "fusion":
fusionnet = Fusionnet(1, 1, arg.ngf, arg.output, arg.clamp).cuda()
model = CNNTrainer(arg,
fusionnet,
recon_loss=nn.L1Loss(),
mse_loss=nn.MSELoss())
elif arg.model == "unet":
# for example
# unet = Unet(1, 1, arg.ngf, arg.output, arg.clamp).cuda()
# model = CNNTrainer(arg, Unet, recon_loss=nn.L1Loss(), mse_loss=nn.MSELoss())
raise NotImplementedError("Not Implemented Unet")
elif arg.model == "pix2pix":
# G = ...Net
# D = ...Discriminator
# model = GANTrainer(...)
raise NotImplementedError("Not Implemented Model")
else:
raise NotImplementedError("Not Implemented Model")
elif arg.model == "unet_nonlocal":
net = Unet_NonLocal2D(feature_scale=arg.feature_scale)
elif arg.model == "unet_gridatt":
net = Unet_GridAttention2D(feature_scale=arg.feature_scale)
elif arg.model == "unet_multiatt":
net = Unet_CT_multi_attention_dsv_2D(feature_scale=arg.feature_scale)
elif arg.model == "unet_dilated":
net = Unet_Dilation_2D(feature_scale=arg.feature_scale,
dilation=arg.dilation)
else:
raise NotImplementedError("Not Implemented Model")
net = nn.DataParallel(net).to(torch_device)
model = CNNTrainer(arg,
net,
torch_device,
recon_loss=nn.BCEWithLogitsLoss(),
metric=jaccard_similarity_score)
if arg.infer:
train_loader = NucleusLoader(train_path,
1,
transform=preprocess,
channel=arg.in_channel,
torch_type=arg.dtype,
cpus=arg.cpus,
shuffle=False,
drop_last=False)
valid_loader = NucleusLoader(valid_path,
1,
transform=preprocess,
channel=arg.in_channel,
#if arg.loss == "l2":
# recon_loss = nn.L2Loss()
#elif arg.loss == "l1":
# recon_loss = nn.L1Loss()
backzero = 1
recon_loss = FocalLoss3d_ver1(backzero=backzero)
val_loss = FocalLoss3d_ver1(gamma=2,
pw=1,
threshold=1.0,
erode=3,
backzero=backzero)
model = CNNTrainer(arg,
net,
torch_device,
recon_loss=recon_loss,
val_loss=val_loss,
logger=logger)
#model.load(filename="epoch[0402]_losssum[0.016887].pth.tar")
##1013 version, quite nice
#arg.save_dir = "nets_1008_unet_glob1007_disdice_FRE_pw10_erode2_feat1_trans64"
#model.load(filename="epoch[0483]_losssum[0.016087].pth.tar")
##1014 version
#model.load(filename="epoch[0418]_losssum[0.049294].pth.tar")
model.load(filename="epoch[0324]_losssum[0.053456].pth.tar")
#filename = "epoch[0401]_losssum[0.059813].pth.tar"
#model.load(filename=filename)
sdir = "/data1/0000Deployed_JY/181013_analCART"