def __init__(self, config, hparams):
self.config = config
self.hparams = hparams
if self.config.dataset == 'CamVid':
train_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtRandomCrop(128, pad_if_needed=True),
utils.ext_transforms.ExtRandomHorizontalFlip(),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
test_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
train_dataset = CamVid(self.config.dataset_path,
split='train',
transform=train_transforms)
test_dataset = CamVid(self.config.dataset_path,
split='val',
transform=test_transforms)
# NYUv2 dataset
if self.config.dataset == 'Nyu':
train_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtRandomCrop(128, pad_if_needed=True),
utils.ext_transforms.ExtRandomHorizontalFlip(),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
test_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
train_dataset = NYUv2(self.config.dataset_path,
split='train',
transform=train_transforms)
test_dataset = NYUv2(self.config.dataset_path,
split=self.config.test_mode,
transform=test_transforms)
elif self.config.dataset == 'Cityscapes':
train_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtRandomCrop(128, pad_if_needed=True),
utils.ext_transforms.ExtRandomHorizontalFlip(),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
test_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
train_dataset = Cityscapes(self.config.dataset_path,
split='train',
mode='fine',
transform=train_transforms)
test_dataset = Cityscapes(self.config.dataset_path,
split='test',
mode='fine',
transform=test_transforms)
self.train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=self.hparams.train_batch_size,
shuffle=True,
num_workers=6,
drop_last=True)
self.test_loader = torch.utils.data.DataLoader(
test_dataset,
batch_size=self.hparams.test_batch_size,
shuffle=False,
num_workers=6)
if self.config.model == 'resnet50_pretrained':
self.model = deeplabv3.deeplabv3_resnet50(num_classes=13,
dropout_p=0.5,
pretrained_backbone=True)
if self.config.model == 'resnet100_pretrained':
self.model = deeplabv3.deeplabv3_resnet101(
num_classes=13, dropout_p=0.5, pretrained_backbone=True)
if self.config.model == 'resnet50':
self.model = deeplabv3.deeplabv3_resnet50(
num_classes=13, dropout_p=0.5, pretrained_backbone=False)
if self.config.model == 'mobilenet':
self.model = deeplabv3.deeplabv3_mobilenet(
num_classes=13, dropout_p=0.5, pretrained_backbone=False)
self.device = torch.device("cuda")
model_checkpoint = torch.load(self.hparams.model_checkpoint)
self.model.load_state_dict(model_checkpoint['state_dict'])
print("Teacher checkpoint loaded succesfully.")
self.model = self.model.to(self.device)
self.model.eval()
self.G = wgan.DCGAN_G(self.hparams.img_size,
self.hparams.nz,
nc=3,
ngf=64,
ngpu=1)
self.D = wgan.DCGAN_D(self.hparams.img_size,
self.hparams.nz,
nc=3,
ndf=64,
ngpu=1)
self.G = self.G.to(self.device)
self.D = self.D.to(self.device)
self.G.apply(self.weights_init)
self.D.apply(self.weights_init)
self.fixed_noise = torch.randn(self.hparams.train_batch_size,
self.hparams.nz,
1,
1,
device=self.device)
self.one = torch.FloatTensor([1]).to(self.device)
self.m_one = torch.FloatTensor([-1]).to(self.device)
self.gen_iterations = 0
self.diversity_gt = torch.ones(
self.config.num_classes,
device=self.device) / self.config.num_classes
# self.optimizerD = torch.optim.Adam(self.D.parameters(),
# lr=self.hparams.lr,
# betas=(0.5, 0.999))
# self.optimizerG = torch.optim.Adam(self.G.parameters(),
# lr=self.hparams.lr,
# betas=(0.5, 0.999))
self.optimizerD = torch.optim.RMSprop(self.D.parameters(),
lr=self.hparams.lr)
self.optimizerG = torch.optim.RMSprop(self.G.parameters(),
lr=self.hparams.lr)
if os.path.exists(self.config.log_dir):
raise Exception("Log directory exists")
self.logger = SummaryWriter(log_dir=self.config.log_dir)
def __init__(self, config, hparams):
self.config = config
self.hparams = hparams
if self.config.dataset == 'CamVid':
train_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtRandomCrop(128, pad_if_needed=True),
utils.ext_transforms.ExtRandomHorizontalFlip(),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5,), (0.5,)),
])
test_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5,), (0.5,)),
])
train_dataset = CamVid(self.config.dataset_path, split='train', transform=train_transforms)
test_dataset = CamVid(self.config.dataset_path, split=self.config.test_mode, transform=test_transforms)
# NYUv2 dataset
if self.config.dataset == 'Nyu':
train_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtRandomCrop(128, pad_if_needed=True),
utils.ext_transforms.ExtRandomHorizontalFlip(),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
test_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
train_dataset = NYUv2(self.config.dataset_path,
split='train',
transform=train_transforms)
test_dataset = NYUv2(self.config.dataset_path,
split=self.config.test_mode,
transform=test_transforms)
elif self.config.dataset == 'Cityscapes':
train_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtRandomCrop(128, pad_if_needed=True),
utils.ext_transforms.ExtRandomHorizontalFlip(),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5,), (0.5,)),
])
test_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5,), (0.5,)),
])
train_dataset = Cityscapes(self.config.dataset_path, split='train', mode='fine', transform=train_transforms)
test_dataset = Cityscapes(self.config.dataset_path, split='val', mode='fine', transform=test_transforms)
self.train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=self.hparams.train_batch_size, shuffle=True, num_workers=6)
self.test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=self.hparams.test_batch_size, shuffle=False, num_workers=6)
if self.config.teacher == 'resnet50_pretrained':
self.teacher = deeplabv3.deeplabv3_resnet50(num_classes=13,
dropout_p=0.5,
pretrained_backbone=True)
if self.config.teacher == 'resnet100_pretrained':
self.teacher = deeplabv3.deeplabv3_resnet101(num_classes=13,
dropout_p=0.5,
pretrained_backbone=True)
if self.config.teacher == 'resnet50':
self.teacher = deeplabv3.deeplabv3_resnet50(
num_classes=13, dropout_p=0.5, pretrained_backbone=False)
if self.config.teacher == 'mobilenet':
self.teacher = deeplabv3.deeplabv3_mobilenet(
num_classes=13, dropout_p=0.5, pretrained_backbone=False)
if self.config.model == 'resnet50_pretrained':
self.model = deeplabv3.deeplabv3_resnet50(num_classes=13,
dropout_p=0.5,
pretrained_backbone=True)
if self.config.model == 'resnet100_pretrained':
self.model = deeplabv3.deeplabv3_resnet101(num_classes=13,
dropout_p=0.5,
pretrained_backbone=True)
if self.config.model == 'resnet50':
self.model = deeplabv3.deeplabv3_resnet50(
num_classes=13, dropout_p=0.5, pretrained_backbone=False)
if self.config.model == 'mobilenet':
self.model = deeplabv3.deeplabv3_mobilenet(
num_classes=13, dropout_p=0.5, pretrained_backbone=False)
# Uses SGD optimizer for better performance than Adam
self.optimizer = torch.optim.SGD(
self.model.parameters(),
lr=self.hparams.lr,
weight_decay=self.hparams.weight_decay,
momentum=self.hparams.momentum)
self.device = torch.device("cuda")
self.model = self.model.to(self.device)
teacher_checkpoint = torch.load(self.config.teacher_checkpoint)
self.teacher.load_state_dict(teacher_checkpoint['state_dict'])
self.teacher = self.teacher.to(self.device)
self.teacher.eval()
print("Teacher checkpoint loaded successfully.")
generator_checkpoint = torch.load(self.config.generator_checkpoint)
self.hparams.nz = generator_checkpoint['hparams']['nz']
self.G = dcgan.DcGanGenerator(nz=generator_checkpoint['hparams']['nz'])
self.G.load_state_dict(generator_checkpoint['g_state_dict'])
self.G.to(self.device)
self.G.eval()
print("Generator model loaded successfully")
if self.hparams.lr_scheduler:
self.scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer, self.hparams.scheduler_step, self.hparams.scheduler_gamma)
if os.path.exists(self.config.log_dir):
raise Exception("Log directory exists")
self.logger = SummaryWriter(log_dir=self.config.log_dir)
def __init__(self, config, hparams):
# Store config and hparams
self.config = config
self.hparams = hparams
# Camvid Dataset
if self.config.dataset == 'CamVid':
train_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtRandomCrop(128, pad_if_needed=True),
utils.ext_transforms.ExtRandomHorizontalFlip(),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
test_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
train_dataset = CamVid(self.config.dataset_path,
split='train',
transform=train_transforms)
test_dataset = CamVid(self.config.dataset_path,
split=self.config.test_mode,
transform=test_transforms)
# NYUv2 dataset
if self.config.dataset == 'Nyu':
train_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtRandomCrop(128, pad_if_needed=True),
utils.ext_transforms.ExtRandomHorizontalFlip(),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
test_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
train_dataset = NYUv2(self.config.dataset_path,
split='train',
transform=train_transforms)
test_dataset = NYUv2(self.config.dataset_path,
split=self.config.test_mode,
transform=test_transforms)
# Cityscapes Dataset
elif self.config.dataset == 'Cityscapes':
train_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtRandomCrop(128, pad_if_needed=True),
utils.ext_transforms.ExtRandomHorizontalFlip(),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
test_transforms = utils.ext_transforms.ExtCompose([
utils.ext_transforms.ExtResize(256),
utils.ext_transforms.ExtToTensor(),
utils.ext_transforms.ExtNormalize((0.5, ), (0.5, )),
])
train_dataset = Cityscapes(self.config.dataset_path,
split='train',
mode='fine',
transform=train_transforms)
test_dataset = Cityscapes(self.config.dataset_path,
split='test',
mode='fine',
transform=test_transforms)
self.train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=self.hparams.train_batch_size,
shuffle=True,
num_workers=6)
self.test_loader = torch.utils.data.DataLoader(
test_dataset,
batch_size=self.hparams.test_batch_size,
shuffle=False,
num_workers=6)
if self.config.model == 'resnet50_pretrained':
self.model = deeplabv3.deeplabv3_resnet50(num_classes=13,
dropout_p=0.5,
pretrained_backbone=True)
if self.config.model == 'resnet100_pretrained':
self.model = deeplabv3.deeplabv3_resnet101(
num_classes=13, dropout_p=0.5, pretrained_backbone=True)
if self.config.model == 'resnet50':
self.model = deeplabv3.deeplabv3_resnet50(
num_classes=13, dropout_p=0.5, pretrained_backbone=False)
if self.config.model == 'mobilenet':
self.model = deeplabv3.deeplabv3_mobilenet(
num_classes=13, dropout_p=0.5, pretrained_backbone=False)
# Uses SGD optimizer for better performance than Adam
self.optimizer = torch.optim.SGD(
self.model.parameters(),
lr=self.hparams.lr,
weight_decay=self.hparams.weight_decay,
momentum=self.hparams.momentum)
# self.optimizer = torch.optim.Adam(self.model.parameters(), lr=self.hparams.lr, weight_decay=self.hparams.weight_decay)
# Run on GPU only. Convert model to GPU
self.device = torch.device("cuda")
self.model = self.model.to(self.device)
# Learning rate scheduler
if self.hparams.lr_scheduler:
self.scheduler = torch.optim.lr_scheduler.StepLR(
self.optimizer, self.hparams.scheduler_step,
self.hparams.scheduler_gamma)
# Initialise a new logging directory and a tensorboard logger
if os.path.exists(self.config.log_dir):
raise Exception("Log directory exists")
self.logger = SummaryWriter(log_dir=self.config.log_dir)