def __init__(self, options, path):
print('Prepare the network and data.')
self._options = options
self._path = path
# Network.
self._net = torch.nn.DataParallel(BCNN()).cuda()
print(self._net)
# Criterion.
self._criterion = torch.nn.CrossEntropyLoss().cuda()
# Solver.
self._solver = torch.optim.SGD(
[
self._net.module.w1, self._net.module.w2, self._net.module.w3,
self._net.module.w4
] + list(self._net.module.fc.parameters()),
lr=self._options['base_lr'],
momentum=0.9,
weight_decay=self._options['weight_decay'])
self._scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
self._solver,
mode='max',
factor=0.1,
patience=5,
verbose=True,
threshold=1e-4)
train_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448), # Let smaller edge match
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.RandomCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
test_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448),
torchvision.transforms.CenterCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
train_data = cub200.CUB200(root=self._path['cub200'],
train=True,
download=True,
transform=train_transforms)
test_data = cub200.CUB200(root=self._path['cub200'],
train=False,
download=True,
transform=test_transforms)
self._train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=self._options['batch_size'],
shuffle=True,
num_workers=16,
pin_memory=True)
self._test_loader = torch.utils.data.DataLoader(test_data,
batch_size=16,
shuffle=False,
num_workers=16,
pin_memory=True)
def __init__(self, options, path):
self._options = options
self._path = path
# Network.
self._net = torch.nn.DataParallel(HBP()).cuda()
print(self._net)
# Criterion.
self._criterion = torch.nn.CrossEntropyLoss().cuda()
# Solver.
param_to_optim = []
for param in self._net.parameters():
if param.requires_grad == False:
continue
param_to_optim.append(param)
self._solver = torch.optim.SGD(
param_to_optim, lr=self._options['base_lr'],
momentum=0.9, weight_decay=self._options['weight_decay'])
milestones = [40,60,80,100]
self._scheduler = torch.optim.lr_scheduler.MultiStepLR(self._solver,milestones = milestones,gamma=0.25)
train_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448), # Let smaller edge match
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.RandomCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
test_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448),
torchvision.transforms.CenterCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
train_data = cub200.CUB200(
root=self._path['cub200'], train=True, download=True,
transform=train_transforms)
test_data = cub200.CUB200(
root=self._path['cub200'], train=False, download=True,
transform=test_transforms)
#added to train on one batch
batch_indices = [536,54,4400,1769,1518,1287,5554,4919,2547,2249,5757,589] #12 indices
subset_sampler = torch.utils.data.sampler.SubsetRandomSampler(batch_indices)
self._train_loader = torch.utils.data.DataLoader(
train_data, batch_size=self._options['batch_size'],
shuffle=False, sampler=subset_sampler, num_workers=4, pin_memory=True)
# self._train_loader = torch.utils.data.DataLoader(
# train_data, batch_size=self._options['batch_size'],
# shuffle=True, num_workers=4, pin_memory=True)
self._test_loader = torch.utils.data.DataLoader(
test_data, batch_size=1,
shuffle=False, num_workers=4, pin_memory=True)
def __init__(self, options, path):
print('Prepare the network and data.')
self._options = options
self._path = path
# Network.
self._net = torch.nn.DataParallel(HBP()).cuda()
print(self._net)
self._net.load_state_dict(torch.load(self._path['model']))
# Criterion.
self._criterion = torch.nn.CrossEntropyLoss().cuda()
# Solver.
param_to_optim = []
for param in self._net.parameters():
param_to_optim.append(param)
self._solver = torch.optim.SGD(
param_to_optim,
lr=self._options['base_lr'],
momentum=0.9,
weight_decay=self._options['weight_decay'])
milestones = [100]
self._scheduler = torch.optim.lr_scheduler.MultiStepLR(
self._solver, milestones=milestones, gamma=0.25)
train_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448), # Let smaller edge match
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.RandomCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
test_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448),
torchvision.transforms.CenterCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
train_data = cub200.CUB200(root=self._path['cub200'],
train=True,
download=True,
transform=train_transforms)
test_data = cub200.CUB200(root=self._path['cub200'],
train=False,
download=True,
transform=test_transforms)
self._train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=self._options['batch_size'],
shuffle=True,
num_workers=4,
pin_memory=True)
self._test_loader = torch.utils.data.DataLoader(test_data,
batch_size=16,
shuffle=False,
num_workers=4,
pin_memory=True)
def __init__(self, paths):
"""Prepare the network and data.
Args:
paths, dict<str, str>: Useful paths.
"""
print('Prepare the network and data.')
# Configurations.
self._paths = paths
# Network.
# self._net = torchvision.models.vgg16(pretrained=True).features
# self._net = torch.nn.Sequential(*list(self._net.children())[:-2])
# self._net = self._net.cuda()
self._net = torchvision.models.resnet50(pretrained=True)
self._net = torch.nn.Sequential(*list(self._net.children())[:-2])
print('resnet50!----finished!')
print(self._net)
# Data.
# NOTE: Resize such that the short edge is 448, and then ceter crop 448.
train_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=(448, 448)),
# torchvision.transforms.CenterCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
test_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=(448, 448)),
# torchvision.transforms.CenterCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
train_data = cub200.CUB200(root=self._paths['cub200'],
train=True,
transform=train_transforms,
download=False)
test_data = cub200.CUB200(root=self._paths['cub200'],
train=False,
transform=test_transforms,
download=False)
self._train_loader = torch.utils.data.DataLoader(train_data,
batch_size=1,
shuffle=False,
num_workers=0,
pin_memory=False)
self._test_loader = torch.utils.data.DataLoader(test_data,
batch_size=1,
shuffle=False,
num_workers=0,
pin_memory=False)
def getStat(self):
"""Get the mean and std value for a certain dataset."""
print('Compute mean and variance for training data.')
# train_data = cub200.CUB200(
# root=self._path['cub200'], train=True,
# transform=torchvision.transforms.ToTensor(), download=True)
train_data = cub200.CUB200(root=self._path[cub200.coarse_class],
train=True,
transform=torchvision.transforms.ToTensor(),
download=True) #change
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=1,
shuffle=False,
num_workers=4,
pin_memory=True)
mean = torch.zeros(3)
std = torch.zeros(3)
for X, _ in train_loader:
for d in range(3):
mean[d] += X[:, d, :, :].mean()
std[d] += X[:, d, :, :].std()
mean.div_(len(train_data))
std.div_(len(train_data))
print(mean)
print(std)
def getStat(self):
print('Compute mean and variance for training data.')
train_data = cub200.CUB200(root=self._path['cub200'],
train=True,
transform=torchvision.transforms.ToTensor(),
download=True)
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=1,
shuffle=False,
num_workers=4,
pin_memory=True)
mean = torch.zeros(3)
std = torch.zeros(3)
for X, _ in train_loader:
for d in range(3):
mean[d] += X[:, d, :, :].mean()
std[d] += X[:, d, :, :].std()
mean.div_(len(train_data))
std.div_(len(train_data))
print("mean")
print(mean)
print("standard deviation")
print(std)
def __init__(self, options, path):
self._options = options
self._path = path
# Network.
self._net = torch.nn.DataParallel(HBP()).cuda()
print(self._net)
# Criterion.
#self._criterion = torch.nn.CrossEntropyLoss().cuda()
self._criterion = torch.nn.CrossEntropyLoss().cuda()
# Solver.
param_to_optim = []
for param in self._net.parameters():
if param.requires_grad == False:
continue
param_to_optim.append(param)
# Using ADAM
self._solver = torch.optim.Adam(
param_to_optim,
lr=self._options['base_lr'],
weight_decay=self._options['weight_decay'])
# Commented
# self._solver = torch.optim.SGD(
# param_to_optim, lr=self._options['base_lr'],
# momentum=0.9, weight_decay=self._options['weight_decay'])
milestones = [40, 60, 80, 100]
self._scheduler = torch.optim.lr_scheduler.MultiStepLR(
self._solver, milestones=milestones, gamma=0.25)
train_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448), # Let smaller edge match
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.RandomCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
test_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448),
torchvision.transforms.CenterCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
train_data = cub200.CUB200(root=self._path['cub200'],
train=True,
download=True,
transform=train_transforms)
test_data = cub200.CUB200(root=self._path['cub200'],
train=False,
download=True,
transform=test_transforms)
batch_indices = [
536, 54, 4400, 1769, 1518, 1287, 5554, 4919, 2547, 2249, 5757, 589
] #12 indices
'''
5508,3672,4026,2859,1701,3799,5087,2493,3256,5808,4979,3464,4302,4653,3412,
5356,4758,2591,3814,2814,2647,3535,4422,4417,1810,1946,2256,2536,3457,5382,
4850,3930,306,2658,3066,298,109,4691,2458,4867,5841,1283,640,3646,5635,4700,
2306,5317,5666,595,4930,4011,906,3592,1372,3972,3415,594,1364,4541,4727,5026,
1939,2917,347,5931,970,622,5380,2673,1145,1339,3743,1192,1606,1075,5740,2909,
4918,3315,3741,1486,149,3765,3123,1756,5098]
'''
subset_sampler = torch.utils.data.sampler.SubsetRandomSampler(
batch_indices)
self._train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=self._options['batch_size'],
shuffle=False,
sampler=subset_sampler,
num_workers=4,
pin_memory=True)
#shuffle=True, num_workers=4, pin_memory=True) #changed to shuffle=false for one batch
self._test_loader = torch.utils.data.DataLoader(
test_data,
batch_size=self._options['test_batch_size'],
shuffle=False,
num_workers=4,
pin_memory=True)
def __init__(self, options, path, ckpt_basename='vgg_16'):
"""Prepare the network, criterion, solver, and data.
Args:
options, dict: Hyperparameters.
"""
print('Prepare the network and data.')
self._options = options
self._path = path
self.ckpt_basename = ckpt_basename
# Network.
self._net = BCNN(freeze_features=True)
#self._net = torch.nn.DataParallel(self._net)
self._net.features = torch.nn.DataParallel(self._net.features)
self._net.cuda()
if 'ckpt_path' in self._path:
if os.path.exists(self._path['ckpt_path']):
print('Continue from', self._path['ckpt_path'])
self._net.load_state_dict(torch.load(self._path['ckpt_path']))
else:
print('Ckpt {} not found!'.format(self._path['ckpt_path']))
print(self._net)
# Criterion.
self._criterion = torch.nn.CrossEntropyLoss().cuda()
# Solver.
self._solver = torch.optim.SGD(
self._net.fc.parameters(),
lr=self._options['base_lr'],
momentum=0.9,
weight_decay=self._options['weight_decay'])
if self._options['lr_scheduler'] == 'reduce_on_plateau':
self._scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
self._solver,
mode='max',
factor=0.1,
patience=5,
verbose=True,
threshold=1e-4,
min_lr=1e-6)
elif self._options['lr_scheduler'] == 'fixed':
self._scheduler = torch.optim.lr_scheduler.LambdaLR(
self._solver, lambda epoch: 1.0)
else:
raise ValueError('Unknown scheduler:',
self._options['lr_scheduler'])
# Imagenet normalization
normalize = torchvision.transforms.Normalize(
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
train_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448), # Let smaller edge match
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.RandomCrop(size=448),
torchvision.transforms.ToTensor(),
normalize
])
test_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448),
torchvision.transforms.CenterCrop(size=448),
torchvision.transforms.ToTensor(), normalize
])
train_data = cub200.CUB200(root=self._path['cub200'],
train=True,
download=True,
transform=train_transforms)
test_data = cub200.CUB200(root=self._path['cub200'],
train=False,
download=True,
transform=test_transforms)
self._train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=self._options['batch_size'],
shuffle=True,
num_workers=4,
pin_memory=True)
self._test_loader = torch.utils.data.DataLoader(test_data,
batch_size=16,
shuffle=False,
num_workers=4,
pin_memory=True)
def __init__(self, options, path):
self._options = options
self._path = path
# Network.
self._net = torch.nn.DataParallel(HBP()).cuda()
print(self._net)
# Criterion.
#self._criterion = torch.nn.CrossEntropyLoss().cuda()
self._criterion = torch.nn.CrossEntropyLoss().cuda()
# Solver.
param_to_optim = []
for param in self._net.parameters():
if param.requires_grad == False:
continue
param_to_optim.append(param)
# Abhi:
self._solver = torch.optim.Adam(
param_to_optim,
lr=self._options['base_lr'],
weight_decay=self._options['weight_decay'])
# Commented by Abhi
# self._solver = torch.optim.SGD(
# param_to_optim, lr=self._options['base_lr'],
# momentum=0.9, weight_decay=self._options['weight_decay'])
milestones = [40, 60, 80, 100]
self._scheduler = torch.optim.lr_scheduler.MultiStepLR(
self._solver, milestones=milestones, gamma=0.25)
train_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448), # Let smaller edge match
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.RandomCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
test_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448),
torchvision.transforms.CenterCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
train_data = cub200.CUB200(root=self._path['cub200'],
train=True,
download=True,
transform=train_transforms)
test_data = cub200.CUB200(root=self._path['cub200'],
train=False,
download=True,
transform=test_transforms)
self._train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=self._options['batch_size'],
shuffle=False,
num_workers=4,
pin_memory=True)
#shuffle=True, num_workers=4, pin_memory=True) #changed to false for one batch
self._test_loader = torch.utils.data.DataLoader(
test_data,
batch_size=self._options['test_batch_size'],
shuffle=False,
num_workers=4,
pin_memory=True)
def __init__(self, options, paths):
"""Prepare the network, criterion, optimizer, and data.
Args:
options, dict<str, float/int>: Hyperparameters.
paths, dict<str, str>: Useful paths.
"""
print('Prepare the network and data.')
# Configurations.
self._options = options
self._paths = paths
# Network.
if self._paths['pretrained'] is not None:
self._net = torch.nn.DataParallel(
model.BCNN(num_classes=200, is_all=True)).cuda()
self._net.load_state_dict(torch.load(self._paths['pretrained']),
strict=False)
else:
self._net = torch.nn.DataParallel(
model.BCNN(num_classes=200, is_all=False)).cuda()
print(self._net)
self._criterion = torch.nn.CrossEntropyLoss().cuda()
# Optimizer.
self._optimizer = torch.optim.SGD(
self._net.parameters(),
lr=self._options['base_lr'],
momentum=0.9,
weight_decay=self._options['weight_decay'])
self._scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
self._optimizer,
mode='max',
factor=0.1,
patience=8,
verbose=True,
threshold=1e-4)
# Data.
if self._paths['pretrained'] is not None:
train_transforms = torchvision.transforms.Compose([
torchvision.transforms.RandomResizedCrop(size=448,
scale=(0.8, 1.0)),
# torchvision.transforms.Resize(size=448),
# torchvision.transforms.RandomCrop(size=448),
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
])
test_transforms = torchvision.transforms.Compose([
# torchvision.transforms.Resize(size=448),
# torchvision.transforms.CenterCrop(size=448),
torchvision.transforms.Resize(size=(448, 448)),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
])
train_data = cub200.CUB200(root=self._paths['cub200'],
train=True,
transform=train_transforms,
download=True)
test_data = cub200.CUB200(root=self._paths['cub200'],
train=False,
transform=test_transforms,
download=True)
else:
train_data = cub200.CUB200ReLU(root=self._paths['cub200'],
train=True)
test_data = cub200.CUB200ReLU(root=self._paths['cub200'],
train=False)
self._train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=self._options['batch_size'],
shuffle=True,
num_workers=4,
pin_memory=False)
self._test_loader = torch.utils.data.DataLoader(
test_data,
batch_size=(64 if self._paths['pretrained'] is not None else 4096),
shuffle=False,
num_workers=4,
pin_memory=False)
def main():
_net = torch.nn.DataParallel(BCNN()).cuda()
train_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=224),
torchvision.transforms.CenterCrop(size=224),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
train_data = cub200.CUB200(root=path,
train=True,
download=True,
transform=train_transforms)
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=8,
shuffle=False,
num_workers=2,
pin_memory=True)
test_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=224),
torchvision.transforms.CenterCrop(size=224),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
test_data = cub200.CUB200(root=path,
train=False,
download=True,
transform=test_transforms)
test_loader = torch.utils.data.DataLoader(test_data,
batch_size=8,
shuffle=False,
num_workers=2,
pin_memory=True)
_net.train(False)
X_numpy = None
y_numpy = None
for X, y in tqdm(train_loader):
X = torch.autograd.Variable(X.cuda())
y = torch.autograd.Variable(y.cuda(async=True))
score = _net(X).cpu().detach().numpy()
y = y.cpu().numpy()
X_numpy = np.append(X_numpy, score)
y_numpy = np.append(y_numpy, y)
np.savez('VGG-D_wo_ft_train_pytorch.npz', X_numpy, y_numpy)
X_numpy = None
y_numpy = None
for X, y in tqdm(test_loader):
X = torch.autograd.Variable(X.cuda())
y = torch.autograd.Variable(y.cuda(async=True))
score = _net(X).cpu().detach().numpy()
y = y.cpu().numpy()
X_numpy = np.append(X_numpy, score)
y_numpy = np.append(y_numpy, y)
np.savez('VGG-D_wo_ft_test_pytorch.npz', X_numpy, y_numpy)
def __init__(self, options, path):
"""Prepare the network, criterion, solver, and data.
Args:
options, dict: Hyperparameters.
"""
print('Prepare the network and data.')
self._options = options
self._path = path
# Network.
if self._options['dataset'] == 'cub200':
num_classes = 200
elif self._options['dataset'] == 'aircraft':
num_classes = 100
else:
raise NotImplementedError("Dataset " + self._options['dataset'] +
" is not implemented.")
self._net = BCNN(num_classes=num_classes,
pretrained=options['target'] == 'fc')
# Load the model from disk.
if options['target'] == 'all':
self._net.load_state_dict(torch.load(self._path['model']))
self._net = torch.nn.parallel.DistributedDataParallel(
self._net.cuda(),
device_ids=[self._options['local_rank']],
output_device=self._options['local_rank'])
if dist.get_rank() == 0:
print(self._net)
# Criterion.
self._criterion = torch.nn.CrossEntropyLoss().cuda()
# Solver.
self._solver = torch.optim.SGD(
self._net.module.trainable_params,
lr=self._options['base_lr'] * dist.get_world_size(),
momentum=0.9,
weight_decay=self._options['weight_decay'])
self._scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
self._solver,
mode='max',
factor=0.1,
patience=3,
verbose=True,
threshold=1e-4)
train_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448), # Let smaller edge match
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.RandomCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
test_transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(size=448),
torchvision.transforms.CenterCrop(size=448),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
if self._options['dataset'] == 'cub200':
train_data = cub200.CUB200(root=self._path['dataset'],
train=True,
download=True,
transform=train_transforms)
test_data = cub200.CUB200(root=self._path['dataset'],
train=False,
download=True,
transform=test_transforms)
elif self._options['dataset'] == 'aircraft':
train_data = aircraft.Aircraft(root=self._path['dataset'],
train=True,
download=True,
transform=train_transforms)
test_data = aircraft.Aircraft(root=self._path['dataset'],
train=False,
download=True,
transform=test_transforms)
else:
raise NotImplementedError("Dataset " + self._options['dataset'] +
" is not implemented.")
# Partition dataset among workers using DistributedSampler
train_sampler = distributed.DistributedSampler(
train_data,
num_replicas=dist.get_world_size(),
rank=dist.get_rank())
test_sampler = distributed.DistributedSampler(
test_data,
num_replicas=dist.get_world_size(),
rank=dist.get_rank())
self._train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=self._options['batch_size'],
shuffle=False,
num_workers=4,
pin_memory=True,
sampler=train_sampler)
self._test_loader = torch.utils.data.DataLoader(
test_data,
batch_size=self._options['batch_size'],
shuffle=False,
num_workers=4,
pin_memory=True,
sampler=test_sampler)
def __init__(self, options, paths):
"""Model preparation."""
# Configurations.
self._options = options
self._paths = paths
# Prepare model and loss function.
self._model = model.Model(architecture=self._options.arch,
input_size=self._options.input_size,
num_classes=options.num_classes)
self._model = torch.nn.DataParallel(self._model).cuda()
print(self._model)
torchsummary.summary(self._model.module,
input_size=(3, self._options.input_size,
self._options.input_size))
# Display configurations and options.
print('PyTorch %s, CUDA %s, cuDNN %s, GPU %s' %
(torch.__version__, torch.version.cuda,
torch.backends.cudnn.version(), torch.cuda.get_device_name(0)))
print(self._options)
print('Code version %s' % __version__)
print('-' * 80)
# Prepare optimizer and optimizer scheduler.
self._loss_function = torch.nn.CrossEntropyLoss().cuda()
self._optimizer = torch.optim.SGD(
self._model.parameters(),
lr=self._options.lr,
momentum=0.9,
weight_decay=self._options.weight_decay)
self._scheduler = torch.optim.lr_scheduler.MultiStepLR(
self._optimizer, milestones=[60, 115], gamma=0.1)
# Prepare dataset.
resize_size = 512 if self._options.input_size == 448 else 256
train_transform = torchvision.transforms.Compose([
torchvision.transforms.Resize((resize_size, resize_size)),
torchvision.transforms.RandomCrop(
(self._options.input_size, self._options.input_size)),
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
])
val_transform = torchvision.transforms.Compose([
torchvision.transforms.Resize((resize_size, resize_size)),
torchvision.transforms.CenterCrop(
(self._options.input_size, self._options.input_size)),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
])
train_data = cub200.CUB200(root=self._paths.data,
train=True,
transform=train_transform)
val_data = cub200.CUB200(root=self._paths.data,
train=False,
transform=val_transform)
print('Dataset size: train=%d, val=%d' %
(len(train_data), len(val_data)))
self._train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=self._options.bs,
shuffle=True,
num_workers=16,
pin_memory=True)
self._val_loader = torch.utils.data.DataLoader(
val_data,
batch_size=self._options.bs,
shuffle=False,
num_workers=16,
pin_memory=True)
# Visulization setup.
self._visdom = visdom.Visdom(env='Term %d' % self._options.term,
use_incoming_socket=False)
assert self._visdom.check_connection()
self._visdom.close()