def set_model(args, n_data):
# set the model
if args.model == 'c3d':
model = C3D(with_classifier=False)
elif args.model == 'r3d':
model = R3DNet(layer_sizes=(1, 1, 1, 1), with_classifier=False)
elif args.model == 'r21d':
model = R2Plus1DNet(layer_sizes=(1, 1, 1, 1), with_classifier=False)
if args.intra_neg:
contrast = NCEAverage(args.feat_dim, n_data, args.nce_k, args.nce_t,
args.nce_m, args.softmax)
else:
contrast = NCEAverage_ori(args.feat_dim, n_data, args.nce_k,
args.nce_t, args.nce_m, args.softmax)
criterion_1 = NCESoftmaxLoss() if args.softmax else NCECriterion(n_data)
criterion_2 = NCESoftmaxLoss() if args.softmax else NCECriterion(n_data)
# GPU mode
model = model.cuda()
contrast = contrast.cuda()
criterion_1 = criterion_1.cuda()
criterion_2 = criterion_2.cuda()
cudnn.benchmark = True
return model, contrast, criterion_1, criterion_2
transform=transform_test)
testloader = torch.utils.data.DataLoader(testset,
batch_size=100,
shuffle=False,
num_workers=2)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse',
'ship', 'truck')
ndata = trainset.__len__()
print('==> Building model..')
net = models.__dict__['ResNet18'](low_dim=args.low_dim)
# define leminiscate
lemniscate = NCEAverage(args.low_dim,
ndata,
args.nce_t,
args.nce_m,
batchSize=args.b)
if device == 'cuda':
net = torch.nn.DataParallel(net,
device_ids=range(torch.cuda.device_count()))
cudnn.benchmark = True
# Model
if args.test_only or len(args.resume) > 0:
# Load checkpoint.
print('==> Resuming from checkpoint..')
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/' + args.resume)
net.load_state_dict(checkpoint['net'])
def main():
global args, best_prec1
args = parser.parse_args()
args.distributed = args.world_size > 1
if args.distributed:
dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
world_size=args.world_size)
# create model
if args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
model = models.__dict__[args.arch](pretrained=True)
else:
print("=> creating model '{}'".format(args.arch))
model = models.__dict__[args.arch](low_dim=args.low_dim)
if not args.distributed:
if args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
else:
model.cuda()
model = torch.nn.parallel.DistributedDataParallel(model)
# Data loading code
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_dataset = datasets.ImageFolderInstance(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224, scale=(0.2,1.)),
transforms.RandomGrayscale(p=0.2),
transforms.ColorJitter(0.4, 0.4, 0.4, 0.4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
else:
train_sampler = None
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None),
num_workers=args.workers, pin_memory=True, sampler=train_sampler)
val_loader = torch.utils.data.DataLoader(
datasets.ImageFolderInstance(valdir, transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True)
# define lemniscate and loss function (criterion)
ndata = train_dataset.__len__()
if args.nce_k > 0:
lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t, args.nce_m).cuda()
criterion = NCECriterion(ndata).cuda()
else:
lemniscate = LinearAverage(args.low_dim, ndata, args.nce_t, args.nce_m).cuda()
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(), args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
lemniscate = checkpoint['lemniscate']
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})"
.format(args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
if args.evaluate:
kNN(0, model, lemniscate, train_loader, val_loader, 200, args.nce_t)
return
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, lemniscate, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = NN(epoch, model, lemniscate, train_loader, val_loader)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint({
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'lemniscate': lemniscate,
'best_prec1': best_prec1,
'optimizer' : optimizer.state_dict(),
}, is_best)
# evaluate KNN after last epoch
kNN(0, model, lemniscate, train_loader, val_loader, 200, args.nce_t)
def main():
global args, best_prec1
args = parser.parse_args()
args.distributed = args.world_size > 1
if args.distributed:
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size)
# create model
# if args.pretrained:
# print("=> using pre-trained model '{}'".format(args.arch))
# model = models.__dict__[args.arch](pretrained=True, finetune=args.finetune, low_dim= args.low_dim)
# else:
# print("=> creating model '{}'".format(args.arch))
#
# model = models.__dict__[args.arch](low_dim=args.low_dim)
# Data loading code
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# train_dataset = datasets.CombinedMaskDataset(
# other_data_path = '/home/saschaho/Simcenter/found_label_imgs',
# csv_root_folder='/home/saschaho/Simcenter/Floor_Elevation_Data/Streetview_Irma/Streetview_Irma/images',
# data_csv='/home/saschaho/Simcenter/Building_Information_Prediction/all_bims_train.csv',
# transform = transforms.Compose([
# transforms.RandomResizedCrop(224, scale=(0.2,1.)),
# transforms.RandomGrayscale(p=0.2),
# transforms.ColorJitter(0.4, 0.4, 0.4, 0.4),
# transforms.RandomHorizontalFlip(),
# transforms.ToTensor(),
# normalize,
# ]),attribute = 'first_floor_elevation_ft', mask_images=True)
# val_dataset = datasets.CombinedMaskDataset(
# csv_root_folder='/home/saschaho/Simcenter/Floor_Elevation_Data/Streetview_Irma/Streetview_Irma/images',
# data_csv='/home/saschaho/Simcenter/Building_Information_Prediction/all_bims_val.csv',
# transform=transforms.Compose([
# transforms.Resize(256),
# transforms.CenterCrop(224),
# transforms.ToTensor(),
# normalize,
# ]),
#attribute = 'first_floor_elevation_ft', mask_images=True)
train_transform = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(0.3, 1.)),
transforms.RandomGrayscale(p=0.5),
transforms.ColorJitter(0.5, 0.5, 0.5, 0.5),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(), normalize
])
val_transform = transforms.Compose(
[transforms.Resize((224, 224)),
transforms.ToTensor(), normalize])
train_dataset = First_Floor_Binary(args.attribute_name,
args.train_data,
args.image_folder,
transform=train_transform,
regression=args.regression,
mask_buildings=args.mask_buildings,
softmask=args.softmask)
val_dataset = First_Floor_Binary(args.attribute_name,
args.val_data,
args.image_folder,
transform=val_transform,
regression=args.regression,
mask_buildings=args.mask_buildings,
softmask=args.softmask)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
model = ResidualAttentionModel_92_Small(args.low_dim, dropout=False)
model = torch.nn.DataParallel(model).cuda()
print('Train dataset instances: {}'.format(len(train_loader.dataset)))
print('Val dataset instances: {}'.format(len(val_loader.dataset)))
# define lemniscate and loss function (criterion)
ndata = train_dataset.__len__()
if args.nce_k > 0:
lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t,
args.nce_m).cuda()
criterion = NCECriterion(ndata).cuda()
else:
lemniscate = LinearAverage(args.low_dim, ndata, args.nce_t,
args.nce_m).cuda()
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
#optimizer = RAdam(model.parameters())
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
keyname = [
keyname for keyname in model.state_dict().keys()
if 'fc.weight' in keyname
][0]
lat_vec_len_model = model.state_dict()[keyname].shape[0]
lat_vec_len_checkpoint = checkpoint['state_dict'][keyname].shape[0]
low_dim_differ = False
if lat_vec_len_model != lat_vec_len_checkpoint:
low_dim_differ = True
print(
'Warning: Latent vector sizes do not match. Assuming finetuning'
)
print(
'Lemniscate will be trained from scratch with new optimizer.'
)
del checkpoint['state_dict'][keyname]
del checkpoint['state_dict'][keyname.replace('weight', 'bias')]
missing_keys, unexpected_keys = model.load_state_dict(
checkpoint['state_dict'], strict=False)
if len(missing_keys) or len(unexpected_keys):
print('Warning: Missing or unexpected keys found.')
print('Missing: {}'.format(missing_keys))
print('Unexpected: {}'.format(unexpected_keys))
if not low_dim_differ:
# The memory bank will be trained from scratch if
# the low dim is different. Maybe later repopulated
lemniscate = checkpoint['lemniscate']
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
if args.evaluate:
kNN(0, model, lemniscate, train_loader, val_loader, 200, args.nce_t)
return
for epoch in range(args.start_epoch, args.epochs):
# if args.distributed:
# train_sampler.set_epoch(epoch)
#adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, lemniscate, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = NN(epoch, model, lemniscate, train_loader, val_loader)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'lemniscate': lemniscate,
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, args.name)
# evaluate KNN after last epoch
kNN(0, model, lemniscate, train_loader, val_loader, 200, args.nce_t)
def main():
global args, best_prec1, best_prec1_past, best_prec1_future
args = parser.parse_args()
args.distributed = args.world_size > 1
if args.distributed:
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size)
# create model
if args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
model = models.__dict__[args.arch](pretrained=True)
else:
print("=> creating model '{}'".format(args.arch))
model = models.__dict__[args.arch](low_dim=args.low_dim)
if not args.distributed:
if args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.to(get_device(args.gpu))
else:
model = torch.nn.DataParallel(model).to(get_device(args.gpu))
else:
model.to(get_device(args.gpu))
model = torch.nn.parallel.DistributedDataParallel(model)
# Data loading code
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_dataset = Dataset(traindir, n_frames)
val_dataset = Dataset(valdir, n_frames)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=True, #(train_sampler is None),
num_workers=args.workers)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers)
# define lemniscate and loss function (criterion)
ndata = train_dataset.__len__()
if args.nce_k > 0:
lemniscate = NCEAverage(args.gpu, args.low_dim, ndata, args.nce_k,
args.nce_t,
args.nce_m).to(get_device(args.gpu))
criterion = NCECriterion(ndata).to(get_device(args.gpu))
else:
lemniscate = LinearAverage(args.low_dim, ndata, args.nce_t,
args.nce_m).to(get_device(args.gpu))
criterion = nn.CrossEntropyLoss().to(get_device(args.gpu))
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
lemniscate = checkpoint['lemniscate']
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
if args.evaluate:
kNN(0, model, lemniscate, train_loader, val_loader, 200, args.nce_t)
return
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, lemniscate, criterion, optimizer, epoch)
# evaluate on validation set
prec1, prec1_past, prec1_future = NN(epoch, model, lemniscate,
train_loader, val_loader)
add_epoch_score('epoch_scores.txt', epoch, prec1)
add_epoch_score('epoch_scores_past.txt', epoch, prec1_past)
add_epoch_score('epoch_scores_future.txt', epoch, prec1_future)
# Sascha: This is a bug because it seems prec1 or best_prec1 is a vector at some point with
# more than one entry
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'lemniscate': lemniscate,
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, epoch)
is_best_past = prec1_past > best_prec1_past
best_prec1_past = max(prec1_past, best_prec1_past)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'lemniscate': lemniscate,
'best_prec1_past': best_prec1_past,
'optimizer': optimizer.state_dict(),
},
is_best_past,
epoch,
best_mod='_past')
is_best_future = prec1_future > best_prec1_future
best_prec1_future = max(prec1_future, best_prec1_future)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'lemniscate': lemniscate,
'best_prec1_future': best_prec1_future,
'optimizer': optimizer.state_dict(),
},
is_best_future,
epoch,
best_mod='_future')
# evaluate KNN after last epoch
kNN(0, model, lemniscate, train_loader, val_loader, 200, args.nce_t)
download=True,
transform=transform_test)
testloader = torch.utils.data.DataLoader(testset,
batch_size=100,
shuffle=False,
num_workers=2)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse',
'ship', 'truck')
ndata = trainset.__len__()
print('==> Building model..')
net = models.__dict__['ResNet34'](low_dim=args.low_dim)
# define leminiscate
if args.nce_k > 0:
lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t,
args.nce_m)
else:
lemniscate = LinearAverage(args.low_dim, ndata, args.nce_t, args.nce_m)
if device == 'cuda':
net = torch.nn.DataParallel(net,
device_ids=range(torch.cuda.device_count()))
cudnn.benchmark = True
# Model
if args.test_only or len(args.resume) > 0:
# Load checkpoint.
print('==> Resuming from checkpoint..')
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/' + args.resume)
net.load_state_dict(checkpoint['net'])
def main():
global args, best_prec1
args = parser.parse_args()
# Initialize distributed processing
args.distributed = args.world_size > 1
if args.distributed:
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size)
# create model
if args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
model = models.__dict__[args.arch](pretrained=True,
low_dim=args.low_dim)
else:
print("=> creating model '{}'".format(args.arch))
model = models.__dict__[args.arch](low_dim=args.low_dim)
if not args.distributed:
if args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
else:
model.cuda()
model = torch.nn.parallel.DistributedDataParallel(model)
# Data loading code
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val')
normalize = transforms.Normalize(
mean=[0.485, 0.456, 0.406], # ImageNet stats
std=[0.229, 0.224, 0.225])
# normalize = transforms.Normalize(mean=[0.234, 0.191, 0.159], # xView stats
# std=[0.173, 0.143, 0.127])
print("Creating datasets")
cj = args.color_jit
train_dataset = datasets.ImageFolderInstance(
traindir,
transforms.Compose([
transforms.Resize((224, 224)),
# transforms.Grayscale(3),
# transforms.ColorJitter(cj, cj, cj, cj), #transforms.ColorJitter(0.4, 0.4, 0.4, 0.4),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.RandomRotation(45),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
elif args.balanced_sampling:
print("Using balanced sampling")
# Here's where we compute the weights for WeightedRandomSampler
class_counts = {v: 0 for v in train_dataset.class_to_idx.values()}
for path, ndx in train_dataset.samples:
class_counts[ndx] += 1
total = float(np.sum([v for v in class_counts.values()]))
class_probs = [
class_counts[ndx] / total for ndx in range(len(class_counts))
]
# make a list of class probabilities corresponding to the entries in train_dataset.samples
reciprocal_weights = [
class_probs[idx]
for i, (_, idx) in enumerate(train_dataset.samples)
]
# weights are the reciprocal of the above
weights = (1 / torch.Tensor(reciprocal_weights))
train_sampler = torch.utils.data.sampler.WeightedRandomSampler(
weights, len(train_dataset), replacement=True)
else:
#if args.red_data is < 1, then the training is done with a subsamle of the total data. Otherwise it's the total data.
data_size = len(train_dataset)
sub_index = np.random.randint(0, data_size,
round(args.red_data * data_size))
sub_index.sort()
train_sampler = torch.utils.data.sampler.SubsetRandomSampler(sub_index)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler)
print("Training on", len(train_dataset.imgs),
"images. Training batch size:", args.batch_size)
if len(train_dataset.imgs) % args.batch_size != 0:
print(
"Warning: batch size doesn't divide the # of training images so ",
len(train_dataset.imgs) % args.batch_size,
"images will be skipped per epoch.")
print("If you don't want to skip images, use a batch size in:",
get_factors(len(train_dataset.imgs)))
val_dataset = datasets.ImageFolderInstance(
valdir,
transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
normalize,
]))
val_bs = [
factor for factor in get_factors(len(val_dataset)) if factor < 500
][-1]
val_bs = 100
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=val_bs,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
print("Validating on", len(val_dataset), "images. Validation batch size:",
val_bs)
# define lemniscate and loss function (criterion)
ndata = train_dataset.__len__()
if args.nce_k > 0:
lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t,
args.nce_m)
criterion = NCECriterion(ndata).cuda()
else:
lemniscate = LinearAverage(args.low_dim, ndata, args.nce_t,
args.nce_m).cuda()
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.static_loss,
verbose=False)
optimizer.load_state_dict(checkpoint['optimizer'])
args.start_epoch = checkpoint['epoch']
# best_prec1 = checkpoint['best_prec1']
lemniscate = checkpoint['lemniscate']
if args.select_load:
pred = checkpoint['prediction']
print("=> loaded checkpoint '{}' (epoch {}, best_prec1 )".format(
args.resume,
checkpoint['epoch'])) #, checkpoint['best_prec1']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
# optionally fine-tune a model trained on a different dataset
elif args.fine_tune:
print("=> loading checkpoint '{}'".format(args.fine_tune))
checkpoint = torch.load(args.fine_tune)
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.static_loss,
verbose=False)
print("=> loaded checkpoint '{}' (epoch {})".format(
args.fine_tune, checkpoint['epoch']))
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.static_loss,
verbose=False)
# Optionally recompute memory. If fine-tuning, then we must recompute memory
if args.recompute_memory or args.fine_tune:
# Aaron - Experiments show that iterating over torch.utils.data.DataLoader will skip the last few
# unless the batch size evenly divides size of the data set. This shouldn't be the case
# according to documentation, there's even a flag for drop_last, but it's not working
# compute a good batch size for re-computing memory
memory_bs = [
factor for factor in get_factors(len(train_loader.dataset))
if factor < 500
][-1]
print("Recomputing memory using", train_dataset.root,
"with a batch size of", memory_bs)
transform_bak = train_loader.dataset.transform
train_loader.dataset.transform = val_loader.dataset.transform
temploader = torch.utils.data.DataLoader(
train_loader.dataset,
batch_size=memory_bs,
shuffle=False,
num_workers=train_loader.num_workers,
pin_memory=True)
lemniscate.memory = torch.zeros(len(train_loader.dataset),
args.low_dim).cuda()
model.eval()
with torch.no_grad():
for batch_idx, (inputs, targets,
indexes) in enumerate(tqdm.tqdm(temploader)):
batchSize = inputs.size(0)
features = model(inputs)
lemniscate.memory[batch_idx * batchSize:batch_idx * batchSize +
batchSize, :] = features.data
train_loader.dataset.transform = transform_bak
model.train()
cudnn.benchmark = True
if args.evaluate:
kNN(model, lemniscate, train_loader, val_loader, args.K, args.nce_t)
return
begin_train_time = datetime.datetime.now()
# my_knn(model, lemniscate, train_loader, val_loader, args.K, args.nce_t, train_dataset, val_dataset)
if args.tsne:
labels = idx_to_name(train_dataset, args.graph_labels)
tsne(lemniscate, args.tsne, labels)
if args.pca:
labels = idx_to_name(train_dataset, args.graph_labels)
pca(lemniscate, labels)
if args.view_knn:
my_knn(model, lemniscate, train_loader, val_loader, args.K, args.nce_t,
train_dataset, val_dataset)
if args.kmeans:
kmeans, yi = kmean(lemniscate, args.kmeans, 500, args.K, train_dataset)
D, I = kmeans.index.search(lemniscate.memory.data.cpu().numpy(), 1)
cent_group = {}
data_cent = {}
for n, i in enumerate(I):
if i[0] not in cent_group.keys():
cent_group[i[0]] = []
cent_group[i[0]].append(n)
data_cent[n] = i[0]
train_sampler = torch.utils.data.sampler.SubsetRandomSampler(
cent_group[0])
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler)
# lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t, args.nce_m)
# criterion = NCECriterion(ndata).cuda()
# lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t, args.nce_m)
if args.tsne_grid:
tsne_grid(val_loader, model)
if args.h_cluster:
for size in range(2, 3):
# size = 20
kmeans, topk = kmean(lemniscate, size, 500, 10, train_dataset)
respred = torch.tensor([]).cuda()
lab, idx = [[] for i in range(2)]
num = 0
'''
for p,index,label in pred:
respred = torch.cat((respred,p))
if num == 0:
lab = label
else:
lab += label
idx.append(index)
num+=1
'''
h_cluster(lemniscate, train_dataset, kmeans, topk,
size) #, respred, lab, idx)
# axis_explore(lemniscate, train_dataset)
# kmeans_opt(lemniscate, 5)
if args.select:
if not args.select_load:
pred = []
if args.select_size:
size = int(args.select_size * ndata)
else:
size = round(ndata / 100.0)
sub_sample = np.random.randint(0, ndata, size=size)
train_sampler = torch.utils.data.sampler.SubsetRandomSampler(
sub_sample)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler)
pred = div_train(train_loader, model, 0, pred)
pred_features = []
pred_labels = []
pred_idx = []
for inst in pred:
feat, idx, lab = list(inst)
pred_features.append(feat)
pred_labels.append(lab)
pred_idx.append(idx.data.cpu())
if args.select_save:
save_checkpoint(
{
'epoch': args.start_epoch,
'arch': args.arch,
'state_dict': model.state_dict(),
'prediction': pred,
'lemniscate': lemniscate,
'optimizer': optimizer.state_dict(),
}, 'select.pth.tar')
min_idx = selection(pred_features, pred_idx, train_dataset,
args.select_num, args.select_thresh)
train_sampler = torch.utils.data.sampler.SubsetRandomSampler(min_idx)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler)
lemniscate = NCEAverage(args.low_dim, ndata, 20, args.nce_t,
args.nce_m)
optimizer = torch.optim.SGD(model.parameters(),
0.1,
momentum=0.1,
weight_decay=0.00001)
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.static_loss,
verbose=False)
for epoch in range(50):
if args.distributed:
train_sampler.set_epoch(epoch)
adjust_learning_rate(optimizer, epoch)
if epoch % 1 == 0:
save_checkpoint({
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'lemniscate': lemniscate,
'optimizer': optimizer.state_dict(),
})
train(train_loader, model, lemniscate, criterion, optimizer, epoch)
train_sampler = torch.utils.data.sampler.SubsetRandomSampler(sub_index)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler)
lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t,
args.nce_m)
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.static_loss,
verbose=False)
if args.kmeans_opt:
kmeans_opt(lemniscate, 500)
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
adjust_learning_rate(optimizer, epoch)
if epoch % 1 == 0:
# evaluate on validation set
#prec1 = NN(epoch, model, lemniscate, train_loader, train_loader) # was evaluating on train
# prec1 = kNN(model, lemniscate, train_loader, val_loader, args.K, args.nce_t)
# prec1 really should be renamed to prec5 as kNN now returns top5 score, but
# it won't be backward's compatible as earlier models were saved with "best_prec1"
# remember best prec@1 and save checkpoint
# is_best = prec1 > best_prec1
# best_prec1 = max(prec1, best_prec1)
save_checkpoint({
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'lemniscate': lemniscate,
# 'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}) # , is_best)
# train for one epoch
train(train_loader, model, lemniscate, criterion, optimizer, epoch)
# kmeans,cent = kmeans()
# group_train(train_loader, model, lemniscate, criterion, optimizer, epoch, kmeans, cent)
# print elapsed time
end_train_time = datetime.datetime.now()
d = end_train_time - begin_train_time
print(
"Trained for %d epochs. Elapsed time: %s days, %.2dh: %.2dm: %.2ds" %
(len(range(args.start_epoch, args.epochs)), d.days, d.seconds // 3600,
(d.seconds // 60) % 60, d.seconds % 60))
testloader = torch.utils.data.DataLoader(testset,
batch_size=int(args.b / 2.),
shuffle=False,
num_workers=2)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse',
'ship', 'truck')
ndata = trainset.__len__()
print('==> Building model..')
net = models.__dict__['ResNet18'](low_dim=args.low_dim)
# define leminiscate
lemniscate = NCEAverage(args.low_dim,
ndata,
args.nce_t,
args.nce_m,
batchSize=args.b)
# define loss function
criterion = NCECriterion()
if device == 'cuda':
net = torch.nn.DataParallel(net.cuda())
lemniscate.cuda()
criterion.cuda()
cudnn.benchmark = True
# Model
if args.test_only or len(args.resume) > 0:
# Load checkpoint.
print('==> Resuming from checkpoint..')
use_gpu = torch.cuda.is_available()
# get model and replace the original fc layer with your fc layer
model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, dataset_sizes)
#model_ft.load_state_dict(torch.load(P_PATH))
if use_gpu:
model_ft = model_ft.cuda()
# define loss function
#criterion = nn.CrossEntropyLoss()
#criterion = NEG_loss(dataset_sizes, dataset_sizes)
lemniscate = NCEAverage(dataset_sizes, dataset_sizes, 100, 0.07, 0.5)
criterion = NCECriterion(dataset_sizes)
# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)
# Decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft,
step_size=7,
gamma=0.1)
model_ft = train_model(model=model_ft,
criterion=criterion,
optimizer=optimizer_ft,
scheduler=exp_lr_scheduler,
num_epochs=60)
shuffle=False)
else:
testloader = torch.utils.data.DataLoader(testset,
batch_size=100,
shuffle=False,
num_workers=4)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse',
'ship', 'truck')
ndata = trainset.__len__()
print('==> Building model..')
net = models.__dict__['ResNet18'](low_dim=args.low_dim)
# define leminiscate
if args.nce_k > 0:
lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t,
args.nce_m, None, lib.get_dev())
else:
lemniscate = LinearAverage(args.low_dim, ndata, args.nce_t, args.nce_m)
LA = LinearAverage(args.low_dim, ndata, args.nce_t, args.nce_m)
if device == 'cuda':
# net = torch.nn.DataParallel(net, device_ids=range(torch.cuda.device_count()))
cudnn.benchmark = True
# Model
if args.test_only or len(args.resume) > 0:
# Load checkpoint.
print('==> Resuming from checkpoint..')
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/' + args.resume)
net.load_state_dict(checkpoint['net'])
def main():
global args, best_prec1
args = parser.parse_args()
args.distributed = args.world_size > 1
if args.distributed:
dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
world_size=args.world_size)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_transform = transforms.Compose(
[
transforms.RandomResizedCrop(224, scale=(0.3, 1.)),
transforms.RandomGrayscale(p=0.5),
transforms.ColorJitter(0.5, 0.5, 0.5, 0.5),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(),
normalize])
val_transform = transforms.Compose(
[
transforms.Resize((224, 224)),
transforms.ToTensor(),
normalize])
train_dataset = Foundation_Type_Binary(args.train_data, transform=train_transform, mask_buildings=args.mask_buildings)
val_dataset = Foundation_Type_Binary(args.val_data, transform=val_transform, mask_buildings=args.mask_buildings)
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True)
model = resnet50(low_dim=args.low_dim)
model = torch.nn.DataParallel(model).cuda()
print ('Train dataset instances: {}'.format(len(train_loader.dataset)))
print ('Val dataset instances: {}'.format(len(val_loader.dataset)))
ndata = train_dataset.__len__()
if args.nce_k > 0:
lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t, args.nce_m).cuda()
criterion = NCECriterion(ndata).cuda()
else:
lemniscate = LinearAverage(args.low_dim, ndata, args.nce_t, args.nce_m).cuda()
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(), args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
args.epochs = args.start_epoch + args.epochs
best_prec1 = checkpoint['best_prec1']
missing_keys, unexpected_keys = model.load_state_dict(checkpoint['state_dict'], strict=False)
if len(missing_keys) or len(unexpected_keys):
print('Warning: Missing or unexpected keys found.')
print('Missing: {}'.format(missing_keys))
print('Unexpected: {}'.format(unexpected_keys))
low_dim_checkpoint = checkpoint['lemniscate'].memory.shape[1]
if low_dim_checkpoint == args.low_dim:
lemniscate = checkpoint['lemniscate']
else:
print('Chosen low dim does not fit checkpoint. Assuming fine-tuning and not loading memory bank.')
try:
optimizer.load_state_dict(checkpoint['optimizer'])
except ValueError:
print('Training optimizer does not fit to checkpoint optimizer. Assuming fine-tuning and load optimizer from scratch. ')
print("=> loaded checkpoint '{}' (epoch {})"
.format(args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
if args.evaluate:
kNN(0, model, lemniscate, train_loader, val_loader, 200, args.nce_t)
return
prec1 = NN(0, model, lemniscate, train_loader, val_loader)
print('Start out precision: {}'.format(prec1))
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, lemniscate, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = NN(epoch, model, lemniscate, train_loader, val_loader)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'lemniscate': lemniscate,
'best_prec1': best_prec1,
'optimizer' : optimizer.state_dict(),
}, is_best, args.name)
download=True,
transform=transform_test)
testloader = torch.utils.data.DataLoader(testset,
batch_size=100,
shuffle=False,
num_workers=2)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse',
'ship', 'truck')
ndata = trainset.__len__()
print('==> Building model..')
net = models.__dict__['ResNet18'](low_dim=args.low_dim)
# define leminiscate
if args.nce_k > 0:
lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t,
args.nce_m)
else:
lemniscate = LinearAverage(args.low_dim, ndata, args.nce_t, args.nce_m)
net = torch.nn.DataParallel(net)
cudnn.benchmark = True
# Model
if args.test_only or len(args.resume) > 0:
# Load checkpoint.
print('==> Resuming from checkpoint..')
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/' + args.resume)
net.load_state_dict(checkpoint['net'])
lemniscate = checkpoint['lemniscate']
best_acc = checkpoint['acc']
# Model
if args.test_only or len(args.resume)>0:
# Load checkpoint.
print('==> Resuming from checkpoint..')
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/'+args.resume)
net = checkpoint['net']
lemniscate = checkpoint['lemniscate']
best_acc = checkpoint['acc']
start_epoch = checkpoint['epoch']
else:
print('==> Building model..')
net = models.__dict__['ResNet18'](low_dim=args.low_dim)
# define leminiscate
if args.nce_k > 0:
lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t, args.nce_m)
else:
lemniscate = LinearAverage(args.low_dim, ndata, args.nce_t, args.nce_m)
# define loss function
if hasattr(lemniscate, 'K'):
criterion = NCECriterion(ndata)
else:
criterion = nn.CrossEntropyLoss()
if use_cuda:
net.cuda()
net = torch.nn.DataParallel(net, device_ids=range(torch.cuda.device_count()))
lemniscate.cuda()
criterion.cuda()
cudnn.benchmark = True
def build_model():
best_acc = 0 # best test accuracy
start_epoch = 0 # start from epoch 0 or last checkpoint epoch
if args.architecture == 'resnet18':
net = models.__dict__['resnet18_cifar'](low_dim=args.low_dim)
elif args.architecture == 'wrn-28-2':
net = models.WideResNet(depth=28,
num_classes=args.low_dim,
widen_factor=2,
dropRate=0).to(args.device)
elif args.architecture == 'wrn-28-10':
net = models.WideResNet(depth=28,
num_classes=args.low_dim,
widen_factor=10,
dropRate=0).to(args.device)
# define leminiscate
if args.nce_k > 0:
lemniscate = NCEAverage(args.low_dim, args.ndata, args.nce_k,
args.nce_t, args.nce_m)
else:
lemniscate = LinearAverage(args.low_dim, args.ndata, args.nce_t,
args.nce_m)
if args.device == 'cuda':
net = torch.nn.DataParallel(net,
device_ids=range(
torch.cuda.device_count()))
cudnn.benchmark = True
optimizer = optim.SGD(net.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=args.weight_decay,
nesterov=True)
# Model
if args.test_only or len(args.resume) > 0:
# Load checkpoint.
print('==> Resuming from checkpoint..')
checkpoint = torch.load(args.resume)
net.load_state_dict(checkpoint['net'])
optimizer.load_state_dict(checkpoint['optimizer'])
lemniscate = checkpoint['lemniscate']
best_acc = checkpoint['acc']
start_epoch = checkpoint['epoch'] + 1
if args.lr_scheduler == 'multi-step':
if args.epochs == 200:
steps = [60, 120, 160]
elif args.epochs == 600:
steps = [180, 360, 480, 560]
else:
raise RuntimeError(
f"need to config steps for epoch = {args.epochs} first.")
scheduler = lr_scheduler.MultiStepLR(optimizer,
steps,
gamma=0.2,
last_epoch=start_epoch - 1)
elif args.lr_scheduler == 'cosine':
scheduler = lr_scheduler.CosineAnnealingLR(optimizer,
args.epochs,
eta_min=0.00001,
last_epoch=start_epoch - 1)
elif args.lr_scheduler == 'cosine-with-restart':
scheduler = CosineAnnealingLRWithRestart(optimizer,
eta_min=0.00001,
last_epoch=start_epoch - 1)
else:
raise ValueError("not supported")
# define loss function
if hasattr(lemniscate, 'K'):
criterion = NCECriterion(args.ndata)
else:
criterion = nn.CrossEntropyLoss()
net.to(args.device)
lemniscate.to(args.device)
criterion.to(args.device)
return net, lemniscate, optimizer, criterion, scheduler, best_acc, start_epoch
def main():
global args, best_prec1, min_avgloss
args = parser.parse_args()
input("Begin the {} time's training".format(args.train_time))
writer_log_dir = "/data/fhz/unsupervised_recommendation/idfe_runs/idfe_train_time:{}".format(
args.train_time)
writer = SummaryWriter(log_dir=writer_log_dir)
if args.dataset == "lung":
# build dataloader,val_dloader will be build in test function
model = idfe.IdFe3d(feature_dim=args.latent_dim)
model.encoder = torch.nn.DataParallel(model.encoder)
model.linear_map = torch.nn.DataParallel(model.linear_map)
model = model.cuda()
train_datalist, test_datalist = multi_cross_validation()
ndata = len(train_datalist)
elif args.dataset == "gland":
dataset_path = "/data/fhz/MICCAI2015/npy"
model = idfe.IdFe2d(feature_dim=args.latent_dim)
model.encoder = torch.nn.DataParallel(model.encoder)
model.linear_map = torch.nn.DataParallel(model.linear_map)
model = model.cuda()
train_datalist = glob(path.join(dataset_path, "train", "*.npy"))
ndata = len(train_datalist)
else:
raise FileNotFoundError("Dataset {} Not Found".format(args.dataset))
if args.nce_k > 0:
"""
Here we use NCE to calculate loss
"""
lemniscate = NCEAverage(args.latent_dim, ndata, args.nce_k, args.nce_t,
args.nce_m).cuda()
criterion = NCECriterion(ndata).cuda()
else:
lemniscate = LinearAverage(args.latent_dim, ndata, args.nce_t,
args.nce_m).cuda()
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
min_avgloss = checkpoint['min_avgloss']
model.encoder.load_state_dict(checkpoint['encoder_state_dict'])
model.linear_map.load_state_dict(
checkpoint['linear_map_state_dict'])
lemniscate = checkpoint['lemniscate']
optimizer.load_state_dict(checkpoint['optimizer'])
train_datalist = checkpoint['train_datalist']
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
if args.dataset == "lung":
train_dset = LungDataSet(data_path_list=train_datalist,
augment_prob=args.aug_prob)
train_dloader = DataLoader(dataset=train_dset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
elif args.dataset == "gland":
train_dset = GlandDataset(data_path_list=train_datalist,
need_seg_label=False,
augment_prob=args.aug_prob)
train_dloader = DataLoader(dataset=train_dset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
else:
raise FileNotFoundError("Dataset {} Not Found".format(args.dataset))
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
epoch_loss = train(train_dloader,
model=model,
lemniscate=lemniscate,
criterion=criterion,
optimizer=optimizer,
epoch=epoch,
writer=writer,
dataset=args.dataset)
if (epoch + 1) % 5 == 0:
if args.dataset == "lung":
"""
Here we define the best point as the minimum average epoch loss
"""
accuracy = list([])
# for i in range(5):
# train_feature = lemniscate.memory.clone()
# test_datalist = train_datalist[five_cross_idx[i]:five_cross_idx[i + 1]]
# test_feature = train_feature[five_cross_idx[i]:five_cross_idx[i + 1], :]
# train_indices = [train_datalist.index(d) for d in train_datalist if d not in test_datalist]
# tmp_train_feature = torch.index_select(train_feature, 0, torch.tensor(train_indices).cuda())
# tmp_train_datalist = [train_datalist[i] for i in train_indices]
# test_label = np.array(
# [int(eval(re.match("(.*)_(.*)_annotations.npy", path.basename(raw_cube_path)).group(2)) > 3)
# for raw_cube_path in test_datalist], dtype=np.float)
# tmp_train_label = np.array(
# [int(eval(re.match("(.*)_(.*)_annotations.npy", path.basename(raw_cube_path)).group(2)) > 3)
# for raw_cube_path in tmp_train_datalist], dtype=np.float)
# accuracy.append(
# kNN(tmp_train_feature, tmp_train_label, test_feature, test_label, K=20, sigma=1 / 10))
# accuracy = mean(accuracy)
is_best = (epoch_loss < min_avgloss)
min_avgloss = min(epoch_loss, min_avgloss)
save_checkpoint(
{
'epoch': epoch + 1,
"train_time": args.train_time,
"encoder_state_dict": model.encoder.state_dict(),
"linear_map_state_dict": model.linear_map.state_dict(),
'lemniscate': lemniscate,
'min_avgloss': min_avgloss,
'dataset': args.dataset,
'optimizer': optimizer.state_dict(),
'train_datalist': train_datalist
}, is_best)
# knn_text = "In epoch :{} the five cross validation accuracy is :{}".format(epoch, accuracy * 100.0)
# # print(knn_text)
# writer.add_text("knn/text", knn_text, epoch)
# writer.add_scalar("knn/accuracy", accuracy, global_step=epoch)
elif args.dataset == "gland":
is_best = (epoch_loss < min_avgloss)
min_avgloss = min(epoch_loss, min_avgloss)
save_checkpoint(
{
'epoch': epoch + 1,
"train_time": args.train_time,
"encoder_state_dict": model.encoder.state_dict(),
"linear_map_state_dict": model.linear_map.state_dict(),
'lemniscate': lemniscate,
'min_avgloss': min_avgloss,
'dataset': args.dataset,
'optimizer': optimizer.state_dict(),
'train_datalist': train_datalist,
}, is_best)