args.scale_range),
ratio=(1, 1))
cropper['val'] = cropper['train']
else:
cropper['train'] = transforms.RandomCrop((args.im_dim, args.im_dim))
cropper['val'] = transforms.CenterCrop((args.im_dim, args.im_dim))
# we will train with data augmentation using random crops, flips, and jittering
# we will validate with central crops
data_transforms = {
'train':
transforms.Compose([
transforms.Resize(int((256 / 224) * args.im_dim)), cropper['train'],
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=.5,
contrast=.5,
saturation=.5,
hue=.5),
transforms.ToTensor(), normalize
]),
'val':
transforms.Compose([
transforms.Resize(int((256 / 224) * args.im_dim)), cropper['val'],
transforms.ToTensor(), normalize
])
}
# load the entry level info for imagenet
with open(f'{DATA_DIR}/imagesets/entrylevel-all_imagenet_info.pkl', 'rb') as f:
entry_dict = pickle.load(f)
if 'subset-' in opt['dataset']:
data_transforms['test'] = data_transforms['val']
phases = ['train', 'val', 'test']
else:
phases = ['train', 'val']
id_phases = ['train', 'val']
if 'gfmt' in opt['dataset'].lower():
id_phases = []
data_transforms = {
'val':
transforms.Compose([
transforms.RandomRotation(degrees=30),
transforms.ColorJitter(brightness=.5,
contrast=.5,
saturation=.5,
hue=.5),
transforms.Pad(padding=(0, 0), padding_mode='square_constant'),
transforms.Resize((opt['im_dim'], opt['im_dim'])),
transforms.ToTensor(), normalize
])
}
image_datasets = {
x: MyImageFolder(os.path.join(data_dir), data_transforms[x])
for x in ['val']
}
dataloaders = {
x: torch.utils.data.DataLoader(image_datasets[x],
batch_size=opt['batch_size'],
shuffle=False,
num_workers=args.num_workers)
for x in ['val']
import numbers
import platform
from contextlib import contextmanager
from matplotlib import colors
from familiarity.transforms import functional as F
import familiarity.transforms as transforms
NORMALIZE = transforms.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]
)
DATA_TRANSFORM = transforms.Compose([
transforms.Pad(padding=(0,0), padding_mode='square_constant'),
transforms.Resize((224, 224)),
transforms.ToTensor(),
NORMALIZE,
])
pltv = platform.version()
ON_CLUSTER = False if 'Darwin' in pltv or 'Ubuntu' in pltv else True
def get_layers_of_interest(net):
if 'vgg16' in net:
layers_of_interest = [0, 5, 10, 17, 24, 31, 33, 36, 38, 39]
layer_names = ['image', 'conv1', 'conv2', 'conv3', 'conv4', 'conv5', 'fc6', 'fc7', 'fc8', 'prob']
elif 'vgg_m_face_bn_dag' in net:
layers_of_interest = [0, 4, 8, 11, 14, 18, 21, 24, 25] #0th layer is the image
layer_names = ['image', 'conv1', 'conv2', 'conv3', 'conv4', 'conv5', 'FC6', 'FC7', 'class-prob']
elif 'vgg_face_dag' in net:
if not opt['no_gpu']:
torch.cuda.manual_seed_all(1)
torch.backends.cudnn.deterministic = True
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# we will train with data augmentation using random crops, flips, and jittering
# we will validate with central crops
data_transforms = {
'train':
transforms.Compose([
transforms.Resize(256),
transforms.RandomCrop((224, 224)),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=.5,
contrast=.5,
saturation=.5,
hue=.5),
transforms.ToTensor(), normalize
]),
'val':
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop((224, 224)),
transforms.ToTensor(), normalize
]),
'test':
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop((224, 224)),
transforms.ToTensor(), normalize
raise NotImplementedError('net:{} not implemented yet'.format(
opt['net']))
#reproducible results
torch.random.manual_seed(1)
np.random.seed(1)
if not opt['no_gpu']:
torch.cuda.manual_seed_all(1)
torch.backends.cudnn.deterministic = True
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
data_transform = transforms.Compose([
transforms.Pad(padding=(0, 0), padding_mode='square_constant'),
transforms.Resize((opt['im_dim'], opt['im_dim'])),
transforms.ToTensor(),
normalize,
])
data_dir, _ = get_data_dir(opt['dataset'],
opt['datasubset'],
exclude_keys,
id_thresh=opt['id_thresh'],
n_val=opt['n_val'])
image_dataset = MyImageFolder(data_dir, data_transform)
dataloader = torch.utils.data.DataLoader(image_dataset,
batch_size=opt['batch_size'],
shuffle=False,
num_workers=4)
dataset_size = len(image_dataset)
class_names = image_dataset.classes