def create_iterator(opt, mode):
if opt.dataset.startswith('CIFAR'):
convert = tnt.transform.compose([
lambda x: x.astype(np.float32),
T.Normalize([125.3, 123.0, 113.9], [63.0, 62.1, 66.7]),
lambda x: x.transpose(2,0,1),
torch.from_numpy,
])
train_transform = tnt.transform.compose([
T.RandomHorizontalFlip(),
T.Pad(opt.randomcrop_pad, cv2.BORDER_REFLECT),
T.RandomCrop(32),
convert,
])
ds = getattr(datasets, opt.dataset)(opt.dataroot, train=mode, download=True)
smode = 'train' if mode else 'test'
ds = tnt.dataset.TensorDataset([getattr(ds, smode + '_data'),
getattr(ds, smode + '_labels')])
ds = ds.transform({0: train_transform if mode else convert})
return ds.parallel(batch_size=opt.batchSize, shuffle=mode,
num_workers=opt.nthread, pin_memory=True)
elif opt.dataset == 'ImageNet':
def cvload(path):
img = cv2.imread(path, cv2.IMREAD_COLOR)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
return img
convert = tnt.transform.compose([
lambda x: x.astype(np.float32) / 255.0,
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
lambda x: x.transpose(2, 0, 1).astype(np.float32),
torch.from_numpy,
])
print("| setting up data loader...")
if mode:
traindir = os.path.join(opt.dataroot, 'train')
ds = datasets.ImageFolder(traindir, tnt.transform.compose([
T.RandomSizedCrop(224),
T.RandomHorizontalFlip(),
convert,
]), loader=cvload)
else:
valdir = os.path.join(opt.dataroot, 'val')
ds = datasets.ImageFolder(valdir, tnt.transform.compose([
T.Scale(256),
T.CenterCrop(224),
convert,
]), loader=cvload)
return torch.utils.data.DataLoader(ds,
batch_size=opt.batchSize, shuffle=mode,
num_workers=opt.nthread, pin_memory=False)
else:
raise ValueError('dataset not understood')
def create_dataset(opt, mode):
convert = tnt.transform.compose([
lambda x: x.astype(np.float32),
T.Normalize(np.asarray([125.3, 123.0, 113.9], dtype=np.float32),
np.asarray([63.0, 62.1, 66.7], dtype=np.float32)),
lambda x: x.transpose(2, 0, 1),
torch.from_numpy,
])
if mode:
convert = tnt.transform.compose([
T.RandomHorizontalFlip(),
T.Pad(opt.randomcrop_pad, cv2.BORDER_REFLECT),
T.RandomCrop(32),
convert,
])
ds = getattr(datasets, opt.dataset)(opt.dataroot,
train=mode,
download=True)
smode = 'train' if mode else 'test'
ds = tnt.dataset.TensorDataset(
[getattr(ds, smode + '_data'),
getattr(ds, smode + '_labels')])
return ds.transform({0: convert})
def main():
# parse input arguments
args = parser.parse_args()
def cvload(path):
img = cv2.imread(path, cv2.IMREAD_COLOR)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
return img
# set up data loader
print("| setting up data loader...")
valdir = os.path.join(args.imagenetpath, 'val')
ds = datasets.ImageFolder(valdir, tnt.transform.compose([
cvtransforms.Scale(256),
cvtransforms.CenterCrop(224),
lambda x: x.astype(np.float32) / 255.0,
cvtransforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
lambda x: x.transpose(2,0,1).astype(np.float32),
torch.from_numpy,
]), loader = cvload)
train_loader = torch.utils.data.DataLoader(ds,
batch_size=256, shuffle=False,
num_workers=args.numthreads, pin_memory=False)
params = hkl.load(args.model)
params = {k: Variable(torch.from_numpy(v).cuda()) for k,v in params.iteritems()}
f = define_model(params)
class_err = tnt.meter.ClassErrorMeter(topk=[1,5], accuracy=True)
for sample in tqdm(train_loader):
inputs = Variable(sample[0].cuda(), volatile=True)
targets = sample[1]
class_err.add(f(inputs, params).data, targets)
print 'Validation top1/top5 accuracy:'
print class_err.value()
def get_iterator(opt, mode):
def cvload(path):
img = cv2.imread(path, cv2.IMREAD_COLOR)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
return img
convert = tnt.transform.compose([
lambda x: x.astype(np.float32) / 255.0,
cvtransforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
lambda x: x.transpose(2, 0, 1).astype(np.float32),
torch.from_numpy,
])
print("| setting up data loader...")
if mode:
traindir = os.path.join(opt.imagenetpath, 'train')
ds = torchvision.datasets.ImageFolder(
traindir,
tnt.transform.compose([
cvtransforms.RandomSizedCrop(224),
cvtransforms.RandomHorizontalFlip(),
convert,
]),
loader=cvload)
else:
valdir = os.path.join(opt.imagenetpath, 'val')
ds = torchvision.datasets.ImageFolder(valdir,
tnt.transform.compose([
cvtransforms.Scale(256),
cvtransforms.CenterCrop(224),
convert,
]),
loader=cvload)
return DataLoader(ds,
batch_size=opt.batchSize,
shuffle=mode,
num_workers=opt.nthread,
pin_memory=False)
def create_dataset(opt, mode):
convert = tnt.transform.compose([
lambda x: x.astype(np.float32),
cvtransforms.Normalize([125.3, 123.0, 113.9], [63.0, 62.1, 66.7]),
lambda x: x.transpose(2, 0, 1).astype(np.float32),
torch.from_numpy,
])
train_transform = tnt.transform.compose([
cvtransforms.RandomHorizontalFlip(),
cvtransforms.Pad(opt.randomcrop_pad, cv2.BORDER_REFLECT),
cvtransforms.RandomCrop(32),
convert,
])
# opt.data_root = './', opt.dataset = 'CIFAR10'
ds = getattr(datasets, opt.dataset)(opt.data_root,
train=mode,
download=True)
# mode can be True and False
# if mode:
# smode = 'train'
# else:
# smode = 'test'
smode = 'train' if mode else 'test'
# for training data, getattr(ds, 'train_data'): numpy array of size 50000 x 3 x 32 x 32
# getattr(ds, 'train_labels'): list of size 5000, each of value from 0 to 9
ds = tnt.dataset.TensorDataset([
getattr(ds, smode + '_data').transpose(0, 2, 3, 1),
getattr(ds, smode + '_labels')
])
return ds.transform({
0: train_transform if mode else convert
}) # mode is True, use train_transform, mode is False, using convert