def trainloader_dct_resized(args):
traindir = os.path.join(args.data, 'train')
train_dataset = ImageFolderDCT(traindir, transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.TransformDCT(), # 28x28x192
transforms.DCTFlatten2D(),
transforms.UpsampleDCT(upscale_ratio_h=4, upscale_ratio_w=4, debug=False),
transforms.ToTensorDCT(),
transforms.SubsetDCT(channels=args.subset),
transforms.Aggregate(),
transforms.NormalizeDCT(
train_dct_subset_mean,
train_dct_subset_std,
channels=args.subset
)
]))
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.train_batch, shuffle=(train_sampler is None),
num_workers=args.workers, pin_memory=True, sampler=train_sampler)
train_loader_len = len(train_loader)
return train_loader, train_sampler, train_loader_len
def get_mean_and_std_dct_resized(dataset, model='mobilenet'):
'''Compute the mean and std value of dataset.'''
# Dataset = ImageFolderDCT(dataset, transforms.Compose([
# transforms.TransformDCT(), # 28x28x192
# transforms.DCTFlatten2D(),
# transforms.UpsampleDCT(upscale_ratio_h=4, upscale_ratio_w=4, debug=False),
# transforms.ToTensorDCT(),
# transforms.Average()
# ]), aggregate=True)
# Dataset = ImageFolderDCT(dataset, transforms.Compose([
# transforms.Upscale(upscale_factor=2),
# transforms.TransformUpscaledDCT(),
# transforms.ToTensorDCT(),
# transforms.Aggregate(),
# transforms.Average()
# ]))
if model == 'mobilenet':
input_size = 896
batchsize = 256
elif model == 'resnet':
input_size = 448
batchsize = 128
else:
raise NotImplementedError
Dataset = ImageFolderDCT(dataset, transforms.Compose([
transforms.DCTFlatten2D(),
transforms.UpsampleDCT(size_threshold=input_size, T=input_size, debug=False),
# transforms.UpsampleDCT(size_threshold=112 * 8, T=112 * 8, debug=False),
transforms.Aggregate2(),
# transforms.RandomResizedCropDCT(112),
# transforms.RandomHorizontalFlip(),
transforms.ToTensorDCT2(),
transforms.Average()
]), backend='dct')
dataloader = torch.utils.data.DataLoader(Dataset, batch_size=batchsize, pin_memory=True, shuffle=False, num_workers=16)
mean, std = torch.zeros(192), torch.zeros(192)
print('==> Computing mean and std..')
# end = time.time()
for i, (inputs, targets) in enumerate(dataloader):
# print('data time: {}'.format(time.time()-end))
print('{}/{}'.format(i, len(dataloader)))
mean += inputs.mean(dim=0)
std += inputs.std(dim=0)
# end = time.time()
mean.div_(i+1)
std.div_(i+1)
return mean, std
def folder2lmdb(dpath, name="train", write_frequency=1):
directory = osp.expanduser(osp.join(dpath, name))
print("Loading dataset from %s" % directory)
dataset = ImageFolderDCT('/ILSVRC2012/train',
transforms.Compose([
transforms.DCTFlatten2D(),
transforms.UpsampleDCT(upscale_ratio_h=4,
upscale_ratio_w=4,
debug=False),
transforms.ToTensorDCT(),
transforms.SubsetDCT(channels=32),
]),
backend='dct')
data_loader = torch.utils.data.DataLoader(
dataset,
num_workers=0,
)
lmdb_path = osp.join(dpath, "%s.lmdb" % name)
isdir = os.path.isdir(lmdb_path)
print("Generate LMDB to %s" % lmdb_path)
db = lmdb.open(
lmdb_path,
subdir=isdir,
map_size=1281167 * 224 * 224 * 32 * 10,
readonly=False,
# map_size=1099511627776 * 2, readonly=False,
meminit=False,
map_async=True)
txn = db.begin(write=True)
for idx, (image, label) in enumerate(data_loader):
image = image.numpy()
label = label.numpy()
txn.put(u'{}'.format(idx).encode('ascii'),
dumps_pyarrow((bz2.compress(image), label)))
if idx % write_frequency == 0:
print("[%d/%d]" % (idx, len(data_loader)))
txn.commit()
txn = db.begin(write=True)
# finish iterating through dataset
txn.commit()
keys = [u'{}'.format(k).encode('ascii') for k in range(idx + 1)]
with db.begin(write=True) as txn:
txn.put(b'__keys__', dumps_pyarrow(keys))
txn.put(b'__len__', dumps_pyarrow(len(keys)))
print("Flushing database ...")
db.sync()
db.close()
def valloader_dct_resized(args):
valdir = os.path.join(args.data, 'val')
val_loader = torch.utils.data.DataLoader(
ImageFolderDCT(valdir, transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.TransformDCT(), # 28x28x192
transforms.DCTFlatten2D(),
transforms.UpsampleDCT(upscale_ratio_h=4, upscale_ratio_w=4, debug=False),
transforms.ToTensorDCT(),
transforms.SubsetDCT(channels=args.subset),
transforms.Aggregate(),
transforms.NormalizeDCT(
train_dct_subset_mean,
train_dct_subset_std,
channels=args.subset
)
])),
batch_size=args.test_batch, shuffle=False,
num_workers=args.workers, pin_memory=True)
return val_loader
std=train_y_std_resized)
input_normalize_cb = transforms.Normalize(mean=train_cb_mean_resized,
std=train_cb_std_resized)
input_normalize_cr = transforms.Normalize(mean=train_cr_mean_resized,
std=train_cr_std_resized)
input_normalize.append(input_normalize_y)
input_normalize.append(input_normalize_cb)
input_normalize.append(input_normalize_cr)
val_loader = torch.utils.data.DataLoader(
# ImageFolderDCT('/mnt/ssd/kai.x/dataset/ILSVRC2012/val', transforms.Compose([
ImageFolderDCT(
'/storage-t1/user/kaixu/datasets/ILSVRC2012/val',
transforms.Compose([
transforms.ToYCrCb(),
transforms.TransformDCT(),
transforms.UpsampleDCT(T=896, debug=False),
transforms.CenterCropDCT(112),
transforms.ToTensorDCT(),
transforms.NormalizeDCT(train_y_mean_resized,
train_y_std_resized,
train_cb_mean_resized,
train_cb_std_resized,
train_cr_mean_resized,
train_cr_std_resized),
])),
batch_size=1,
shuffle=False,
num_workers=1,
pin_memory=False)
# train_dataset = ImageFolderDCT('/mnt/ssd/kai.x/dataset/ILSVRC2012/train', transforms.Compose([
def get_mean_and_std_dct(dataset, sublist=None):
'''Compute the mean and std value of dataset.'''
import datasets.cvtransforms as transforms
# jpeg_encoder = TurboJPEG('/home/kai.x/work/local/lib/libturbojpeg.so')
# Dataset = ImageFolderDCT(dataset, transforms.Compose([
# # transforms.RandomResizedCrop(224),
# # # transforms.RandomHorizontalFlip(),
# # transforms.TransformDCT(),
# # transforms.UpsampleDCT(896),
# # transforms.ToTensorDCT()
# # # transforms.RandomResizedCrop(256),
# # # transforms.RandomHorizontalFlip(),
# # # transforms.TransformDCT(),
# # # transforms.UpsampleDCT(256, 256),
# # # transforms.CenterCropDCT(112),
# # # transforms.ToTensorDCT()
# # ]))
# Dataset = ImageFolderDCT(dataset, transforms.Compose([
# transforms.Upscale(),
# transforms.TransformDCT(),
# transforms.ToTensorDCT(),
# ]))
Dataset = ImageFolderDCT(dataset, transforms.Compose([
transforms.DCTFlatten2D(),
transforms.UpsampleDCT(upscale_ratio_h=4, upscale_ratio_w=4, debug=False),
transforms.ToTensorDCT(),
transforms.Average()
]), backend='dct')
dataloader = torch.utils.data.DataLoader(Dataset, batch_size=1, pin_memory=True, shuffle=False, num_workers=0)
mean_y, mean_cb, mean_cr = torch.zeros(64), torch.zeros(64), torch.zeros(64)
std_y, std_cb, std_cr = torch.zeros(64), torch.zeros(64), torch.zeros(64)
print('==> Computing mean and std..')
end = time.time()
if sublist is None:
for i, (inputs_y, inputs_cb, inputs_cr, targets) in enumerate(dataloader):
print('data time: {}'.format(time.time()-end))
print('{}/{}'.format(i, len(dataloader)))
mean_y += inputs_y.mean(dim=0)
std_y += inputs_y.std(dim=0)
mean_cb += inputs_cb.mean(dim=0)
std_cb += inputs_cb.std(dim=0)
mean_cr += inputs_cr.mean(dim=0)
std_cr += inputs_cr.std(dim=0)
end = time.time()
mean_y.div_(i+1)
std_y.div_(i+1)
mean_cb.div_(i+1)
std_cb.div_(i+1)
mean_cr.div_(i+1)
std_cr.div_(i+1)
else:
dataloader_iterator = iter(dataloader)
for i in range(sublist):
try:
inputs_y, inputs_cb, inputs_cr, targets = next(dataloader_iterator)
except:
print('error')
print('{}/{}'.format(i, sublist))
for i in range(64):
mean_y[i] += inputs_y[:, i, :, :].mean()
std_y[i] += inputs_y[:, i, :, :].std()
mean_cb[i] += inputs_cb[:, i, :, :].mean()
std_cb[i] += inputs_cb[:, i, :, :].std()
mean_cr[i] += inputs_cr[:, i, :, :].mean()
std_cr[i] += inputs_cr[:, i, :, :].std()
mean_y.div_(sublist)
std_y.div_(sublist)
mean_cb.div_(sublist)
std_cb.div_(sublist)
mean_cr.div_(sublist)
std_cr.div_(sublist)
return mean_y, std_y, mean_cb, std_cb, mean_cr, std_cr
# transforms.SubsetDCT2(channels=32),
# transforms.Aggregate2(),
# transforms.RandomResizedCropDCT(112),
# transforms.RandomHorizontalFlip(),
# transforms.ToTensorDCT2(),
# transforms.NormalizeDCT(
# train_upscaled_static_mean,
# train_upscaled_static_std,
# channels=32
# )
# ])
transform6 = transforms.Compose([
transforms.DCTFlatten2D(mux=0b011),
transforms.UpsampleCbCr(),
transforms.UpsampleDCT(T=512, debug=False),
transforms.SubsetDCT2(channels=64),
transforms.Aggregate2(),
transforms.CenterCrop(448 // 8),
transforms.ToTensorDCT2(),
transforms.NormalizeDCT(
train_upscaled_static_dct_direct_mean,
train_upscaled_static_dct_direct_std,
channels=64
)
])
transform7 = transforms.Compose([
transforms.UpsampleCbCr(),
transforms.SubsetDCT2(channels=64),
transforms.RandomResizedCropDCT(size=448),
transforms.RandomHorizontalFlip(),
transforms.Upscale(upscale_factor=2),
transforms.TransformUpscaledDCT(),
transforms.ToTensorDCT(),
transforms.SubsetDCT(channels='24'),
transforms.Aggregate(),
transforms.NormalizeDCT(
train_upscaled_static_mean,
train_upscaled_static_std,
channels='24'
)
])
transform5 = transforms.Compose([
transforms.DCTFlatten2D(),
transforms.UpsampleDCT(size_threshold=112 * 8, T=112 * 8, debug=False),
transforms.SubsetDCT2(channels='32'),
transforms.Aggregate2(),
transforms.RandomResizedCropDCT(112),
transforms.RandomHorizontalFlip(),
transforms.ToTensorDCT2(),
transforms.NormalizeDCT(
train_upscaled_static_mean,
train_upscaled_static_std,
channels='32'
)
])
# train_dataset = ImageFolderDCT('/ILSVRC2012/train', transform1, backend='opencv')
# train_dataset = ImageFolderDCT('/ILSVRC2012/train', transform2
# , backend='dct')
train_dataset = ImageFolderDCT('/ILSVRC2012/train', transform5, backend='dct')