def main():
import dataset
X = dataset.ImageFolder('./images')
ind = np.random.permutation(len(X))[:500]
X2 = []
for i in ind:
X2.append(X[i][0].numpy())
X2 = np.array(X2).transpose(1, 0, 2, 3).reshape(3, -1).T
cid, cc = kmeans(X2, 4, 50, -1, 1)
print(cid.shape)
print(cc)
out = {'cc': cc}
np.savez('cc.npz', **out)
def dataloader(batch_size, data_root, shuffle=False, num_workers=8):
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
testset = dataset.ImageFolder(data_root, transform_test)
testloader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=shuffle,
num_workers=num_workers)
return testloader
def get_batches(args, flows):
args.warp = False
dset = dataset.ImageFolder(
is_train=False,
root=args.output_path,
mv_dir='',
args=args
)
batches = []
assert len(flows)==dset.__len__()
for idx in range(dset.__len__()):
_, ctx_frames, fn = dset.__getitem__(idx)
ctx_frames = ctx_frames[None,:,:,:]
# print("frames")
# print(ctx_frames.size())
batches.append((flows[idx], ctx_frames, fn))
return batches
def __init__(self, cfg, logger, model_dir, tensorboard_log_dir, **kwargs):
super(RETrainer, self).__init__(cfg, logger, model_dir,
tensorboard_log_dir, **kwargs)
self.imp_value = torch.ones(cfg['CODE_CHNS']).cuda()
self.order = torch.randperm(cfg['CODE_CHNS']).cuda()
_, self.reorder = torch.sort(self.order)
self.log.info('initialized imp order: ')
self.log.info(self.order)
imp_transform = transforms.Compose([
transforms.RandomCrop((256, 256)),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(),
])
imp_set = dataset.ImageFolder(root=cfg['DATASET']['TRAIN_SET'],
transform=imp_transform)
self.imp_loader = data.DataLoader(dataset=imp_set,
batch_size=1,
shuffle=True,
num_workers=cfg['WORKERS'])
self.end_batch = cfg['RE_END_BATCH']
def get_batches_from_file(args):
# TODO: torch stack batch, ctx frames for better performance
dset = dataset.ImageFolder(
is_train=False,
root=args.output_path,
mv_dir=MV_PATH,
args=args
)
batches = []
for idx in range(dset.__len__()):
batch, ctx_frames, fn = dset.__getitem__(idx)
batch = batch[None, :, :, :]
ctx_frames = ctx_frames[None, :, :, :]
with torch.no_grad():
batch = batch.cuda()
cooked_batch = util.prepare_batch(batch, args.v_compress, args.warp)
_,_,_,flows = cooked_batch
batches.append((flows, ctx_frames, fn))
return batches
parser.add_argument('--iterations',
type=int,
default=24,
help='unroll iterations')
parser.add_argument('--checkpoint', type=int, help='unroll iterations')
args = parser.parse_args()
## load 32x32 patches from images
import dataset
train_transform = transforms.Compose([
transforms.RandomCrop((32, 32)),
#transforms.ToTensor(),
])
train_set = dataset.ImageFolder(root=args.train, transform=train_transform)
train_loader = data.DataLoader(dataset=train_set,
batch_size=args.batch_size,
shuffle=True,
num_workers=1)
print('total images: {}; total batches: {}'.format(len(train_set),
len(train_loader)))
## load networks on GPU
import network
encoder = network.EncoderCell().cuda()
binarizer = network.Binarizer().cuda()
decoder = network.DecoderCell().cuda()
def __init__(self, cfg, logger, model_dir, tensorboard_log_dir, **kwargs):
torch.cuda.set_device(cfg['GPU_DEVICE'])
self.nums_epoch = cfg['TRAIN']['NUM_EPOCH']
self.epoch = 0
self.wr_train_idx = 1
self.wr_eval_idx = 1
self.scale = 16
self.max_psnr = 0.
self.max_ssim = 0.
self.its_bpp = 24.
self.log = logger
self.alpha = cfg['TRAIN']['ALPHA']
self.beta = cfg['TRAIN']['BETA']
self.print_freq = cfg['PRINT_FREQ']
self.quant_level = cfg['QUA_LEVELS']
self.model_dir = model_dir
self.writer = SummaryWriter(tensorboard_log_dir)
self.encoder = Encoder(cfg).cuda()
self.decoder = Decoder(cfg).cuda()
self.quantizer = nn.ModuleList()
self.entropy = nn.ModuleList()
self.clip_value = cfg['TRAIN']['CLIP_VALUE']
self.part = [0]
[
self.part.append(int(x * cfg['CODE_CHNS']) + self.part[i])
for i, x in enumerate(cfg['ENP']['PART'])
]
for i in range(len(cfg['ENP']['PART'])):
self.quantizer.append(
GMMQuantizer(cfg['QUA_LEVELS'][i], cfg['QUA']['STD'][i],
cfg['QUA']['PI'][i]).cuda())
self.entropy.append(
ContextEstimater3d(cfg['QUA_LEVELS'][i],
cfg['ENP']['FEAT_NUMS'][i]).cuda())
self.code_nums = cfg['CODE_CHNS']
self.solver_enc = optim.Adam(self.encoder.parameters(),
lr=cfg['TRAIN']['LR_ENC'])
Q_params = list()
for i in range(len(cfg['QUA_LEVELS'])):
Q_params.append({
'params': self.quantizer[i].mean,
'lr': cfg['TRAIN']['LR_QUA']
})
Q_params.append({
'params': self.quantizer[i].log_pi,
'lr': cfg['TRAIN']['LR_QUA'] * 0.2
})
Q_params.append({
'params': self.quantizer[i].log_std,
'lr': cfg['TRAIN']['LR_QUA'] * 0.2
})
self.solver_qua = optim.Adam(Q_params)
self.solver_etp = optim.Adam(self.entropy.parameters(),
lr=cfg['TRAIN']['LR_ENP'])
self.solver_dec = optim.Adam(self.decoder.parameters(),
lr=cfg['TRAIN']['LR_DEC'])
self.scheduler_enc = LS.MultiStepLR(
self.solver_enc,
milestones=cfg['TRAIN']['MILESTONES'],
gamma=cfg['TRAIN']['GAMMA'])
self.scheduler_qua = LS.MultiStepLR(
self.solver_qua,
milestones=cfg['TRAIN']['MILESTONES'],
gamma=cfg['TRAIN']['GAMMA'])
self.scheduler_etp = LS.MultiStepLR(
self.solver_etp,
milestones=cfg['TRAIN']['MILESTONES'],
gamma=cfg['TRAIN']['GAMMA'])
self.scheduler_dec = LS.MultiStepLR(
self.solver_dec,
milestones=cfg['TRAIN']['MILESTONES'],
gamma=cfg['TRAIN']['GAMMA'])
train_transform = transforms.Compose([
transforms.RandomCrop(
(cfg['DATASET']['PATCH_SIZE'], cfg['DATASET']['PATCH_SIZE'])),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(),
])
train_set = dataset.ImageFolder(root=cfg['DATASET']['TRAIN_SET'],
transform=train_transform)
self.train_loader = data.DataLoader(
dataset=train_set,
batch_size=cfg['TRAIN']['BATCH_SIZE'],
shuffle=True,
num_workers=cfg['WORKERS'])
test_transform = transforms.Compose([
transforms.ToTensor(),
])
test_set = dataset.ImageFolder(root=cfg['DATASET']['TEST_SET'],
transform=test_transform)
self.test_loader = data.DataLoader(dataset=test_set,
batch_size=1,
shuffle=False,
num_workers=cfg['WORKERS'])
self.entropy_loss = nn.CrossEntropyLoss()
def main():
global args
args = parser.parse_args()
train_transform = transforms.Compose([
transforms.RandomCrop((32, 32)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor()
])
train_dataset = dataset.ImageFolder(roots=args.train_paths,
transform=train_transform)
train_loader = data.DataLoader(dataset = train_dataset,\
batch_size = args.batch_size,\
shuffle = True,\
num_workers = args.num_workers
)
encoder = compression_net.CompressionEncoder(rnn_type=args.rnn_type).cuda()
binarizer = compression_net.CompressionBinarizer(
code_size=args.code_size).cuda()
decoder = compression_net.CompressionDecoder(
rnn_type=args.rnn_type, code_size=args.code_size).cuda()
optimizer = optim.Adam([
{
'params': encoder.parameters()
},
{
'params': binarizer.parameters()
},
{
'params': decoder.parameters()
},
],
lr=args.lr)
start_epoch = args.start_epoch
if args.resume:
if os.path.isdir('checkpoint'):
print("=> loading checkpoint '{}'".format(args.resume))
encoder.load_state_dict(
torch.load('checkpoint/{}_{}_{}/encoder_{:08d}.pth'.format(
args.rnn_type, args.loss_type, args.code_size,
args.resume)))
binarizer.load_state_dict(
torch.load('checkpoint/{}_{}_{}/binarizer_{:08d}.pth'.format(
args.rnn_type, args.loss_type, args.code_size,
args.resume)))
decoder.load_state_dict(
torch.load('checkpoint/{}_{}_{}/decoder_{:08d}.pth'.format(
args.rnn_type, args.loss_type, args.code_size,
args.resume)))
#args.start_epoch = checkpoint['epoch']
#model.load_state_dict(checkpoint['state_dict'])
#optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}'".format(args.resume))
start_epoch = args.resume + 1
else:
print("=> no checkpoint found at '{}'".format(args.resume))
torch.manual_seed(23)
scheduler = LS.MultiStepLR(optimizer,
milestones=[3, 10, 20, 50, 100],
gamma=0.5)
for epoch in range(start_epoch, start_epoch + args.epochs):
scheduler.step()
train(train_loader, encoder, binarizer, decoder, epoch, optimizer)
if epoch % args.save_freq == 0 or epoch == args.epochs - 1:
if not os.path.exists('checkpoint'):
os.mkdir('checkpoint')
if not os.path.exists('checkpoint/{}_{}_{}'.format(
args.rnn_type, args.loss_type, args.code_size)):
os.mkdir('checkpoint/{}_{}_{}'.format(args.rnn_type,
args.loss_type,
args.code_size))
torch.save(
encoder.state_dict(),
'checkpoint/{}_{}_{}/encoder_{:08d}.pth'.format(
args.rnn_type, args.loss_type, args.code_size, epoch))
torch.save(
binarizer.state_dict(),
'checkpoint/{}_{}_{}/binarizer_{:08d}.pth'.format(
args.rnn_type, args.loss_type, args.code_size, epoch))
torch.save(
decoder.state_dict(),
'checkpoint/{}_{}_{}/decoder_{:08d}.pth'.format(
args.rnn_type, args.loss_type, args.code_size, epoch))
def load_from_image_folder():
train_set = dataset.ImageFolder(root=args.train, transform=train_transform)
train_loader = data.DataLoader(dataset=train_set, batch_size=args.batch_size, shuffle=True, num_workers=1)
return train_set, train_loader
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
#trainset = torchvision.datasets.CIFAR10(root='/home/rzding/DATA', train=True, download=True, transform=transform_train)
trainset = torchvision.datasets.CIFAR10(root='/home/rzding/DATA', train=True, download=True, transform=None)
trainset = dataset.ImageFolder(root=, transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=1, shuffle=True, num_workers=2)
#testset = torchvision.datasets.CIFAR10(root='/home/rzding/DATA', train=False, download=True, transform=transform_train)
testset = torchvision.datasets.CIFAR10(root='/home/rzding/DATA', train=False, download=True, transform=None)
testloader = torch.utils.data.DataLoader(testset, batch_size=1, shuffle=False, num_workers=2)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
#classes = ('airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
classes_f2c = {}
for idx,a_class in enumerate(classes):
if a_class in ['bird', 'plane', 'car', 'ship', 'truck']:
classes_f2c[idx] = 0
elif a_class in ['cat', 'deer', 'dog', 'frog', 'horse']:
classes_f2c[idx] = 1
'--iterations', type=int, default=16, help='unroll iterations')
parser.add_argument(
'--directory', '-d', required=True, type=str, help='directory of checkpoint')
parser.add_argument('--checkpoint', type=int, help='unroll iterations')
parser.add_argument('--evaluatePickle', '-p', required=True, type=str, help='folder of training images')
args = parser.parse_args()
from torchvision import transforms
eval_transform = transforms.Compose([
transforms.ToTensor()
])
# evaluatePickle = "0to20.p"
evaluatePickle = args.evaluatePickle
txtFile = "resultTxt/"+ evaluatePickle[:-1]+"txt" +"." +args.directory
eval_set = dataset.ImageFolder(root=args.eval,file_name = evaluatePickle ,train=False)
eval_loader = data.DataLoader(
dataset=eval_set, batch_size=args.batch_size, shuffle=True, num_workers=0)
print('total images: {}; total batches: {}'.format(
len(eval_set), len(eval_loader)))
hypernet = network.HyperNetwork(eval_set.vid_count).cuda()
encoder = network.EncoderCell().cuda()
binarizer = network.Binarizer().cuda()
decoder = network.DecoderCell().cuda()
hypernet.eval()
encoder.eval()