def __getitem__(self, index):
if self.is_train:
img, target = self.train_img[index], self.train_label[index]
if len(img.shape) == 2:
img = np.stack([img] * 3, 2)
img = Image.fromarray(img, mode="RGB")
img = transforms.Resize((256, 256), Image.BILINEAR)(img)
img = transforms.RandomCrop(INPUT_SIZE)(img)
img = transforms.RandomHorizontalFlip()(img)
img = transforms.ToTensor()(img)
img = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])(img)
else:
img, target = self.test_img[index], self.test_label[index]
if len(img.shape) == 2:
img = np.stack([img] * 3, 2)
img = Image.fromarray(img, mode="RGB")
img = transforms.Resize((256, 256), Image.BILINEAR)(img)
img = transforms.CenterCrop(INPUT_SIZE)(img)
img = transforms.ToTensor()(img)
img = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])(img)
return img, target
def load_data(args):
normalize = t.Normalize(mean=[0.445, 0.287, 0.190],
std=[0.31, 0.225, 0.168])
im_transform = t.Compose([t.ToTensor(), normalize])
# Use the following code fo co_transformations e.g. random rotation or random flip etc.
# co_transformer = cot.Compose([cot.RandomRotate(45)])
dsetTrain = GIANA(args.imgdir,
args.gtdir,
input_size=(args.input_width, args.input_height),
train=True,
transform=im_transform,
co_transform=None,
target_transform=t.ToLabel())
train_data_loader = data.DataLoader(dsetTrain,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers)
dsetVal = GIANA(args.imgdir,
args.gtdir,
train=False,
transform=im_transform,
co_transform=None,
target_transform=t.ToLabel())
val_data_loader = data.DataLoader(dsetVal,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers)
return train_data_loader, val_data_loader
def __init__(self, cms, margin_size, margin_time):
self.IMAGE_SCALE = 20
self.IMAGE_SIZE = self.IMAGE_SCALE * self.IMAGE_SCALE
self.train_transform = transforms.Compose([
transforms.Resize((self.IMAGE_SCALE, self.IMAGE_SCALE)),
transforms.ToTensor()
])
self.test_transform = transforms.Compose([
transforms.Resize((self.IMAGE_SCALE, self.IMAGE_SCALE)),
transforms.ToTensor()
])
self.CMS = cms
self.MARGIN_SIZE = margin_size
self.MARGIN_TIME = margin_time
def get_transform():
transforms = []
transforms.append(custom_T.ToTensor())
transforms.append(custom_T.FasterRCNNResizer())
return custom_T.Compose(transforms)
def get_transform(train):
transforms = []
# converts the image, a PIL image, into a PyTorch Tensor
transforms.append(T.ToTensor())
if train:
# during training, randomly flip the training images
# and ground-truth for data augmentation
transforms.append(T.RandomHorizontalFlip(0.5))
return T.Compose(transforms)
def get_transform(train=False):
transforms = []
if train:
transforms.append(custom_T.RandomHorizontalFlip())
transforms.append(custom_T.RandomCrop())
transforms.append(custom_T.ToTensor())
transforms.append(custom_T.FasterRCNNResizer())
return custom_T.Compose(transforms)
def __init__(self, root=None, dataloader=default_loader):
self.transform1 = transforms.Compose([
transforms.RandomRotation(30),
transforms.Resize([256, 256]),
transforms.RandomCrop(INPUT_SIZE),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=(0.9, 1.1),
contrast=(0.9, 1.1),
saturation=(0.9, 1.1)),
transforms.ToTensor(),
transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
transforms.RandomErasing(probability=0.5, sh=0.05)
])
# 增强方法2： 关注更小的区域
self.transform2 = transforms.Compose([
transforms.RandomRotation(30),
transforms.Resize([336, 336]),
transforms.RandomCrop(INPUT_SIZE),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=(0.9, 1.1),
contrast=(0.9, 1.1),
saturation=(0.9, 1.1)),
transforms.ToTensor(),
transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
transforms.RandomErasing(probability=0.5, sh=0.05)
])
self.dataloader = dataloader
self.root = root
with open(os.path.join(self.root, TRAIN_DATASET), 'r') as fid:
self.imglist = fid.readlines()
self.labels = []
for line in self.imglist:
image_path, label = line.strip().split()
self.labels.append(int(label))
self.labels = np.array(self.labels)
self.labels = torch.LongTensor(self.labels)
def visualization(x_train, y_train, x_val, y_val, x_test, y_test, nb_class,
ckpt, opt):
# no augmentations used for linear evaluation
transform_lineval = transforms.Compose([transforms.ToTensor()])
train_set_lineval = UCR2018(data=x_train,
targets=y_train,
transform=transform_lineval)
val_set_lineval = UCR2018(data=x_val,
targets=y_val,
transform=transform_lineval)
test_set_lineval = UCR2018(data=x_test,
targets=y_test,
transform=transform_lineval)
train_loader_lineval = torch.utils.data.DataLoader(train_set_lineval,
batch_size=128,
shuffle=True)
val_loader_lineval = torch.utils.data.DataLoader(val_set_lineval,
batch_size=128,
shuffle=False)
test_loader_lineval = torch.utils.data.DataLoader(test_set_lineval,
batch_size=128,
shuffle=False)
signal_length = x_train.shape[1]
# loading the saved backbone
backbone_lineval = SimConv4().cuda() # defining a raw backbone model
checkpoint = torch.load(ckpt, map_location='cpu')
backbone_lineval.load_state_dict(checkpoint)
print('Visualization')
for epoch in range(opt.epochs_test):
backbone_lineval.eval()
embeds = None
labels = None
for i, (data, target) in enumerate(test_loader_lineval):
data = data.cuda()
target = target.cuda().view(-1, 1)
output = backbone_lineval(data).detach()
if embeds is None:
embeds = output.cpu().numpy()
labels = target.cpu().numpy()
else:
embeds = np.vstack([embeds, output.cpu().numpy()])
labels = np.vstack([labels, target.cpu().numpy()])
return embeds, labels
def pretrain_InterSampleRel(x_train, y_train, opt):
K = opt.K
batch_size = opt.batch_size # 128 has been used in the paper
tot_epochs = opt.epochs # 400 has been used in the paper
feature_size = opt.feature_size
ckpt_dir = opt.ckpt_dir
prob = 0.2 # Transform Probability
raw = transforms.Raw()
cutout = transforms.Cutout(sigma=0.1, p=prob)
jitter = transforms.Jitter(sigma=0.2, p=prob)
scaling = transforms.Scaling(sigma=0.4, p=prob)
magnitude_warp = transforms.MagnitudeWrap(sigma=0.3, knot=4, p=prob)
time_warp = transforms.TimeWarp(sigma=0.2, knot=8, p=prob)
window_slice = transforms.WindowSlice(reduce_ratio=0.8, p=prob)
window_warp = transforms.WindowWarp(window_ratio=0.3, scales=(0.5, 2), p=prob)
transforms_list = {'jitter': [jitter],
'cutout': [cutout],
'scaling': [scaling],
'magnitude_warp': [magnitude_warp],
'time_warp': [time_warp],
'window_slice': [window_slice],
'window_warp': [window_warp],
'G0': [jitter, magnitude_warp, window_slice],
'G1': [jitter, time_warp, window_slice],
'G2': [jitter, time_warp, window_slice, window_warp, cutout],
'none': [raw]}
transforms_targets = list()
for name in opt.aug_type:
for item in transforms_list[name]:
transforms_targets.append(item)
train_transform = transforms.Compose(transforms_targets + [transforms.ToTensor()])
backbone = SimConv4().cuda()
model = RelationalReasoning(backbone, feature_size).cuda()
train_set = MultiUCR2018(data=x_train, targets=y_train, K=K, transform=train_transform)
train_loader = torch.utils.data.DataLoader(train_set,
batch_size=batch_size,
shuffle=True)
torch.save(model.backbone.state_dict(), '{}/backbone_init.tar'.format(ckpt_dir))
acc_max, epoch_max = model.train(tot_epochs=tot_epochs, train_loader=train_loader, opt=opt)
torch.save(model.backbone.state_dict(), '{}/backbone_last.tar'.format(ckpt_dir))
return acc_max, epoch_max
def eval(args, model):
if not os.path.exists(args.savedir):
os.makedirs(args.savedir)
listImgFiles = [
k.split('/')[-1].split('.')[0]
for k in glob.glob(os.path.join(args.imgdir, '*.bmp'))
]
for currFile in tqdm(listImgFiles):
img = Image.open(os.path.join(args.imgdir, currFile + '.bmp'))
img = t.ToTensor()(img)
if args.cuda:
img.cuda()
output = model(Variable(img, volatile=True).unsqueeze(0))
output = t.ToPILImage()(output[0].cpu().data)
newfilename = os.path.join(args.savedir,
currFile + '_' + args.model + '.bmp')
output.save(newfilename, 'BMP')
def get_dataloader():
# TODO(xwd): Adaptive normalization by some large image.
# E.g. In medical image processing, WSI image is very large and different to ordinary images.
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
train_transform = transform.Compose([
transform.RandomScale([cfg['scale_min'], cfg['scale_max']]),
transform.RandomRotate([cfg['rotate_min'], cfg['rotate_max']],
padding=mean,
ignore_label=cfg['ignore_label']),
transform.RandomGaussianBlur(),
transform.RandomHorizontallyFlip(),
transform.RandomCrop([cfg['train_h'], cfg['train_w']],
crop_type='rand',
padding=mean,
ignore_label=cfg['ignore_label']),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)
])
train_data = cityscapes.Cityscapes(cfg['data_path'],
split='train',
transform=train_transform)
# Use data sampler to make sure each GPU loads specific parts of dataset to avoid data reduntant.
train_sampler = DistributedSampler(train_data)
train_loader = DataLoader(train_data,
batch_size=cfg['batch_size'] //
cfg['world_size'],
shuffle=(train_sampler is None),
num_workers=4,
pin_memory=True,
sampler=train_sampler,
drop_last=True)
return train_loader, train_sampler
def get_transform(train=False, yolo=False, aug=None):
assert aug == 'dirty_camera_lens' or aug == 'gan' or aug is None, "Aug parameter not valid"
transforms = []
if yolo:
transforms.append(custom_T.PadToSquare())
transforms.append(custom_T.Resize(img_size=None))
if train:
transforms.append(custom_T.RandomHorizontalFlip())
transforms.append(custom_T.RandomCrop())
if aug == 'dirty_camera_lens':
print("Augmentation: Dirty Camera Lens")
transforms.append(custom_T.DirtyCameraLens())
transforms.append(custom_T.ToTensor())
# transforms.append(custom_T.FasterRCNNResizer())
return custom_T.Compose(transforms)
def __init__(self, root=None, dataloader=default_loader):
self.transform = transforms.Compose([
transforms.Resize([256, 256]),
transforms.RandomCrop(INPUT_SIZE),
transforms.ToTensor(),
transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
])
self.dataloader = dataloader
self.root = root
self.imgs = []
self.labels = []
with open(os.path.join(self.root, EVAL_DATASET), 'r') as fid:
for line in fid.readlines():
img_path, label = line.strip().split()
img = self.dataloader(img_path)
label = int(label)
self.imgs.append(img)
self.labels.append(label)
def deploy(path):
assert os.path.exists(path), f'{path} not found : ('
dataset = 'YOUR_DATASET_NAME'
img_size = 256
test_transform = transforms.Compose([
transforms.Resize((img_size, img_size)),
transforms.ToTensor(),
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])
testA = ImageFolder(os.path.join('dataset', dataset, 'testA'), test_transform)
with fluid.dygraph.guard():
testA_loader = DataLoader(testA, batch_size=1, shuffle=False)
real_A, _ = next(iter(testA_loader))
in_np = real_A.numpy()
# load model
place = fluid.CPUPlace()
exe = fluid.Executor(place)
program, feed_vars, fetch_vars = fluid.io.load_inference_model(path, exe)
# inference
fetch, = exe.run(program, feed={feed_vars[0]: in_np}, fetch_list=fetch_vars)
def img_postprocess(img):
assert isinstance(img, np.ndarray), type(img)
img = img * 0.5 + 0.5
img = img.squeeze(0).transpose((1, 2, 0))
# BGR to RGB
img = img[:, :, ::-1]
return img
in_img = img_postprocess(in_np)
out_img = img_postprocess(fetch)
plt.subplot(121)
plt.title('real A')
plt.imshow(in_img)
plt.subplot(122)
plt.title('A to B')
plt.imshow(out_img)
plt.show()
def main():
net = AFENet(classes=21, pretrained_model_path=None).cuda()
net.load_state_dict(torch.load(os.path.join(args['model_save_path'], args['snapshot'])))
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
test_transform = transform.Compose([transform.ToTensor()])
test_data = voc2012.VOC2012(split='test', data_root=args['dataset_root'], data_list=args['test_list'],
transform=test_transform)
test_loader = DataLoader(test_data, batch_size=1, shuffle=False, num_workers=1)
gray_folder = os.path.join(args['test_result_save_path'])
# gray_folder = os.path.join(args['test_result_save_path'], 'gray')
color_folder = os.path.join(args['test_result_save_path'], 'color')
colors = np.loadtxt(args['colors_path']).astype('uint8')
test(test_loader, test_data.data_list, net, 21, mean, std, 512, 480, 480, args['scales'],
gray_folder, color_folder, colors)
def main(args=None):
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
# torch.cuda.set_device(0)
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, args.log_train))
else:
sys.stdout = Logger(osp.join(args.save_dir, args.log_test))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_dataset(name=args.dataset)
transform_train = T.Compose([
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
T.RandomErasing(),
])
transform_train2 = T.Compose([
T.Resize((args.height, args.width)),
T.Random2DTranslation(args.height, args.width),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = False
trainloader = DataLoader(
VideoDataset(dataset.train,
data_name=args.dataset,
seq_len=args.seq_len,
sample='random',
transform=transform_train,
transform2=transform_train2,
type="train"),
sampler=RandomIdentitySampler(dataset.train,
num_instances=args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
VideoDataset(dataset.query,
data_name=args.dataset,
seq_len=args.seq_len,
sample='dense',
transform=transform_test,
type="test"),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
VideoDataset(dataset.gallery,
data_name=args.dataset,
seq_len=args.seq_len,
sample='dense',
transform=transform_test,
type="test"),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing models: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
final_dim=args.feat_dim)
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
crossEntropyLoss = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
tripletLoss = TripletLoss(margin=args.margin)
regularLoss = RegularLoss(use_gpu=use_gpu)
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
scheduler = WarmupMultiStepLR(optimizer, args.stepsize, args.gamma,
args.warmup_factor, args.warmup_items,
args.warmup_method)
start_epoch = args.start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
atest(model, queryloader, galleryloader, use_gpu)
return
start_time = time.time()
best_rank1 = -np.inf
for epoch in range(start_epoch, args.max_epoch):
print("==> Epoch {}/{}".format(epoch + 1, args.max_epoch))
train(model, crossEntropyLoss, tripletLoss, regularLoss, optimizer,
trainloader, use_gpu)
# if args.stepsize > 0:
scheduler.step()
if (epoch + 1) >= 200 and (epoch + 1) % args.eval_step == 0:
print("==> Test")
rank1 = atest(model, queryloader, galleryloader, use_gpu)
is_best = rank1 > best_rank1
if is_best: best_rank1 = rank1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
}, is_best,
osp.join(
args.save_dir,
args.model_name + str(epoch + 1) + '.pth.tar'))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
print("Finished. Total elapsed time (h:m:s): {}".format(elapsed))
def main(args=None):
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
# torch.cuda.set_device(0)
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, args.log_train))
else:
sys.stdout = Logger(osp.join(args.save_dir, args.log_test))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_dataset(name=args.dataset)
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = False
queryloader = DataLoader(
VideoDataset(dataset.query,
data_name=args.dataset,
seq_len=args.seq_len,
sample='dense',
transform=transform_test,
type="test"),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
VideoDataset(dataset.gallery,
data_name=args.dataset,
seq_len=args.seq_len,
sample='dense',
transform=transform_test,
type="test"),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing models: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
final_dim=args.feat_dim)
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
checkpoint = torch.load(args.model_path)
model.load_state_dict(checkpoint['state_dict'])
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
atest(model, queryloader, galleryloader, use_gpu)
return
def main():
net = AFENet(classes=21,
pretrained_model_path=args['pretrained_model_path']).cuda()
net_ori = [net.layer0, net.layer1, net.layer2, net.layer3, net.layer4]
net_new = [
net.ppm, net.cls, net.aux, net.ppm_reduce, net.aff1, net.aff2, net.aff3
]
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
train_transform = transform.Compose([
transform.RandScale([0.75, 2.0]),
transform.RandomHorizontalFlip(),
transform.Crop([480, 480],
crop_type='rand',
padding=mean,
ignore_label=255),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)
])
train_data = voc2012.VOC2012(split='train',
data_root=args['dataset_root'],
data_list=args['train_list'],
transform=train_transform)
train_loader = DataLoader(train_data,
batch_size=args['train_batch_size'],
shuffle=True,
num_workers=8,
drop_last=True)
val_transform = transform.Compose([
transform.Crop([480, 480],
crop_type='center',
padding=mean,
ignore_label=255),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)
])
val_data = voc2012.VOC2012(split='val',
data_root=args['dataset_root'],
data_list=args['val_list'],
transform=val_transform)
val_loader = DataLoader(val_data,
batch_size=args['val_batch_size'],
shuffle=False,
num_workers=8)
if len(args['snapshot']) == 0:
curr_epoch = 1
args['best_record'] = {
'epoch': 0,
'val_loss': 1e10,
'acc': 0,
'acc_cls': 0,
'mean_iu': 0,
'fwavacc': 0
}
else:
print('training resumes from ' + args['snapshot'])
net.load_state_dict(
torch.load(os.path.join(args['model_save_path'],
args['snapshot'])))
split_snapshot = args['snapshot'].split('_')
curr_epoch = int(split_snapshot[1]) + 1
args['best_record'] = {
'epoch': int(split_snapshot[1]),
'val_loss': float(split_snapshot[3]),
'acc': float(split_snapshot[5]),
'acc_cls': float(split_snapshot[7]),
'mean_iu': float(split_snapshot[9]),
'fwavacc': float(split_snapshot[11])
}
params_list = []
for module in net_ori:
params_list.append(dict(params=module.parameters(), lr=args['lr']))
for module in net_new:
params_list.append(dict(params=module.parameters(),
lr=args['lr'] * 10))
args['index_split'] = 5
criterion = torch.nn.CrossEntropyLoss(ignore_index=255)
optimizer = torch.optim.SGD(params_list,
lr=args['lr'],
momentum=args['momentum'],
weight_decay=args['weight_decay'])
if len(args['snapshot']) > 0:
optimizer.load_state_dict(
torch.load(
os.path.join(args['model_save_path'],
'opt_' + args['snapshot'])))
check_makedirs(args['model_save_path'])
all_iter = args['epoch_num'] * len(train_loader)
for epoch in range(curr_epoch, args['epoch_num'] + 1):
train(train_loader, net, optimizer, epoch, all_iter)
validate(val_loader, net, criterion, optimizer, epoch)
def get_transform(train=False):
transform = []
transform.append(transforms.ToTensor())
if train:
transform.append(transforms.RandomHorizontalFlip(0.5))
return transforms.Compose(transform)
def get_transform(train):
transforms = []
transforms.append(T.ToTensor())
if train:
transforms.append(T.RandomHorizontalFlip(0.5))
return T.Compose(transforms)
def build_model(self):
""" DataLoader """
pad = int(30 * self.img_size // 256)
train_transform = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.Resize((self.img_size + pad, self.img_size + pad)),
transforms.RandomCrop(self.img_size),
transforms.ToTensor(),
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])
test_transform = transforms.Compose([
transforms.Resize((self.img_size, self.img_size)),
transforms.ToTensor(),
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])
self.trainA = ImageFolder(
os.path.join('dataset', self.dataset, 'trainA'), train_transform)
self.trainB = ImageFolder(
os.path.join('dataset', self.dataset, 'trainB'), train_transform)
self.testA = ImageFolder(
os.path.join('dataset', self.dataset, 'testA'), test_transform)
self.testB = ImageFolder(
os.path.join('dataset', self.dataset, 'testB'), test_transform)
self.trainA_loader = DataLoader(self.trainA,
batch_size=self.batch_size,
shuffle=True)
self.trainB_loader = DataLoader(self.trainB,
batch_size=self.batch_size,
shuffle=True)
self.testA_loader = DataLoader(self.testA, batch_size=1, shuffle=False)
self.testB_loader = DataLoader(self.testB, batch_size=1, shuffle=False)
""" Define Generator, Discriminator """
self.genA2B = ResnetGenerator(input_nc=3,
output_nc=3,
ngf=self.ch,
n_blocks=self.n_res,
img_size=self.img_size,
light=self.light)
self.genB2A = ResnetGenerator(input_nc=3,
output_nc=3,
ngf=self.ch,
n_blocks=self.n_res,
img_size=self.img_size,
light=self.light)
self.disGA = Discriminator(input_nc=3, ndf=self.ch, n_layers=7)
self.disGB = Discriminator(input_nc=3, ndf=self.ch, n_layers=7)
self.disLA = Discriminator(input_nc=3, ndf=self.ch, n_layers=5)
self.disLB = Discriminator(input_nc=3, ndf=self.ch, n_layers=5)
""" Define Loss """
self.L1_loss = loss.L1Loss()
self.MSE_loss = loss.MSELoss()
self.BCE_loss = loss.BCEWithLogitsLoss()
""" Trainer """
def get_params(block):
out = []
for name, param in block.named_parameters():
if 'instancenorm' in name or 'weight_u' in name or 'weight_v' in name:
continue
out.append(param)
return out
genA2B_parameters = get_params(self.genA2B)
genB2A_parameters = get_params(self.genB2A)
disGA_parameters = get_params(self.disGA)
disGB_parameters = get_params(self.disGB)
disLA_parameters = get_params(self.disLA)
disLB_parameters = get_params(self.disLB)
G_parameters = genA2B_parameters + genB2A_parameters
D_parameters = disGA_parameters + disGB_parameters + disLA_parameters + disLB_parameters
self.G_optim = fluid.optimizer.Adam(
parameter_list=G_parameters,
learning_rate=self.lr,
beta1=0.5,
beta2=0.999,
regularization=fluid.regularizer.L2Decay(self.weight_decay))
self.D_optim = fluid.optimizer.Adam(
parameter_list=D_parameters,
learning_rate=self.lr,
beta1=0.5,
beta2=0.999,
regularization=fluid.regularizer.L2Decay(self.weight_decay))
""" Define Rho clipper to constraint the value of rho in AdaILN and ILN"""
def evaluation(x_train,
y_train,
x_val,
y_val,
x_test,
y_test,
nb_class,
ckpt,
opt,
ckpt_tosave=None):
# no augmentations used for linear evaluation
transform_lineval = transforms.Compose([transforms.ToTensor()])
train_set_lineval = UCR2018(data=x_train,
targets=y_train,
transform=transform_lineval)
val_set_lineval = UCR2018(data=x_val,
targets=y_val,
transform=transform_lineval)
test_set_lineval = UCR2018(data=x_test,
targets=y_test,
transform=transform_lineval)
train_loader_lineval = torch.utils.data.DataLoader(train_set_lineval,
batch_size=128,
shuffle=True)
val_loader_lineval = torch.utils.data.DataLoader(val_set_lineval,
batch_size=128,
shuffle=False)
test_loader_lineval = torch.utils.data.DataLoader(test_set_lineval,
batch_size=128,
shuffle=False)
signal_length = x_train.shape[1]
# loading the saved backbone
backbone_lineval = SimConv4().cuda() # defining a raw backbone model
# backbone_lineval = OS_CNN(signal_length).cuda() # defining a raw backbone model
# 64 are the number of output features in the backbone, and 10 the number of classes
linear_layer = torch.nn.Linear(opt.feature_size, nb_class).cuda()
# linear_layer = torch.nn.Linear(backbone_lineval.rep_dim, nb_class).cuda()
checkpoint = torch.load(ckpt, map_location='cpu')
backbone_lineval.load_state_dict(checkpoint)
if ckpt_tosave:
torch.save(backbone_lineval.state_dict(), ckpt_tosave)
optimizer = torch.optim.Adam(linear_layer.parameters(),
lr=opt.learning_rate_test)
CE = torch.nn.CrossEntropyLoss()
early_stopping = EarlyStopping(opt.patience_test, verbose=True)
best_acc = 0
best_epoch = 0
print('Linear evaluation')
for epoch in range(opt.epochs_test):
linear_layer.train()
backbone_lineval.eval()
acc_trains = list()
for i, (data, target) in enumerate(train_loader_lineval):
optimizer.zero_grad()
data = data.cuda()
target = target.cuda()
output = backbone_lineval(data).detach()
output = linear_layer(output)
loss = CE(output, target)
loss.backward()
optimizer.step()
# estimate the accuracy
prediction = output.argmax(-1)
correct = prediction.eq(target.view_as(prediction)).sum()
accuracy = (100.0 * correct / len(target))
acc_trains.append(accuracy.item())
print('[Train-{}][{}] loss: {:.5f}; \t Acc: {:.2f}%' \
.format(epoch + 1, opt.model_name, loss.item(), sum(acc_trains) / len(acc_trains)))
acc_vals = list()
acc_tests = list()
linear_layer.eval()
with torch.no_grad():
for i, (data, target) in enumerate(val_loader_lineval):
data = data.cuda()
target = target.cuda()
output = backbone_lineval(data).detach()
output = linear_layer(output)
# estimate the accuracy
prediction = output.argmax(-1)
correct = prediction.eq(target.view_as(prediction)).sum()
accuracy = (100.0 * correct / len(target))
acc_vals.append(accuracy.item())
val_acc = sum(acc_vals) / len(acc_vals)
if val_acc > best_acc:
best_acc = val_acc
best_epoch = epoch
for i, (data, target) in enumerate(test_loader_lineval):
data = data.cuda()
target = target.cuda()
output = backbone_lineval(data).detach()
output = linear_layer(output)
# estimate the accuracy
prediction = output.argmax(-1)
correct = prediction.eq(target.view_as(prediction)).sum()
accuracy = (100.0 * correct / len(target))
acc_tests.append(accuracy.item())
test_acc = sum(acc_tests) / len(acc_tests)
print('[Test-{}] Val ACC:{:.2f}%, Best Test ACC.: {:.2f}% in Epoch {}'.
format(epoch, val_acc, test_acc, best_epoch))
early_stopping(val_acc, None)
if early_stopping.early_stop:
print("Early stopping")
break
return test_acc, best_epoch
#dataset prepare
#---------------------------------
print('Loading dataset...')
cache_size = 256
if args.image_size == 448:
cache_size = 256 * 2
if args.image_size == 352:
cache_size = 402
transform_train = transforms.Compose([
transforms.Resize((cache_size,cache_size)),
#transforms.Resize((args.image_size,args.image_size)),
#transforms.RandomRotation(10),
transforms.RandomCrop(args.image_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485,0.456,0.406], [0.229,0.224,0.225]),
])
transform_test = transforms.Compose([
transforms.Resize((cache_size,cache_size)),
transforms.CenterCrop(args.image_size),
transforms.ToTensor(),
transforms.Normalize([0.485,0.456,0.406], [0.229,0.224,0.225]),
])
print("Dataset Initializing...")
trainset = SRDataset.SRDataset(max_person=args.max_person,image_dir=args.images_root, \
images_list=args.train_file_pre + '_images.txt',bboxes_list=args.train_file_pre + '_bbox.json', \
relations_list=args.train_file_pre + '_relation.json', image_size=args.image_size,input_transform=transform_train)
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_imgreid_dataset(
root=args.root, name=args.dataset,
)
transform_train = T.Compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = True if use_gpu else False
trainloader = DataLoader(
ImageDataset(dataset.train, transform=transform_train),
sampler=RandomIdentitySampler(dataset.train, num_instances=args.num_instances),
batch_size=args.train_batch, num_workers=args.workers,
pin_memory=pin_memory, drop_last=True,
)
queryloader = DataLoader(
ImageDataset(dataset.test_3000_query, transform=transform_test),
batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
pin_memory=pin_memory, drop_last=False,
)
galleryloader = DataLoader(
ImageDataset(dataset.test_3000_gallery, transform=transform_test),
batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
pin_memory=pin_memory, drop_last=False,
)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch, num_classes=dataset.train_vehicle_nums, loss={'xent', 'htri'})
print("Model size: {:.3f} M".format(sum(p.numel() for p in model.parameters()) / 1000000.0))
modelid = 0
if args.upload:
modelid = upload_data.updatemodel(args.arch, args.lr, args.stepsize, args.gamma, args.loss, args.dataset,
args.height, args.width, args.seq_len, args.train_batch, args.other)
# criterion_xent = CrossEntropyLabelSmooth(num_classes=dataset.train_vehicle_nums, use_gpu=use_gpu)
criterion_xent = nn.CrossEntropyLoss()
criterion_htri = TripletLoss(margin=args.margin)
optimizer = init_optim(args.optim, model.parameters(), args.lr, args.weight_decay)
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=args.stepsize, gamma=args.gamma)
start_epoch = args.start_epoch
if args.load_weights:
# load pretrained weights but ignore layers that don't match in size
print("Loading pretrained weights from '{}'".format(args.load_weights))
checkpoint = torch.load(args.load_weights)
pretrain_dict = checkpoint['state_dict']
model_dict = model.state_dict()
pretrain_dict = {k: v for k, v in pretrain_dict.items() if k in model_dict and model_dict[k].size() == v.size()}
model_dict.update(pretrain_dict)
model.load_state_dict(model_dict)
if args.resume:
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
start_epoch = checkpoint['epoch']
rank1 = checkpoint['rank1']
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(start_epoch, rank1))
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
test(model, queryloader, galleryloader, use_gpu, return_distmat=True)
return
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
for epoch in range(start_epoch, args.max_epoch):
start_train_time = time.time()
train(modelid, epoch, model, criterion_xent, criterion_htri, optimizer, trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
scheduler.step()
if (epoch + 1) > args.start_eval and args.eval_step > 0 and (epoch + 1) % args.eval_step == 0 or (
epoch + 1) == args.max_epoch:
print("==> Test")
cmc, mAP = test(model, queryloader, galleryloader, use_gpu)
rank1 = cmc[0]
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best, osp.join(args.save_dir, 'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
if args.upload:
upload_data.updatetest(modelid, epoch + 1, mAP.item(), cmc[0].item(), cmc[4].item(), cmc[9].item(),
cmc[19].item())
upload_data.updaterank(modelid, best_rank1.item())
print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(best_rank1, best_epoch))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
train_time = str(datetime.timedelta(seconds=train_time))
print("Finished. Total elapsed time (h:m:s): {}. Training time (h:m:s): {}.".format(elapsed, train_time))
def main():
# Initialize arguments
args = init_arguments()
# Initialize random
init_seed(args.seed)
# Initialize the model
model, parameters, losses, accuracy_dict, accuracies, start_epoch = init_model(
args)
IMAGE_SCALE = 20
transform = transforms.Compose(
[transforms.Resize((IMAGE_SCALE, IMAGE_SCALE)),
transforms.ToTensor()])
omniglot_loader = loader.OmniglotLoader('data/omniglot',
classify=False,
partition=0.8,
classes=True)
train_loader = torch.utils.data.DataLoader(
OMNIGLOT('data/omniglot',
train=True,
transform=transform,
download=True,
omniglot_loader=omniglot_loader,
batch_size=parameters.episode_length),
batch_size=parameters.batch_size,
shuffle=True)
test_loader = torch.utils.data.DataLoader(OMNIGLOT(
'data/omniglot',
train=False,
transform=transform,
download=False,
omniglot_loader=omniglot_loader,
batch_size=parameters.episode_length),
batch_size=parameters.batch_size,
shuffle=False)
criterion = nn.CrossEntropyLoss(reduce=False)
optimizer = optim.Adam(model.parameters())
init_logging()
accuracy_dict, losses, accuracies = train_model(model, parameters,
train_loader, criterion,
optimizer, args,
accuracy_dict, losses,
accuracies, start_epoch)
loss_plot.plot([accuracies], ["Training Accuracy Percentage"],
"training_stats", args.name + "/", "Percentage")
loss_plot.plot([losses], ["Training Loss"], "training_loss",
args.name + "/", "Average Loss")
predictions, labels = evaluate(model, test_loader, criterion, parameters,
args, 10)
matrix_plot.plot_matrix(predictions,
labels,
args.name + "/",
title="matrix_plot_test")
scatterplot.plot(accuracy_dict,
args.name + "/",
parameters.batch_size,
title="Prediction Accuracy after Testing")
import matplotlib.pyplot as plt
from tools import prediction
from utils.metrics import Evaluator
args = get_args()
rng = np.random.RandomState(seed=args.seed)
torch.manual_seed(seed=args.seed)
transform_train = trans.Compose([
trans.RandomHorizontalFlip(),
#trans.FixScale((args.crop_size,args.crop_size)),
trans.RandomScale((0.5, 2.0)),
#trans.FixScale(args.crop_size),
trans.RandomCrop(args.crop_size),
trans.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
trans.ToTensor(),
])
transform_val = trans.Compose([
#trans.FixScale((args.crop_size,args.crop_size)),
trans.FixScale(args.crop_size),
trans.CenterCrop(args.crop_size),
trans.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
trans.ToTensor(),
])
if (args.aug == True):
voc_train = VOCSegmentation(root='./data',
set_name='train',
transform=transform_train)
else:
voc_train = VOCSegmentation(root='./data',
def get_data(data_dir,
source,
target,
source_train_path,
target_train_path,
source_extension,
target_extension,
height,
width,
batch_size,
re=0,
workers=8):
dataset = DA(data_dir, source, target, source_train_path,
target_train_path, source_extension, target_extension)
normalizer = T.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
source_num_classes = dataset.num_source_train_ids
train_transformer = T.Compose([
T.RandomSizedRectCrop(height, width),
T.RandomHorizontalFlip(),
T.ToTensor(),
normalizer,
T.RandomErasing(EPSILON=re),
])
test_transformer = T.Compose([
T.Resize((height, width), interpolation=3),
T.ToTensor(),
normalizer,
])
source_train_loader = DataLoader(Preprocessor(
dataset.source_train,
root=osp.join(dataset.source_images_dir, dataset.source_train_path),
transform=train_transformer),
batch_size=batch_size,
num_workers=0,
shuffle=True,
pin_memory=False,
drop_last=True)
target_train_loader = DataLoader(Preprocessor(
dataset.target_train,
root=osp.join(dataset.target_images_dir, dataset.target_train_path),
transform=train_transformer),
batch_size=batch_size,
num_workers=0,
shuffle=True,
pin_memory=False,
drop_last=True)
# source_train_loader = DataLoader(
# UnsupervisedCamStylePreprocessor(dataset.source_train, root=osp.join(dataset.source_images_dir, dataset.source_train_path),
# camstyle_root=osp.join(dataset.source_images_dir, dataset.source_train_path),
# transform=train_transformer),
# batch_size=batch_size, num_workers=0,
# shuffle=True, pin_memory=False, drop_last=True)
# target_train_loader = DataLoader(
# UnsupervisedCamStylePreprocessor(dataset.target_train,
# root=osp.join(dataset.target_images_dir, dataset.target_train_path),
# camstyle_root=osp.join(dataset.target_images_dir,
# dataset.target_train_camstyle_path),
# num_cam=dataset.target_num_cam, transform=train_transformer),
# batch_size=batch_size, num_workers=workers,
# shuffle=True, pin_memory=True, drop_last=True)
query_loader = DataLoader(Preprocessor(dataset.query,
root=osp.join(
dataset.target_images_dir,
dataset.query_path),
transform=test_transformer),
batch_size=batch_size,
num_workers=workers,
shuffle=False,
pin_memory=True)
gallery_loader = DataLoader(Preprocessor(dataset.gallery,
root=osp.join(
dataset.target_images_dir,
dataset.gallery_path),
transform=test_transformer),
batch_size=batch_size,
num_workers=workers,
shuffle=False,
pin_memory=True)
return dataset, source_num_classes, source_train_loader, target_train_loader, query_loader, gallery_loader
def semisupervised_train(x_train, y_train, x_val, y_val, x_test, y_test, opt):
# construct data loader
# Those are the transformations used in the paper
prob = 0.2 # Transform Probability
cutout = transforms_ts.Cutout(sigma=0.1, p=prob)
jitter = transforms_ts.Jitter(sigma=0.2, p=prob) # CIFAR10
scaling = transforms_ts.Scaling(sigma=0.4, p=prob)
magnitude_warp = transforms_ts.MagnitudeWrap(sigma=0.3, knot=4, p=prob)
time_warp = transforms_ts.TimeWarp(sigma=0.2, knot=8, p=prob)
window_slice = transforms_ts.WindowSlice(reduce_ratio=0.8, p=prob)
window_warp = transforms_ts.WindowWarp(window_ratio=0.3,
scales=(0.5, 2),
p=prob)
transforms_list = {
'jitter': [jitter],
'cutout': [cutout],
'scaling': [scaling],
'magnitude_warp': [magnitude_warp],
'time_warp': [time_warp],
'window_slice': [window_slice],
'window_warp': [window_warp],
'G0': [jitter, magnitude_warp, window_slice],
'G1': [jitter, time_warp, window_slice],
'G2': [jitter, time_warp, window_slice, window_warp, cutout],
'none': []
}
transforms_targets = list()
for name in opt.aug_type:
for item in transforms_list[name]:
transforms_targets.append(item)
if '2C' in opt.class_type:
cut_piece = transforms.CutPiece2C(sigma=opt.piece_size)
temp_class = 2
elif '3C' in opt.class_type:
cut_piece = transforms.CutPiece3C(sigma=opt.piece_size)
temp_class = 3
elif '4C' in opt.class_type:
cut_piece = transforms.CutPiece4C(sigma=opt.piece_size)
temp_class = 4
elif '5C' in opt.class_type:
cut_piece = transforms.CutPiece5C(sigma=opt.piece_size)
temp_class = 5
elif '6C' in opt.class_type:
cut_piece = transforms.CutPiece6C(sigma=opt.piece_size)
temp_class = 6
elif '7C' in opt.class_type:
cut_piece = transforms.CutPiece7C(sigma=opt.piece_size)
temp_class = 7
elif '8C' in opt.class_type:
cut_piece = transforms.CutPiece8C(sigma=opt.piece_size)
temp_class = 8
tensor_transform = transforms.ToTensor()
train_transform_peice = transforms.Compose(transforms_targets)
train_transform = transforms_ts.Compose(transforms_targets +
[transforms_ts.ToTensor()])
transform_eval = transforms.Compose([transforms.ToTensor()])
train_set_labeled = UCR2018(data=x_train,
targets=y_train,
transform=train_transform)
train_set_unlabeled = MultiUCR2018_Intra(
data=x_train,
targets=y_train,
K=opt.K,
transform=train_transform_peice,
transform_cut=cut_piece,
totensor_transform=tensor_transform)
val_set = UCR2018(data=x_val, targets=y_val, transform=transform_eval)
test_set = UCR2018(data=x_test, targets=y_test, transform=transform_eval)
train_dataset_size = len(train_set_labeled)
partial_size = int(opt.label_ratio * train_dataset_size)
train_ids = list(range(train_dataset_size))
np.random.shuffle(train_ids)
train_sampler = SubsetRandomSampler(train_ids[:partial_size])
trainloader_label = torch.utils.data.DataLoader(train_set_labeled,
batch_size=128,
sampler=train_sampler)
trainloader_unlabel = torch.utils.data.DataLoader(train_set_unlabeled,
batch_size=128)
val_loader_lineval = torch.utils.data.DataLoader(val_set,
batch_size=128,
shuffle=False)
test_loader_lineval = torch.utils.data.DataLoader(test_set,
batch_size=128,
shuffle=False)
# loading the saved backbone
backbone = SimConv4().cuda() # defining a raw backbone model
# 64 are the number of output features in the backbone, and 10 the number of classes
linear_layer = torch.nn.Linear(opt.feature_size, opt.nb_class).cuda()
# linear_layer = torch.nn.Sequential(
# torch.nn.Linear(opt.feature_size, 256),
# torch.nn.BatchNorm1d(256),
# torch.nn.LeakyReLU(),
# torch.nn.Linear(256, nb_class),
# torch.nn.Softmax(),
# ).cuda()
cls_head = torch.nn.Sequential(
torch.nn.Linear(opt.feature_size * 2, 256),
torch.nn.BatchNorm1d(256),
torch.nn.LeakyReLU(),
torch.nn.Linear(256, temp_class),
torch.nn.Softmax(),
).cuda()
optimizer = torch.optim.Adam([{
'params': backbone.parameters()
}, {
'params': linear_layer.parameters()
}, {
'params': cls_head.parameters()
}],
lr=opt.learning_rate)
CE = torch.nn.CrossEntropyLoss()
early_stopping = EarlyStopping(
opt.patience,
verbose=True,
checkpoint_pth='{}/backbone_best.tar'.format(opt.ckpt_dir))
torch.save(backbone.state_dict(),
'{}/backbone_init.tar'.format(opt.ckpt_dir))
best_acc = 0
best_epoch = 0
acc_epoch_cls = 0
print('Semi-supervised Train')
for epoch in range(opt.epochs):
backbone.train()
linear_layer.train()
cls_head.train()
acc_epoch = 0
acc_epoch_cls = 0
loss_epoch_label = 0
loss_epoch_unlabel = 0
loss = 0
for i, data_labeled in enumerate(trainloader_label):
optimizer.zero_grad()
# labeled sample
(x, target) = data_labeled
x = x.cuda()
target = target.cuda()
output = backbone(x)
output = linear_layer(output)
loss_label = CE(output, target)
loss_epoch_label += loss_label.item()
loss_label.backward()
optimizer.step()
# estimate the accuracy
prediction = output.argmax(-1)
correct = prediction.eq(target.view_as(prediction)).sum()
accuracy = (100.0 * correct / len(target))
acc_epoch += accuracy.item()
for i, data_unlabeled in enumerate(trainloader_unlabel):
optimizer.zero_grad()
# unlabeled time piece
(x_piece1, x_piece2, target_temp, _) = data_unlabeled
x_piece1 = torch.cat(x_piece1, 0).cuda()
x_piece2 = torch.cat(x_piece2, 0).cuda()
target_temp = torch.cat(target_temp, 0).cuda()
features_cut0 = backbone(x_piece1)
features_cut1 = backbone(x_piece2)
features_cls = torch.cat([features_cut0, features_cut1], 1)
c_output = cls_head(features_cls)
correct_cls, length_cls = run_test(c_output, target_temp)
loss_unlabel = CE(c_output, target_temp)
loss_unlabel.backward()
optimizer.step()
loss_epoch_unlabel += loss_unlabel.item()
accuracy_cls = 100. * correct_cls / length_cls
acc_epoch_cls += accuracy_cls.item()
acc_epoch /= len(trainloader_label)
acc_epoch_cls /= len(trainloader_unlabel)
loss_epoch_label /= len(trainloader_label)
loss_epoch_unlabel /= len(trainloader_unlabel)
print('[Train-{}][{}] loss_label: {:.5f}; \tloss_unlabel: {:.5f}; \t Acc label: {:.2f}% \t Acc unlabel: {:.2f}%' \
.format(epoch + 1, opt.model_name, loss_epoch_label, loss_epoch_unlabel, acc_epoch, acc_epoch_cls))
acc_vals = list()
acc_tests = list()
backbone.eval()
linear_layer.eval()
with torch.no_grad():
for i, (x, target) in enumerate(val_loader_lineval):
x = x.cuda()
target = target.cuda()
output = backbone(x).detach()
output = linear_layer(output)
# estimate the accuracy
prediction = output.argmax(-1)
correct = prediction.eq(target.view_as(prediction)).sum()
accuracy = (100.0 * correct / len(target))
acc_vals.append(accuracy.item())
val_acc = sum(acc_vals) / len(acc_vals)
if val_acc >= best_acc:
best_acc = val_acc
best_epoch = epoch
for i, (x, target) in enumerate(test_loader_lineval):
x = x.cuda()
target = target.cuda()
output = backbone(x).detach()
output = linear_layer(output)
# estimate the accuracy
prediction = output.argmax(-1)
correct = prediction.eq(target.view_as(prediction)).sum()
accuracy = (100.0 * correct / len(target))
acc_tests.append(accuracy.item())
test_acc = sum(acc_tests) / len(acc_tests)
print('[Test-{}] Val ACC:{:.2f}%, Best Test ACC.: {:.2f}% in Epoch {}'.
format(epoch, val_acc, test_acc, best_epoch))
early_stopping(val_acc, backbone)
if early_stopping.early_stop:
print("Early stopping")
break
torch.save(backbone.state_dict(),
'{}/backbone_last.tar'.format(opt.ckpt_dir))
return test_acc, best_epoch
kwargs = {'num_workers': 1, 'pin_memory': True} if args.cuda else {}
### PARAMETERS ###
# LSTM & Q Learning
IMAGE_SCALE = 28
IMAGE_SIZE = IMAGE_SCALE*IMAGE_SCALE
HIDDEN_LAYERS = 1
HIDDEN_NODES = 200
OUTPUT_CLASSES = args.class_vector_size
##################
train_transform = transforms.Compose([
transforms.Resize((IMAGE_SCALE, IMAGE_SCALE)),
transforms.ToTensor()
])
test_transform = transforms.Compose([
transforms.Resize((IMAGE_SCALE, IMAGE_SCALE)),
transforms.ToTensor()
])
print("Loading trainingsets...")
omniglot_loader = loader.OmniglotLoader('data/omniglot', classify=False, partition=0.8, classes=True)
train_loader = torch.utils.data.DataLoader(
OMNIGLOT('data/omniglot', train=True, transform=train_transform, download=True, omniglot_loader=omniglot_loader, batch_size=args.episode_size),
batch_size=args.mini_batch_size, shuffle=True, **kwargs)
print("Loading testset...")
test_loader = torch.utils.data.DataLoader(
OMNIGLOT('data/omniglot', train=False, transform=test_transform, omniglot_loader=omniglot_loader, batch_size=args.episode_size),
batch_size=args.mini_batch_size, shuffle=True, **kwargs)
def supervised_train2(x_train, y_train, x_val, y_val, x_test, y_test, nb_class,
opt):
# construct data loader
# Those are the transformations used in the paper
prob = 0.2 # Transform Probability
cutout = transforms_ts.Cutout(sigma=0.1, p=prob)
jitter = transforms_ts.Jitter(sigma=0.2, p=prob) # CIFAR10
scaling = transforms_ts.Scaling(sigma=0.4, p=prob)
magnitude_warp = transforms_ts.MagnitudeWrap(sigma=0.3, knot=4, p=prob)
time_warp = transforms_ts.TimeWarp(sigma=0.2, knot=8, p=prob)
window_slice = transforms_ts.WindowSlice(reduce_ratio=0.8, p=prob)
window_warp = transforms_ts.WindowWarp(window_ratio=0.3,
scales=(0.5, 2),
p=prob)
transforms_list = {
'jitter': [jitter],
'cutout': [cutout],
'scaling': [scaling],
'magnitude_warp': [magnitude_warp],
'time_warp': [time_warp],
'window_slice': [window_slice],
'window_warp': [window_warp],
'G0': [jitter, magnitude_warp, window_slice],
'G1': [jitter, time_warp, window_slice],
'G2': [jitter, time_warp, window_slice, window_warp, cutout],
'none': []
}
transforms_targets = list()
for name in opt.aug_type:
for item in transforms_list[name]:
transforms_targets.append(item)
train_transform = transforms_ts.Compose(transforms_targets +
[transforms_ts.ToTensor()])
transform_lineval = transforms.Compose([transforms.ToTensor()])
train_set_lineval = UCR2018(data=x_train,
targets=y_train,
transform=train_transform)
val_set_lineval = UCR2018(data=x_val,
targets=y_val,
transform=transform_lineval)
test_set_lineval = UCR2018(data=x_test,
targets=y_test,
transform=transform_lineval)
train_loader_lineval = torch.utils.data.DataLoader(train_set_lineval,
batch_size=128,
shuffle=True)
val_loader_lineval = torch.utils.data.DataLoader(val_set_lineval,
batch_size=128,
shuffle=False)
test_loader_lineval = torch.utils.data.DataLoader(test_set_lineval,
batch_size=128,
shuffle=False)
# loading the saved backbone
backbone_lineval = SimConv4().cuda() # defining a raw backbone model
# 64 are the number of output features in the backbone, and 10 the number of classes
linear_layer = torch.nn.Linear(opt.feature_size, nb_class).cuda()
optimizer = torch.optim.Adam([{
'params': backbone_lineval.parameters()
}, {
'params': linear_layer.parameters()
}],
lr=opt.learning_rate)
CE = torch.nn.CrossEntropyLoss()
early_stopping = EarlyStopping(
opt.patience,
verbose=True,
checkpoint_pth='{}/backbone_best.tar'.format(opt.ckpt_dir))
torch.save(backbone_lineval.state_dict(),
'{}/backbone_init.tar'.format(opt.ckpt_dir))
best_acc = 0
best_epoch = 0
print('Supervised Train')
for epoch in range(opt.epochs):
backbone_lineval.train()
linear_layer.train()
acc_trains = list()
for i, (data, target) in enumerate(train_loader_lineval):
optimizer.zero_grad()
data = data.cuda()
target = target.cuda()
output = backbone_lineval(data)
output = linear_layer(output)
loss = CE(output, target)
loss.backward()
optimizer.step()
# estimate the accuracy
prediction = output.argmax(-1)
correct = prediction.eq(target.view_as(prediction)).sum()
accuracy = (100.0 * correct / len(target))
acc_trains.append(accuracy.item())
print('[Train-{}][{}] loss: {:.5f}; \t Acc: {:.2f}%' \
.format(epoch + 1, opt.model_name, loss.item(), sum(acc_trains) / len(acc_trains)))
acc_vals = list()
acc_tests = list()
backbone_lineval.eval()
linear_layer.eval()
with torch.no_grad():
for i, (data, target) in enumerate(val_loader_lineval):
data = data.cuda()
target = target.cuda()
output = backbone_lineval(data).detach()
output = linear_layer(output)
# estimate the accuracy
prediction = output.argmax(-1)
correct = prediction.eq(target.view_as(prediction)).sum()
accuracy = (100.0 * correct / len(target))
acc_vals.append(accuracy.item())
val_acc = sum(acc_vals) / len(acc_vals)
if val_acc >= best_acc:
best_acc = val_acc
best_epoch = epoch
for i, (data, target) in enumerate(test_loader_lineval):
data = data.cuda()
target = target.cuda()
output = backbone_lineval(data).detach()
output = linear_layer(output)
# estimate the accuracy
prediction = output.argmax(-1)
correct = prediction.eq(target.view_as(prediction)).sum()
accuracy = (100.0 * correct / len(target))
acc_tests.append(accuracy.item())
test_acc = sum(acc_tests) / len(acc_tests)
print('[Test-{}] Val ACC:{:.2f}%, Best Test ACC.: {:.2f}% in Epoch {}'.
format(epoch, val_acc, test_acc, best_epoch))
early_stopping(val_acc, backbone_lineval)
if early_stopping.early_stop:
print("Early stopping")
break
torch.save(backbone_lineval.state_dict(),
'{}/backbone_last.tar'.format(opt.ckpt_dir))
return test_acc, best_epoch