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 __init__(self, root: str, annotation: str, targetHeight: int,
targetWidth: int, numClass: int):
self.root = root
self.coco = COCO(annotation)
self.ids = list(self.coco.imgs.keys())
self.targetHeight = targetHeight
self.targetWidth = targetWidth
self.numClass = numClass
self.transforms = T.Compose([
T.RandomOrder(
[T.RandomHorizontalFlip(),
T.RandomSizeCrop(numClass)]),
T.Resize((targetHeight, targetWidth)),
T.ColorJitter(brightness=.2, contrast=.1, saturation=.1, hue=0),
T.Normalize()
])
self.newIndex = {}
classes = []
for i, (k, v) in enumerate(self.coco.cats.items()):
self.newIndex[k] = i
classes.append(v['name'])
with open('classes.txt', 'w') as f:
f.write(str(classes))
def __init__(self,
annFile: str,
imageDir: str,
targetHeight: int,
targetWidth: int,
numClass: int,
train: bool = True):
self.annotations = {}
self.table = {
'不良-機械傷害': 0,
'不良-著色不佳': 1,
'不良-炭疽病': 2,
'不良-乳汁吸附': 3,
'不良-黑斑病': 4
}
self.imageDir = imageDir
self.numClass = numClass
with open(annFile, 'r', encoding='utf-8-sig') as f:
for line in f.readlines():
arr = line.rstrip().split(',')
ans = []
for idx in range(1, len(arr), 5):
tlx, tly, w, h, c = arr[idx:idx + 5]
if tlx:
tlx, tly, w, h = list(map(float, (tlx, tly, w, h)))
if c not in self.table:
self.table[c] = len(self.table)
cx = tlx + w / 2
cy = tly + h / 2
c = self.table[c]
ans.append(list(map(int, (cx, cy, w, h, c))))
self.annotations[arr[0]] = ans
self.names = list(self.annotations)
with open('table.txt', 'w') as f:
f.write(str(self.table))
print(self.table)
if train:
self.transforms = T.Compose([
T.RandomOrder([
T.RandomHorizontalFlip(),
T.RandomVerticalFlip(),
T.RandomSizeCrop(numClass)
]),
T.Resize((targetHeight, targetWidth)),
T.ColorJitter(brightness=.2, contrast=0, saturation=0, hue=0),
T.Normalize()
])
else:
self.transforms = T.Compose(
[T.Resize((targetHeight, targetWidth)),
T.Normalize()])
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 get_transforms(args):
transforms_list = [
custom_transforms.RandomSizedCrop(args.image_size, args.resize_min,
args.resize_max),
custom_transforms.RandomHorizontalFlip(),
custom_transforms.toTensor()
]
if args.color_space == 'lab':
transforms_list.insert(2, custom_transforms.toLAB())
elif args.color_space == 'rgb':
transforms_list.insert(2, custom_transforms.toRGB('RGB'))
transforms = custom_transforms.Compose(transforms_list)
return transforms
def __init__(self,
root: str,
targetHeight: int,
targetWidth: int,
numClass: int,
train: bool = True):
"""
:param root: should contain .jpg files and corresponding .txt files
:param targetHeight: desired height for model input
:param targetWidth: desired width for model input
:param numClass: number of classes in the given dataset
"""
self.cache = {}
imagePaths = glob(os.path.join(root, '*.jpg'))
for path in imagePaths:
name = path.split('/')[-1].split('.jpg')[0]
self.cache[path] = os.path.join(root, f'{name}.txt')
self.paths = list(self.cache.keys())
self.targetHeight = targetHeight
self.targetWidth = targetWidth
self.numClass = numClass
if train:
self.transforms = T.Compose([
T.RandomOrder([
T.RandomHorizontalFlip(),
T.RandomVerticalFlip(),
T.RandomSizeCrop(numClass)
]),
T.Resize((targetHeight, targetWidth)),
T.ColorJitter(brightness=.2, contrast=.1, saturation=.1,
hue=0),
T.Normalize()
])
else:
self.transforms = T.Compose(
[T.Resize((targetHeight, targetWidth)),
T.Normalize()])
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.RandomHorizontalFlip(),
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, TRAIN_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 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 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"""
#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', \
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)
if __name__ == "__main__":
dataset_root_path_1 = ""
dataset_root_path_2 = ""
split = "train"
percentage = 15
NUM_WORKERS = 0
BATCH_SIZE = 4
TGT_IMAGES_IN_BATCH = 1
DEVICE = "cuda"
DATASET_NAME_1 = ""
DATASET_NAME_2 = ""
transforms = None
if split == "train":
transforms = custom_T.Compose([
custom_T.RandomHorizontalFlip(),
custom_T.RandomCrop(),
custom_T.ToTensor(),
custom_T.FasterRCNNResizer()
])
elif split == "val" or split == "test":
transforms = custom_T.Compose(
[custom_T.ToTensor(),
custom_T.FasterRCNNResizer()])
dataset_1 = CustomYoloAnnotatedDataset(dataset_root_path_1,
transforms=transforms,
dataset_name=DATASET_NAME_1,
percentage=percentage,
split=split)
dataset_2 = CustomYoloAnnotatedDataset(dataset_root_path_2,
def main():
"""Create the model and start the training."""
or_nyu_dict = {
0: 255,
1: 16,
2: 40,
3: 39,
4: 7,
5: 14,
6: 39,
7: 12,
8: 38,
9: 40,
10: 10,
11: 6,
12: 40,
13: 39,
14: 39,
15: 40,
16: 18,
17: 40,
18: 4,
19: 40,
20: 40,
21: 5,
22: 40,
23: 40,
24: 30,
25: 36,
26: 38,
27: 40,
28: 3,
29: 40,
30: 40,
31: 9,
32: 38,
33: 40,
34: 40,
35: 40,
36: 34,
37: 37,
38: 40,
39: 40,
40: 39,
41: 8,
42: 3,
43: 1,
44: 2,
45: 22
}
or_nyu_map = lambda x: or_nyu_dict.get(x, x) - 1
or_nyu_map = np.vectorize(or_nyu_map)
device = torch.device("cuda" if not args.cpu else "cpu")
w, h = map(int, args.input_size.split(','))
input_size = (w, h)
w, h = map(int, args.input_size_target.split(','))
input_size_target = (w, h)
cudnn.enabled = True
args.or_nyu_map = or_nyu_map
# Create network
if args.model == 'DeepLab':
model = DeeplabMulti(num_classes=args.num_classes)
if args.restore_from[:4] == 'http':
saved_state_dict = model_zoo.load_url(args.restore_from)
elif args.restore_from == "":
saved_state_dict = None
else:
saved_state_dict = torch.load(args.restore_from)
new_params = model.state_dict().copy()
for i in saved_state_dict:
# Scale.layer5.conv2d_list.3.weight
i_parts = i.split('.')
# print i_parts
if not args.num_classes == 40 or not i_parts[1] == 'layer5':
new_params['.'.join(i_parts[1:])] = saved_state_dict[i]
# print i_parts
model.load_state_dict(new_params)
model.train()
model.to(device)
cudnn.benchmark = True
if args.mode != "baseline" and args.mode != "baseline_tar":
# init D
model_D1 = FCDiscriminator(num_classes=args.num_classes).to(device)
model_D2 = FCDiscriminator(num_classes=args.num_classes).to(device)
model_D1.train()
model_D1.to(device)
model_D2.train()
model_D2.to(device)
if not os.path.exists(args.snapshot_dir):
os.makedirs(args.snapshot_dir)
scale_min = 0.5
scale_max = 2.0
rotate_min = -10
rotate_max = 10
ignore_label = 255
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]
args.width = w
args.height = h
train_transform = transforms.Compose([
# et.ExtResize( 512 ),
transforms.RandScale([scale_min, scale_max]),
transforms.RandRotate([rotate_min, rotate_max],
padding=IMG_MEAN_RGB,
ignore_label=ignore_label),
transforms.RandomGaussianBlur(),
transforms.RandomHorizontalFlip(),
transforms.Crop([args.height + 1, args.width + 1],
crop_type='rand',
padding=IMG_MEAN_RGB,
ignore_label=ignore_label),
#et.ExtRandomCrop(size=(opts.crop_size, opts.crop_size)),
#et.ExtColorJitter(brightness=0.5, contrast=0.5, saturation=0.5),
#et.ExtRandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=IMG_MEAN, std=[1, 1, 1]),
])
val_transform = transforms.Compose([
# et.ExtResize( 512 ),
transforms.Crop([args.height + 1, args.width + 1],
crop_type='center',
padding=IMG_MEAN_RGB,
ignore_label=ignore_label),
transforms.ToTensor(),
transforms.Normalize(mean=IMG_MEAN, std=[1, 1, 1]),
])
if args.mode != "baseline_tar":
src_train_dst = OpenRoomsSegmentation(root=args.data_dir,
opt=args,
split='train',
transform=train_transform,
imWidth=args.width,
imHeight=args.height,
remap_labels=args.or_nyu_map)
else:
src_train_dst = NYU_Labelled(root=args.data_dir_target,
opt=args,
split='train',
transform=train_transform,
imWidth=args.width,
imHeight=args.height,
phase="TRAIN",
randomize=True)
tar_train_dst = NYU(root=args.data_dir_target,
opt=args,
split='train',
transform=train_transform,
imWidth=args.width,
imHeight=args.height,
phase="TRAIN",
randomize=True,
mode=args.mode)
tar_val_dst = NYU(root=args.data_dir,
opt=args,
split='val',
transform=val_transform,
imWidth=args.width,
imHeight=args.height,
phase="TRAIN",
randomize=False)
trainloader = data.DataLoader(src_train_dst,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True)
trainloader_iter = enumerate(trainloader)
targetloader = data.DataLoader(tar_train_dst,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True)
targetloader_iter = enumerate(targetloader)
# implement model.optim_parameters(args) to handle different models' lr setting
optimizer = optim.SGD(model.optim_parameters(args),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
optimizer.zero_grad()
if args.mode != "baseline" and args.mode != "baseline_tar":
optimizer_D1 = optim.Adam(model_D1.parameters(),
lr=args.learning_rate_D,
betas=(0.9, 0.99))
optimizer_D1.zero_grad()
optimizer_D2 = optim.Adam(model_D2.parameters(),
lr=args.learning_rate_D,
betas=(0.9, 0.99))
optimizer_D2.zero_grad()
if args.gan == 'Vanilla':
bce_loss = torch.nn.BCEWithLogitsLoss()
elif args.gan == 'LS':
bce_loss = torch.nn.MSELoss()
seg_loss = torch.nn.CrossEntropyLoss(ignore_index=255)
interp = nn.Upsample(size=(input_size[1] + 1, input_size[0] + 1),
mode='bilinear',
align_corners=True)
interp_target = nn.Upsample(size=(input_size_target[1] + 1,
input_size_target[0] + 1),
mode='bilinear',
align_corners=True)
# labels for adversarial training
source_label = 0
target_label = 1
# set up tensor board
if args.tensorboard:
if not os.path.exists(args.log_dir):
os.makedirs(args.log_dir)
writer = SummaryWriter(args.log_dir)
for i_iter in range(args.num_steps):
loss_seg_value1 = 0
loss_seg_value1_tar = 0
loss_adv_target_value1 = 0
loss_D_value1 = 0
loss_seg_value2 = 0
loss_seg_value2_tar = 0
loss_adv_target_value2 = 0
loss_D_value2 = 0
optimizer.zero_grad()
adjust_learning_rate(optimizer, i_iter)
if args.mode != "baseline" and args.mode != "baseline_tar":
optimizer_D1.zero_grad()
optimizer_D2.zero_grad()
adjust_learning_rate_D(optimizer_D1, i_iter)
adjust_learning_rate_D(optimizer_D2, i_iter)
sample_src = None
sample_tar = None
sample_res_src = None
sample_res_tar = None
sample_gt_src = None
sample_gt_tar = None
for sub_i in range(args.iter_size):
# train G
if args.mode != "baseline" and args.mode != "baseline_tar":
# don't accumulate grads in D
for param in model_D1.parameters():
param.requires_grad = False
for param in model_D2.parameters():
param.requires_grad = False
# train with source
try:
_, batch = trainloader_iter.__next__()
except:
trainloader_iter = enumerate(trainloader)
_, batch = trainloader_iter.__next__()
images, labels, _ = batch
sample_src = images.clone()
sample_gt_src = labels.clone()
images = images.to(device)
labels = labels.long().to(device)
pred1, pred2 = model(images)
pred1 = interp(pred1)
pred2 = interp(pred2)
sample_pred_src = pred2.detach().cpu()
loss_seg1 = seg_loss(pred1, labels)
loss_seg2 = seg_loss(pred2, labels)
loss = loss_seg2 + args.lambda_seg * loss_seg1
# proper normalization
loss = loss / args.iter_size
loss.backward()
loss_seg_value1 += loss_seg1.item() / args.iter_size
loss_seg_value2 += loss_seg2.item() / args.iter_size
# train with target
try:
_, batch = targetloader_iter.__next__()
except:
targetloader_iter = enumerate(targetloader)
_, batch = targetloader_iter.__next__()
images, tar_labels, _, labelled = batch
n_labelled = labelled.sum().detach().item()
batch_size = images.shape[0]
sample_tar = images.clone()
sample_gt_tar = tar_labels.clone()
images = images.to(device)
pred_target1, pred_target2 = model(images)
pred_target1 = interp_target(pred_target1)
pred_target2 = interp_target(pred_target2)
#print("N_labelled {}".format(n_labelled))
if args.mode == "sda" and n_labelled != 0:
labelled = labelled.to(device) == 1
tar_labels = tar_labels.to(device)
loss_seg1_tar = seg_loss(pred_target1[labelled],
tar_labels[labelled])
loss_seg2_tar = seg_loss(pred_target2[labelled],
tar_labels[labelled])
loss_tar_labelled = loss_seg2_tar + args.lambda_seg * loss_seg1_tar
loss_tar_labelled = loss_tar_labelled / args.iter_size
loss_seg_value1_tar += loss_seg1_tar.item() / args.iter_size
loss_seg_value2_tar += loss_seg2_tar.item() / args.iter_size
else:
loss_tar_labelled = torch.zeros(
1, requires_grad=True).float().to(device)
# proper normalization
sample_pred_tar = pred_target2.detach().cpu()
if args.mode != "baseline" and args.mode != "baseline_tar":
D_out1 = model_D1(F.softmax(pred_target1))
D_out2 = model_D2(F.softmax(pred_target2))
loss_adv_target1 = bce_loss(
D_out1,
torch.FloatTensor(
D_out1.data.size()).fill_(source_label).to(device))
loss_adv_target2 = bce_loss(
D_out2,
torch.FloatTensor(
D_out2.data.size()).fill_(source_label).to(device))
loss = args.lambda_adv_target1 * loss_adv_target1 + args.lambda_adv_target2 * loss_adv_target2
loss = loss / args.iter_size + loss_tar_labelled
#loss = loss_tar_labelled
loss.backward()
loss_adv_target_value1 += loss_adv_target1.item(
) / args.iter_size
loss_adv_target_value2 += loss_adv_target2.item(
) / args.iter_size
# train D
# bring back requires_grad
for param in model_D1.parameters():
param.requires_grad = True
for param in model_D2.parameters():
param.requires_grad = True
# train with source
pred1 = pred1.detach()
pred2 = pred2.detach()
D_out1 = model_D1(F.softmax(pred1))
D_out2 = model_D2(F.softmax(pred2))
loss_D1 = bce_loss(
D_out1,
torch.FloatTensor(
D_out1.data.size()).fill_(source_label).to(device))
loss_D2 = bce_loss(
D_out2,
torch.FloatTensor(
D_out2.data.size()).fill_(source_label).to(device))
loss_D1 = loss_D1 / args.iter_size / 2
loss_D2 = loss_D2 / args.iter_size / 2
loss_D1.backward()
loss_D2.backward()
loss_D_value1 += loss_D1.item()
loss_D_value2 += loss_D2.item()
# train with target
pred_target1 = pred_target1.detach()
pred_target2 = pred_target2.detach()
D_out1 = model_D1(F.softmax(pred_target1))
D_out2 = model_D2(F.softmax(pred_target2))
loss_D1 = bce_loss(
D_out1,
torch.FloatTensor(
D_out1.data.size()).fill_(target_label).to(device))
loss_D2 = bce_loss(
D_out2,
torch.FloatTensor(
D_out2.data.size()).fill_(target_label).to(device))
loss_D1 = loss_D1 / args.iter_size / 2
loss_D2 = loss_D2 / args.iter_size / 2
loss_D1.backward()
loss_D2.backward()
loss_D_value1 += loss_D1.item()
loss_D_value2 += loss_D2.item()
optimizer.step()
if args.mode != "baseline" and args.mode != "baseline_tar":
optimizer_D1.step()
optimizer_D2.step()
if args.tensorboard:
scalar_info = {
'loss_seg1': loss_seg_value1,
'loss_seg2': loss_seg_value2,
'loss_adv_target1': loss_adv_target_value1,
'loss_adv_target2': loss_adv_target_value2,
'loss_D1': loss_D_value1,
'loss_D2': loss_D_value2,
'loss_seg1_tar': loss_seg_value1_tar,
'loss_seg2_tar': loss_seg_value2_tar,
}
if i_iter % 10 == 0:
for key, val in scalar_info.items():
writer.add_scalar(key, val, i_iter)
if i_iter % 1000 == 0:
img = sample_src.cpu()[:, [2, 1, 0], :, :] + torch.from_numpy(
np.array(IMG_MEAN_RGB).reshape(1, 3, 1, 1)).float()
img = img.type(torch.uint8)
writer.add_images("Src/Images", img, i_iter)
label = tar_train_dst.decode_target(sample_gt_src).transpose(
0, 3, 1, 2)
writer.add_images("Src/Labels", label, i_iter)
preds = sample_pred_src.permute(0, 2, 3, 1).cpu().numpy()
preds = np.asarray(np.argmax(preds, axis=3), dtype=np.uint8)
preds = tar_train_dst.decode_target(preds).transpose(
0, 3, 1, 2)
writer.add_images("Src/Preds", preds, i_iter)
tar_img = sample_tar.cpu()[:,
[2, 1, 0], :, :] + torch.from_numpy(
np.array(IMG_MEAN_RGB).reshape(
1, 3, 1, 1)).float()
tar_img = tar_img.type(torch.uint8)
writer.add_images("Tar/Images", tar_img, i_iter)
tar_label = tar_train_dst.decode_target(
sample_gt_tar).transpose(0, 3, 1, 2)
writer.add_images("Tar/Labels", tar_label, i_iter)
tar_preds = sample_pred_tar.permute(0, 2, 3, 1).cpu().numpy()
tar_preds = np.asarray(np.argmax(tar_preds, axis=3),
dtype=np.uint8)
tar_preds = tar_train_dst.decode_target(tar_preds).transpose(
0, 3, 1, 2)
writer.add_images("Tar/Preds", tar_preds, i_iter)
print('exp = {}'.format(args.snapshot_dir))
print(
'iter = {0:8d}/{1:8d}, loss_seg1 = {2:.3f} loss_seg2 = {3:.3f} loss_adv1 = {4:.3f}, loss_adv2 = {5:.3f} loss_D1 = {6:.3f} loss_D2 = {7:.3f} loss_seg1_tar={8:.3f} loss_seg2_tar={9:.3f}'
.format(i_iter, args.num_steps, loss_seg_value1, loss_seg_value2,
loss_adv_target_value1, loss_adv_target_value2,
loss_D_value1, loss_D_value2, loss_seg_value1_tar,
loss_seg_value2_tar))
if i_iter >= args.num_steps_stop - 1:
print('save model ...')
torch.save(
model.state_dict(),
osp.join(args.snapshot_dir,
'OR_' + str(args.num_steps_stop) + '.pth'))
if args.mode != "baseline" and args.mode != "baseline_tar":
torch.save(
model_D1.state_dict(),
osp.join(args.snapshot_dir,
'OR_' + str(args.num_steps_stop) + '_D1.pth'))
torch.save(
model_D2.state_dict(),
osp.join(args.snapshot_dir,
'OR_' + str(args.num_steps_stop) + '_D2.pth'))
break
if i_iter % args.save_pred_every == 0 and i_iter != 0:
print('taking snapshot ...')
torch.save(
model.state_dict(),
osp.join(args.snapshot_dir, 'OR_' + str(i_iter) + '.pth'))
if args.mode != "baseline" and args.mode != "baseline_tar":
torch.save(
model_D1.state_dict(),
osp.join(args.snapshot_dir,
'OR_' + str(i_iter) + '_D1.pth'))
torch.save(
model_D2.state_dict(),
osp.join(args.snapshot_dir,
'OR_' + str(i_iter) + '_D2.pth'))
if args.tensorboard:
writer.close()
def build_transform(deterministic, color=True, daug=0):
transforms = []
if not deterministic:
transforms = [csl_tf.RandomHorizontalFlip(0.5)]
if daug == 1:
transforms += [
csl_tf.RandomAffine(
3,
translate=[0.025, 0.025],
scale=[0.975, 1.025],
shear=0,
keep_aspect=False,
)
]
elif daug == 2:
transforms += [
csl_tf.RandomAffine(
3,
translate=[0.035, 0.035],
scale=[0.970, 1.030],
shear=2,
keep_aspect=False,
)
]
elif daug == 3:
transforms += [
csl_tf.RandomAffine(
20,
translate=[0.035, 0.035],
scale=[0.970, 1.030],
shear=0,
keep_aspect=False,
)
]
elif daug == 4:
transforms += [
csl_tf.RandomAffine(
45,
translate=[0.035, 0.035],
scale=[0.940, 1.030],
shear=5,
keep_aspect=False,
)
]
elif daug == 5:
transforms += [
csl_tf.RandomAffine(
60,
translate=[0.035, 0.035],
scale=[0.940, 1.030],
shear=5,
keep_aspect=False,
)
]
elif daug == 6: # CVPR landmark training
transforms += [
csl_tf.RandomAffine(
30,
translate=[0.04, 0.04],
scale=[0.940, 1.050],
shear=5,
keep_aspect=False,
)
]
elif daug == 7:
transforms += [
csl_tf.RandomAffine(
0,
translate=[0.04, 0.04],
scale=[0.940, 1.050],
shear=5,
keep_aspect=False,
)
]
return tf.Compose(transforms)
import os
from data_provider import VOCSegmentation
from experiment_builder import ExperimentBuilder
from utils.arg_extractor import get_args
import utils.transforms as trans
from model.deeplab import DeepLab
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(),
])
def generate_dataloader(model, eval_type, args):
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.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]),
])
if eval_type == 'train':
data = ImageDataset(max_person=args.max_person + 1,
image_dir=args.images_root,
images_list=args.train_file_pre + '_images.txt',
bboxes_list=args.train_file_pre + '_bbox.json',
image_size=args.image_size,
input_transform=transform_train)
elif eval_type == 'valid':
data = ImageDataset(max_person=args.max_person + 1, image_dir=args.images_root, \
images_list=args.valid_file_pre + '_images.txt', \
bboxes_list=args.valid_file_pre + '_bbox.json', \
image_size=args.image_size, \
input_transform=transform_test)
else:
data = ImageDataset(max_person=args.max_person + 1, image_dir=args.images_root, \
images_list=args.test_file_pre + '_images.txt', \
bboxes_list=args.test_file_pre + '_bbox.json', \
image_size=args.image_size, \
input_transform=transform_test)
loader = torch.utils.data.DataLoader(data,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=0,
worker_init_fn=np.random.seed(
args.manualSeed))
model.eval()
batch_time = AverageMeter.AverageMeter()
end_time = time.time()
union_filename = './para_data/' + args.dataset + '_' + eval_type + '_union' + '.npy'
feat_filename = './para_data/' + args.dataset + '_' + eval_type + '_ssl.npy'
ssl_model = load_ssl_model()
ssl_model.eval()
if args.cuda:
ssl_model.cuda()
ssl_model = torch.nn.DataParallel(ssl_model)
if os.path.exists(feat_filename) and os.path.exists(
union_filename) and not args.regenerate_roifeat:
all_feat = np.load(feat_filename)
logger.info('loadding RoI feature npy from {} successfully'.format(
feat_filename))
all_union_feat = np.load(union_filename,
mmap_mode='r') # [B, N, N, 2048]
logger.info(
'loading union npy from {} successfully'.format(union_filename))
else:
all_feat, all_union_feat = [], []
for batch_idx, (img, image_bboxes) in enumerate(loader):
if args.cuda:
img, image_bboxes = img.cuda(), image_bboxes.cuda()
img, image_bboxes = Variable(img), Variable(image_bboxes)
node_num = image_bboxes.shape[1]
union_boxes = get_union_box(image_bboxes)
if args.cuda:
union_boxes = union_boxes.cuda()
image_bboxes = torch.cat((image_bboxes, union_boxes), axis=1)
del union_boxes
image_bboxes = Variable(image_bboxes)
# [batcn, node_num, 2048]
rois_feature_all = model(img, image_bboxes)
feature_num = rois_feature_all.shape[2]
rois_feature = rois_feature_all[:, :node_num, :]
union_feature = rois_feature_all[:, node_num:, :].reshape(
-1, node_num, node_num, feature_num)
if args.load_ssl_model:
img_feat = ssl_model(
img, image_bboxes) # [batch_size, max_person+1, feat_dim]
rois_feature = torch.cat((rois_feature, img_feat[:, -1:, :]),
dim=1)
all_feat.append(rois_feature.cpu().data.numpy())
all_union_feat.append(union_feature.cpu().data.numpy())
batch_time.update(time.time() - end_time)
end_time = time.time()
if batch_idx % args.print_freq == 0:
logger.info('%s Epoch: [%d/%d] '
'Time %.3f (%.3f)\t' %
(eval_type, batch_idx, len(loader), batch_time.val,
batch_time.avg))
all_feat = np.concatenate(all_feat)
all_union_feat = np.concatenate(all_union_feat)
np.save(feat_filename, all_feat)
np.save(union_filename, all_union_feat)
class_weight, class_count = [], []
if eval_type == 'train':
dataset = SRDataset(all_feat, all_union_feat, max_person=args.max_person+1, image_dir=args.images_root, \
images_list=args.train_file_pre + '_images.txt',
relations_list=args.train_file_pre + '_relation.json', image_size=args.image_size)
class_weight, class_count = dataset.class_weight()
batch_size = args.batch_size
is_shuffle = True
elif eval_type == 'valid':
dataset = SRDataset(all_feat, all_union_feat, max_person=args.max_person+1,image_dir=args.images_root, \
images_list=args.valid_file_pre + '_images.txt',
relations_list=args.valid_file_pre + '_relation.json', image_size=args.image_size)
batch_size = args.test_batch_size
is_shuffle = False
else:
dataset = SRDataset(all_feat, all_union_feat, max_person=args.max_person+1,image_dir=args.images_root, \
images_list=args.test_file_pre + '_images.txt', \
relations_list=args.test_file_pre + '_relation.json', image_size=args.image_size
)
batch_size = args.test_batch_size
is_shuffle = False
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
shuffle=is_shuffle,
num_workers=args.num_workers,
worker_init_fn=np.random.seed(
args.manualSeed))
return dataloader, class_weight, class_count
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 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():
global BEST_LOSS
cudnn.benchmark = True
start_epoch = cfg.OPOSE.start_epoch # start from epoch 0 or last checkpoint epoch
# Create ckpt & vis folder
if not os.path.isdir(cfg.OPOSE.ckpt):
os.makedirs(cfg.OPOSE.ckpt)
if not os.path.exists(os.path.join(cfg.OPOSE.ckpt, 'vis')):
os.makedirs(os.path.join(cfg.OPOSE.ckpt, 'vis'))
if args.cfg_file is not None and not cfg.OPOSE.evaluate:
shutil.copyfile(
args.cfg_file,
os.path.join(cfg.OPOSE.ckpt,
args.cfg_file.split('/')[-1]))
model = pose_estimation.PoseModel(num_point=19,
num_vector=19,
pretrained=True)
# # Calculate FLOPs & Param
# n_flops, n_convops, n_params = measure_model(model, cfg.OPOSE.input_size, cfg.OPOSE.input_size)
criterion = nn.MSELoss().cuda()
# Dataset and Loader
train_dataset = dataset.CocoOpenposeData(
cfg,
cfg.OPOSE.data_root,
cfg.OPOSE.info_root,
'train2017',
transformer=transforms.Compose([
transforms.RandomResized(),
transforms.RandomRotate(40),
transforms.RandomCrop(368),
transforms.RandomHorizontalFlip(),
]))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=int(cfg.workers),
pin_memory=True)
if cfg.OPOSE.validate or cfg.OPOSE.evaluate:
val_dataset = dataset.CocoOpenposeData(
cfg,
cfg.OPOSE.data_root,
cfg.OPOSE.info_root,
'val2017',
transformer=transforms.Compose(
[transforms.TestResized(cfg.OPOSE.input_size)]))
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=int(cfg.workers),
pin_memory=True)
# Load nets into gpu
if NUM_GPUS > 1:
model = torch.nn.DataParallel(model, device_ids=gpu).cuda()
# Set up optimizers
params, multiple = get_parameters(model, cfg, False)
optimizer = torch.optim.SGD(params,
cfg.OPOSE.base_lr,
momentum=cfg.OPOSE.momentum,
weight_decay=cfg.OPOSE.weight_decay)
# Resume training
title = 'Pytorch-OPOSE-{}-{}'.format(cfg.OPOSE.arch_encoder,
cfg.OPOSE.arch_decoder)
if cfg.OPOSE.resume:
# Load checkpoint.
print("==> Resuming from checkpoint '{}'".format(cfg.OPOSE.resume))
assert os.path.isfile(
cfg.OPOSE.resume), 'Error: no checkpoint directory found!'
ckpt = torch.load(cfg.OPOSE.resume)
BEST_LOSS = ckpt['best_loss']
start_epoch = ckpt['epoch']
try:
model.module.load_state_dict(ckpt['state_dict'])
except:
model.load_state_dict(ckpt['state_dict'])
optimizer.load_state_dict(ckpt['optimizer'])
logger = Logger(os.path.join(cfg.OPOSE.ckpt, 'log.txt'),
title=title,
resume=True)
else:
logger = Logger(os.path.join(cfg.OPOSE.ckpt, 'log.txt'), title=title)
logger.set_names(
['epoch', 'Learning Rate', 'Train Loss', 'Valid Loss'])
# Train and val
for epoch in range(start_epoch, cfg.OPOSE.epochs):
print('\nEpoch: [{}/{}] | LR: {:.8f} '.format(epoch + 1,
cfg.OPOSE.epochs,
cfg.OPOSE.base_lr))
train_loss = train(train_loader, model, criterion, optimizer, epoch,
USE_CUDA)
if cfg.OPOSE.validate:
test_loss = test(val_loader, model, criterion, optimizer, epoch,
USE_CUDA)
else:
test_loss = 0.0, 0.0
# Append logger file
logger.append([epoch, cfg.OPOSE.base_lr, train_loss, test_loss])
# Save model
save_checkpoint(model, optimizer, test_loss, epoch)
# Adjust learning rate
adjust_learning_rate(optimizer, epoch)
# Draw curve
try:
draw_curve('model', cfg.OPOSE.ckpt)
print('==> Success saving log curve...')
except:
print('==> Saving log curve error...')
logger.close()
try:
savefig(os.path.join(cfg.OPOSE.ckpt, 'log.eps'))
shutil.copyfile(
os.path.join(cfg.OPOSE.ckpt, 'log.txt'),
os.path.join(
cfg.OPOSE.ckpt, 'log{}.txt'.format(
datetime.datetime.now().strftime('%Y%m%d%H%M%S'))))
except:
print('Copy log error.')
print('==> Training Done!')
print('==> Best acc: {:.4f}%'.format(BEST_LOSS))
def main(args):
if args.gpu is None:
args.gpu = str(os.environ["CUDA_VISIBLE_DEVICES"])
else:
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
device = torch.device('cuda')
best_acc = 0
torch.manual_seed(0)
np.random.seed(0)
random.seed(0)
num_gpu = len(str(args.gpu).split(','))
args.batch_size = num_gpu * args.batch_size
print('=> Effective BatchSize = %d' % args.batch_size)
args.img_path, args.model_path, args.exp_path = set_path(args)
### classifier model ###
num_class_dict = {
'ucf101': 101,
'hmdb51': 51,
'k400': 400,
'ucf101-f': 101,
'hmdb51-f': 51,
'k400-f': 400
}
args.num_class = num_class_dict[args.dataset]
if args.train_what == 'last': # for linear probe
args.final_bn = True
args.final_norm = True
args.use_dropout = False
else: # for training the entire network
args.final_bn = False
args.final_norm = False
args.use_dropout = True
if args.model == 'lincls':
model = LinearClassifier(network=args.net,
num_class=args.num_class,
dropout=args.dropout,
use_dropout=args.use_dropout,
use_final_bn=args.final_bn,
use_l2_norm=args.final_norm)
else:
raise NotImplementedError
model.to(device)
### optimizer ###
if args.train_what == 'last':
print('=> [optimizer] only train last layer')
params = []
for name, param in model.named_parameters():
if 'backbone' in name:
param.requires_grad = False
else:
params.append({'params': param})
elif args.train_what == 'ft':
print('=> [optimizer] finetune backbone with smaller lr')
params = []
for name, param in model.named_parameters():
if 'backbone' in name:
params.append({'params': param, 'lr': args.lr / 10})
else:
params.append({'params': param})
else: # train all
params = []
print('=> [optimizer] train all layer')
for name, param in model.named_parameters():
params.append({'params': param})
if args.train_what == 'last':
print('\n===========Check Grad============')
for name, param in model.named_parameters():
if param.requires_grad:
print(name, param.requires_grad)
print('=================================\n')
if args.optim == 'adam':
optimizer = optim.Adam(params, lr=args.lr, weight_decay=args.wd)
elif args.optim == 'sgd':
optimizer = optim.SGD(params,
lr=args.lr,
weight_decay=args.wd,
momentum=0.9)
else:
raise NotImplementedError
model = torch.nn.DataParallel(model)
model_without_dp = model.module
ce_loss = nn.CrossEntropyLoss()
args.iteration = 1
### test: higher priority ###
if args.test:
if os.path.isfile(args.test):
print("=> loading testing checkpoint '{}'".format(args.test))
checkpoint = torch.load(args.test,
map_location=torch.device('cpu'))
epoch = checkpoint['epoch']
state_dict = checkpoint['state_dict']
if args.retrieval_ucf or args.retrieval_full: # if directly test on pretrained network
new_dict = {}
for k, v in state_dict.items():
k = k.replace('encoder_q.0.', 'backbone.')
new_dict[k] = v
state_dict = new_dict
try:
model_without_dp.load_state_dict(state_dict)
except:
neq_load_customized(model_without_dp, state_dict, verbose=True)
else:
print("[Warning] no checkpoint found at '{}'".format(args.test))
epoch = 0
print("[Warning] if test random init weights, press c to continue")
import ipdb
ipdb.set_trace()
args.logger = Logger(path=os.path.dirname(args.test))
args.logger.log('args=\n\t\t' + '\n\t\t'.join(
['%s:%s' % (str(k), str(v)) for k, v in vars(args).items()]))
transform_test_cuda = transforms.Compose([
T.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
channel=1)
])
if args.retrieval:
test_retrieval(model, ce_loss, transform_test_cuda, device, epoch,
args)
elif args.center_crop or args.five_crop or args.ten_crop:
transform = get_transform('test', args)
test_dataset = get_data(transform, 'test', args)
test_10crop(test_dataset, model, ce_loss, transform_test_cuda,
device, epoch, args)
else:
raise NotImplementedError
sys.exit(0)
### data ###
transform_train = get_transform('train', args)
train_loader = get_dataloader(get_data(transform_train, 'train', args),
'train', args)
transform_val = get_transform('val', args)
val_loader = get_dataloader(get_data(transform_val, 'val', args), 'val',
args)
transform_train_cuda = transforms.Compose([
T.RandomHorizontalFlip(),
T.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
channel=1)
]) # ImageNet
transform_val_cuda = transforms.Compose([
T.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
channel=1)
]) # ImageNet
print('===================================')
### restart training ###
if args.resume:
if os.path.isfile(args.resume):
checkpoint = torch.load(args.resume, map_location='cpu')
args.start_epoch = checkpoint['epoch'] + 1
args.iteration = checkpoint['iteration']
best_acc = checkpoint['best_acc']
state_dict = checkpoint['state_dict']
try:
model_without_dp.load_state_dict(state_dict)
except:
print('[WARNING] resuming training with different weights')
neq_load_customized(model_without_dp, state_dict, verbose=True)
print("=> load resumed checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
try:
optimizer.load_state_dict(checkpoint['optimizer'])
except:
print(
'[WARNING] failed to load optimizer state, initialize optimizer'
)
else:
print("[Warning] no checkpoint found at '{}', use random init".
format(args.resume))
elif args.pretrain:
if os.path.isfile(args.pretrain):
checkpoint = torch.load(args.pretrain, map_location='cpu')
state_dict = checkpoint['state_dict']
new_dict = {}
for k, v in state_dict.items():
k = k.replace('encoder_q.0.', 'backbone.')
new_dict[k] = v
state_dict = new_dict
try:
model_without_dp.load_state_dict(state_dict)
except:
neq_load_customized(model_without_dp, state_dict, verbose=True)
print("=> loaded pretrained checkpoint '{}' (epoch {})".format(
args.pretrain, checkpoint['epoch']))
else:
print("[Warning] no checkpoint found at '{}', use random init".
format(args.pretrain))
else:
print("=> train from scratch")
torch.backends.cudnn.benchmark = True
# plot tools
writer_val = SummaryWriter(logdir=os.path.join(img_path, 'val'))
writer_train = SummaryWriter(logdir=os.path.join(img_path, 'train'))
args.val_plotter = TB.PlotterThread(writer_val)
args.train_plotter = TB.PlotterThread(writer_train)
args.logger = Logger(path=args.img_path)
args.logger.log('args=\n\t\t' + '\n\t\t'.join(
['%s:%s' % (str(k), str(v)) for k, v in vars(args).items()]))
# main loop
for epoch in range(args.start_epoch, args.epochs):
np.random.seed(epoch)
random.seed(epoch)
adjust_learning_rate(optimizer, epoch, args)
train_one_epoch(train_loader, model, ce_loss, optimizer,
transform_train_cuda, device, epoch, args)
if epoch % args.eval_freq == 0:
_, val_acc = validate(val_loader, model, ce_loss,
transform_val_cuda, device, epoch, args)
# save check_point
is_best = val_acc > best_acc
best_acc = max(val_acc, best_acc)
state_dict = model_without_dp.state_dict()
save_dict = {
'epoch': epoch,
'state_dict': state_dict,
'best_acc': best_acc,
'optimizer': optimizer.state_dict(),
'iteration': args.iteration
}
save_checkpoint(save_dict,
is_best,
1,
filename=os.path.join(args.model_path,
'epoch%d.pth.tar' % epoch),
keep_all=False)
print('Training from ep %d to ep %d finished' %
(args.start_epoch, args.epochs))
sys.exit(0)
