def set_transformers(self):
self.transforms = {
self.augmentation_name[0]:
transforms.Compose([ToTensor()]),
self.augmentation_name[1]:
transforms.Compose([
# Choose from tranforms_list randomly
transforms.RandomChoice(self.randomchoice['choice1']),
ToTensor()
]),
self.augmentation_name[2]:
transforms.Compose([
transforms.RandomChoice(self.randomchoice['choice2']),
ToTensor()
]),
self.augmentation_name[3]:
transforms.Compose([
transforms.RandomChoice(self.randomchoice['choice3']),
ToTensor()
]),
self.augmentation_name[4]:
transforms.Compose([
transforms.RandomChoice(self.randomchoice['choice4']),
ToTensor()
])
}
def preprocess(self, phase):
if phase == 'train':
preprocess = transforms.Compose([
FixedResize(self.output_size),
Translate(50, 255),
# RandomScale(),
CutOut(64),
RandomHorizontalFlip(),
AdjustColor(self.factor),
RandomGaussianBlur(self.radius),
Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
ToTensor(),
])
elif phase == 'val':
preprocess = transforms.Compose([
FixedResize(self.output_size),
Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
ToTensor(),
])
elif phase == 'test':
preprocess = transforms.Compose([
FixedResize(self.output_size, is_resize=False),
Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
ToTensor(),
])
else:
raise NotImplementedError
return preprocess
def load_dataset(self):
self.dataset = self.dataset_class(txt_file='testing.txt',
root_dir=self.root_dir,
transform=transforms.Compose([
Resize((64, 64)),
ToTensor()
])
)
self.dataloader = DataLoader(self.dataset, batch_size=self.batch_size,
shuffle=self.is_shuffle, num_workers=self.num_workers)
def __init__(self, clean_file_path, noisy_file_path, frame_size,
frame_shift):
self.clean_test_name = os.listdir(clean_file_path)
self.noisy_test_name = os.listdir(noisy_file_path)
self.noisy_file_path = noisy_file_path
self.clean_file_path = clean_file_path
self.get_frames = SignalToFrames(frame_size=frame_size,
frame_shift=frame_shift)
self.to_tensor = ToTensor()
def set_transformers(self):
self.transforms = {
self.augmentation_name[0]:
transforms.Compose(
[ToPILImage(),
ToTensor(),
Resize(self.resize),
OrigPad()]),
self.augmentation_name[1]:
transforms.Compose([
ToPILImage(),
# Choose from tranforms_list randomly
transforms.RandomChoice(self.randomchoice['choice1']),
ToTensor(),
Resize(self.resize),
OrigPad()
]),
self.augmentation_name[2]:
transforms.Compose([
ToPILImage(),
transforms.RandomChoice(self.randomchoice['choice2']),
ToTensor(),
Resize(self.resize),
OrigPad()
]),
self.augmentation_name[3]:
transforms.Compose([
ToPILImage(),
transforms.RandomChoice(self.randomchoice['choice3']),
ToTensor(),
Resize(self.resize),
OrigPad()
]),
self.augmentation_name[4]:
transforms.Compose([
ToPILImage(),
transforms.RandomChoice(self.randomchoice['choice4']),
ToTensor(),
Resize(self.resize),
OrigPad()
])
}
def __init__(self, file_path, frame_size=512, frame_shift=256):
#self.filename = filename
with open(file_path, 'r') as validation_file_list:
self.file_list = [
line.strip() for line in validation_file_list.readlines()
]
self.get_frames = SignalToFrames(frame_size=frame_size,
frame_shift=frame_shift)
self.to_tensor = ToTensor()
def __init__(self, video_path, txt_path, max_frame_len=14, trainflag=True):
'''
video_path:视频路径
txt_path:文本路径
trainflag:数据集类型：决定样本上的可用的可选变换
'''
self.video_path = video_path
self.txt_path = txt_path
self.max_frame_len = max_frame_len
self.trainflag = trainflag
self.common_transform = transforms.Compose([
Rescale(114),
CenterCrop((112, 224)),
ToTensor(),
ColorNormalize()
])
self.data = self._load_data(self.txt_path)
def __init__(self,
file_path,
frame_size=512,
frame_shift=256,
nsamples=64000):
# file_path is the path of training dataset
# option1: '/data/KaiWang/pytorch_learn/pytorch_for_speech/dataset/timit_mix/trainset/two_data'
# option2 : .txt file format file_path='/data/KaiWang/pytorch_learn/pytorch_for_speech/DDAEC/train_file_list'
#self.file_list = glob.glob(os.path.join(file_path, '*'))
with open(file_path, 'r') as train_file_list:
self.file_list = [
line.strip() for line in train_file_list.readlines()
]
self.nsamples = nsamples
self.get_frames = SignalToFrames(frame_size=frame_size,
frame_shift=frame_shift)
self.to_tensor = ToTensor()
help='path to save output')
args = parser.parse_args()
print(args)
device = torch.device("cuda:2" if torch.cuda.is_available() else "cpu")
model = Hybird(args).to(device)
path = "/home/yinzi/data4/fcn_res/exp_stage2/checkpoints7fc8d2bc/best.pth.tar"
state = torch.load(path, map_location=device)
model.load_state_dict(state['model'])
print(path + " loaded")
root_dir = {'test': "/home/yinzi/data4/warped/test"}
transforms_list = {'test': transforms.Compose([ToTensor()])}
# DataLoader
Dataset = Warped_HelenDataset(root_dir=root_dir['test'],
mode='test',
transform=transforms_list['test'])
dataloader = DataLoader(Dataset, batch_size=16, shuffle=True, num_workers=4)
for batch in dataloader:
image = batch['image'].to(device)
name = batch['name']
boxes = batch['boxes']
warp_box = batch['warp_boxes']
N = image.shape[0]
inverse_box = torch.round(warp_box * 256. + 256.).float().numpy()
assert boxes.shape == (N, 4, 4)
# Dataset Read_in Part
root_dir = {
'train': "/home/yinzi/data4/warped/train",
'val': "/home/yinzi/data4/warped/val",
'test': "/home/yinzi/data4/warped/test"
}
class DataLoaderX(DataLoader):
def __iter__(self):
return BackgroundGenerator(super().__iter__())
transforms_list = {
'train': transforms.Compose([PrepareLabels((128, 128)),
ToTensor()]),
'val': transforms.Compose([PrepareLabels((128, 128)),
ToTensor()]),
'test': transforms.Compose([PrepareLabels((128, 128)),
ToTensor()])
}
Dataset = {
'val':
Warped_HelenDataset(root_dir=root_dir['val'],
mode='val',
transform=transforms_list['val'])
}
if args.datamore:
# Stage 1 augmentation
stage1_augmentation = Stage2DataAugmentation(dataset=Warped_HelenDataset,
device = torch.device("cuda")
else:
device = torch.device("cpu")
# Load data
def make_dataset(file, dir_name, trans, fg_indexs=None, bg_indexs=set([])):
return ImageDataset(txt_file=file, root_dir='data/facial_parts/'+dir_name, fg_indexs=fg_indexs, bg_indexs=bg_indexs,
transform=transforms.Compose(trans))
train_datasets = {}
valid_datasets = {}
test_datasets = {}
## Training set
train_datasets['eyebrow'] = ConcatDataset([make_dataset('exemplars.txt', 'eyebrow1', fg_indexs=set([2]), trans=[DataArg(),ToTensor()]),
make_dataset('exemplars.txt', 'eyebrow2', fg_indexs=set([3]), trans=[Invert(), DataArg(),ToTensor()])])
train_datasets['eye'] = ConcatDataset([make_dataset('exemplars.txt', 'eye1', fg_indexs=set([4]), trans=[DataArg(),ToTensor()]),
make_dataset('exemplars.txt', 'eye2', fg_indexs=set([5]), trans=[Invert(), DataArg(), ToTensor()])])
train_datasets['nose'] = make_dataset('exemplars.txt', 'nose', fg_indexs=set([6]), trans=[DataArg(),ToTensor()])
train_datasets['mouth'] = make_dataset('exemplars.txt', 'mouth', fg_indexs=set([7,8,9]), trans=[DataArg(), ToTensor()])
## Validation set
valid_datasets['eyebrow'] = ConcatDataset([make_dataset('tuning.txt', 'eyebrow1', fg_indexs=set([2]), trans=[ToTensor()]),
make_dataset('tuning.txt', 'eyebrow2', fg_indexs=set([3]), trans=[Invert(), ToTensor()])])
valid_datasets['eye'] = ConcatDataset([make_dataset('tuning.txt', 'eye1', fg_indexs=set([4]), trans=[ToTensor()]),
make_dataset('tuning.txt', 'eye2', fg_indexs=set([5]), trans=[Invert(), ToTensor()])])
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
resize_num = 64
part_width = 64
part_mouth = 80
## Load test data
test_dataset = ImageDataset(txt_file='testing.txt',
root_dir='data/SmithCVPR2013_dataset_resized',
bg_indexs=set([0,1,10]),
transform=transforms.Compose([
Rescale((resize_num,resize_num)),
ToTensor(),
]))
test_loader = DataLoader(test_dataset, batch_size=args.batch_size,
shuffle=True, num_workers=4)
unresized_dataset = ImageDataset(txt_file='testing.txt',
root_dir='data/SmithCVPR2013_dataset_resized',
transform=ToTensor(), calc_bg=False)
## Load models
model = pickle.load(open('res/saved-model.pth', 'rb'))
model = model.to(device)
names = ['eyebrow', 'eye', 'nose', 'mouth']
models={}
args = parser.parse_args()
print(args)
# Dataset Read_in Part
img_root_dir = "/data1/yinzi/datas"
# root_dir = '/home/yinzi/Downloads/datas'
part_root_dir = "/data1/yinzi/facial_parts"
root_dir = {'image': img_root_dir, 'parts': part_root_dir}
txt_file_names = {'train': "exemplars.txt", 'val': "tuning.txt"}
twostage_Dataset = {
x: TwoStepData(txt_file=txt_file_names[x],
root_dir=root_dir,
transform=transforms.Compose(
[Resize((64, 64)),
ToTensor(), Normalize()]))
for x in ['train', 'val']
}
two_dataloader = {
x: DataLoader(twostage_Dataset[x],
batch_size=args.batch_size,
shuffle=True,
num_workers=16)
for x in ['train', 'val']
}
class TrainModel(TemplateModel):
def __init__(self):
super(TrainModel, self).__init__()
def main():
valid_datasets = {}
## Validation set
valid_datasets['eyebrow'] = ConcatDataset([
make_dataset('tuning.txt',
'eyebrow1',
fg_indexs=set([2]),
trans=[ToTensor()]),
make_dataset('tuning.txt',
'eyebrow2',
fg_indexs=set([3]),
trans=[ToTensor(), Invert()])
])
valid_datasets['eye'] = ConcatDataset([
make_dataset('tuning.txt',
'eye1',
fg_indexs=set([4]),
trans=[ToTensor()]),
make_dataset('tuning.txt',
'eye2',
fg_indexs=set([5]),
trans=[ToTensor(), Invert()])
])
valid_datasets['nose'] = make_dataset('tuning.txt',
'nose',
fg_indexs=set([6]),
trans=[ToTensor()])
valid_datasets['mouth'] = make_dataset('tuning.txt',
'mouth',
fg_indexs=set([7, 8, 9]),
trans=[ToTensor()])
names = ['eyebrow', 'eye', 'nose', 'mouth']
l = {'eyebrow': 2, 'eye': 2, 'nose': 2, 'mouth': 4}
for name in names:
print(name, "...")
model = pickle.load(open('res/saved-model-%s.pth' % name, 'rb'))
#model model.to(device)
mod = Modulator(l[name])
mod = mod.to(device)
valid_loader = DataLoader(valid_datasets[name],
batch_size=230,
shuffle=True,
num_workers=4)
batch = [b for b in valid_loader][0]
image, labels = batch['image'].to(device), batch['labels'].to(device)
observations = model(image)
loss_before = mod.loss_fn(observations, labels)
mod.fit(observations, labels)
loss_after = mod.loss_fn(mod.forward(observations), labels)
print("Loss Before", loss_before)
print("Loss After", loss_after)
pickle.dump(mod, open('res/saved-modulator-%s.pth' % name, 'wb'))
# Dataset and Dataloader
# Dataset Read_in Part
root_dir = "/data1/yinzi/datas"
parts_root_dir = "/home/yinzi/data3/recroped_parts"
txt_file_names = {
'train': "exemplars.txt",
'val': "tuning.txt",
'test': "testing.txt"
}
transforms_list = {
'train':
transforms.Compose([
ToTensor(),
Resize((128, 128)),
OrigPad()
]),
'val':
transforms.Compose([
ToTensor(),
Resize((128, 128)),
OrigPad()
]),
'test':
transforms.Compose([
ToTensor(),
Resize((128, 128)),
OrigPad()
])
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
resize_num = 64
part_width = 64
part_mouth = 80
## Load test data
test_dataset = ImageDataset(txt_file='testing.txt',
root_dir='data/SmithCVPR2013_dataset_resized',
bg_indexs=set([0, 1, 10]),
transform=transforms.Compose([
Rescale((resize_num, resize_num)),
ToTensor(),
]))
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
unresized_dataset = ImageDataset(txt_file='testing.txt',
root_dir='data/SmithCVPR2013_dataset_resized',
transform=ToTensor(),
calc_bg=False)
## Load models
model = pickle.load(open('res/saved-model.pth', 'rb'))
model = model.to(device)
'train': "/home/yinzi/data4/warped/train",
'val': "/home/yinzi/data4/warped/val",
'test': "/home/yinzi/data4/warped/test"
}
class DataLoaderX(DataLoader):
def __iter__(self):
return BackgroundGenerator(super().__iter__())
transforms_list = {
'train':
transforms.Compose([
PrepareLabels((128, 128)),
ToTensor()
]),
'val':
transforms.Compose([
PrepareLabels((128, 128)),
ToTensor()
]),
'test':
transforms.Compose([
PrepareLabels((128, 128)),
ToTensor()
])
}
Dataset = {'val': Warped_HelenDataset(root_dir=root_dir['val'],
mode='val',
args = parser.parse_args()
print(args)
# Dataset and Dataloader
# Dataset Read_in Part
root_dir = "/data1/yinzi/datas"
parts_root_dir = "/home/yinzi/data3/recroped_parts"
txt_file_names = {
'train': "exemplars.txt",
'val': "tuning.txt",
'test': "testing.txt"
}
transforms_list = {
'train': transforms.Compose([ToTensor(),
Resize((128, 128)),
OrigPad()]),
'val': transforms.Compose([ToTensor(),
Resize((128, 128)),
OrigPad()]),
'test': transforms.Compose([ToTensor(),
Resize((128, 128)),
OrigPad()])
}
# DataLoader
Dataset = {
x: HelenDataset(txt_file=txt_file_names[x],
root_dir=root_dir,
parts_root_dir=parts_root_dir,
args = parser.parse_args()
print(args)
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
resize_num = 64
# Load data
train_dataset = ImageDataset(txt_file='exemplars.txt',
root_dir='data/SmithCVPR2013_dataset_resized_' +
str(resize_num),
bg_indexs=set([0, 1, 10]),
transform=transforms.Compose([ToTensor()]))
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
valid_dataset = ImageDataset(txt_file='tuning.txt',
root_dir='data/SmithCVPR2013_dataset_resized_' +
str(resize_num),
bg_indexs=set([0, 1, 10]),
transform=transforms.Compose([ToTensor()]))
valid_loader = DataLoader(valid_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
