def __call__(self, images, gts, w, h):
if random.uniform(0., 1.) < self.p:
i, j, new_h, new_w = self.get_params(w, h, self.scale, self.ratio)
images = [
transforms.functional.resized_crop(image, i, j, new_h, new_w,
self.size,
self.interpolation)
for image in images
]
# gt transform
offset = np.array([j, i, j, i, 0])
scale = np.array([
self.size[0] / new_w, self.size[1] / new_h,
self.size[0] / new_w, self.size[1] / new_h, 1
])
new_fullframe = np.array([0, 0, self.size[0], self.size[1]])
for i, frame in enumerate(gts):
new_frame = (frame - offset) * scale
# remove bbox out of the frame
inside_mask = overlap_numpy(new_frame[:, :4],
new_fullframe) - self.threshold
inside_mask = inside_mask > 0.
cleared_frame = np.compress(inside_mask, new_frame, axis=0)
if len(cleared_frame) == 0:
cleared_frame = np.array(
[[0., 0., self.size[0], self.size[1], 0.]],
dtype=float)
# clip into the frame
cleared_frame = cleared_frame.reshape(5, -1)
np.clip(cleared_frame[0],
0.,
self.size[0],
out=cleared_frame[0])
np.clip(cleared_frame[2],
0.,
self.size[0],
out=cleared_frame[2])
np.clip(cleared_frame[1],
0.,
self.size[1],
out=cleared_frame[1])
np.clip(cleared_frame[3],
0.,
self.size[1],
out=cleared_frame[3])
cleared_frame = cleared_frame.reshape(-1, 5)
gts[i] = cleared_frame
else:
# fall back to simple resize
images = transforms.Lambda(lambda frames: [
transforms.Resize(self.size)(frame) for frame in frames
])(images)
w_ratio = float(self.size[0]) / w
h_ratio = float(self.size[1]) / h
ratio = np.array([w_ratio, h_ratio, w_ratio, h_ratio, 1.],
dtype=np.float32)
for i, frame in enumerate(gts):
new_frame = frame * ratio
gts[i] = new_frame
return images, gts, w, h
print('Best val Acc: {:4f}'.format(best_acc))
model.load_state_dict(best_model_wts)
return model, val_acc_history
crop_size = 44
transform_train = transforms.Compose([
transforms.RandomCrop(crop_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor()
])
transform_test = transforms.Compose([
transforms.TenCrop(crop_size),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops]))
])
if args.checkpoint is None:
start_epoch = 0
model = VGG(args.model_name)
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=5e-4)
else:
checkpoint = torch.load(args.checkpoint)
start_epoch = checkpoint['epoch'] + 1
print('\nLoaded checkpoint from epoch %d.\n' % start_epoch)
model = VGG(args.model_name)
model.load_state_dict(checkpoint["model_weights"])
VALIDATION_IMAGE_LIST,
transform=transform_val)
validation_dataloader = DataLoader(dataset=validation_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=4,
pin_memory=True,
drop_last=True)
test_dataset = ChestXrayDataSetMultiLabel(
data_dir=DATA_DIR,
image_list_file=TEST_IMAGE_LIST,
transform=transforms.Compose([
transforms.Resize(256),
transforms.TenCrop(224),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(lambda crops: torch.stack([
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
(crop) for crop in crops
]))
]))
test_dataloader = DataLoader(dataset=test_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=4,
pin_memory=True,
drop_last=True)
# cached_m = torch.load('/Users/haigangliu/Desktop/ML_model_cache/multiclass/multi_class_.pth.tar')
(len(dataset_list), len(activation_function_list), num_epochs))
validation_loss_array = np.zeros(
(len(dataset_list), len(activation_function_list), num_epochs))
test_accuracy_array = np.zeros(
(len(dataset_list), len(activation_function_list)))
test_batch_size = 1000
num_parameters = np.zeros((len(activation_function_list)))
criterion = nn.CrossEntropyLoss()
for index_dataset, (dataset, dataset_name, train_range, validation_range,
test_range, mean_std) in enumerate(
zip(dataset_list, dataset_name_list,
train_range_list, validation_range_list,
test_range_list, mean_std_list)):
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Lambda(lambda x: torch.cat([x, x, x], 0)),
transforms.Normalize(mean_std[0], mean_std[1])
])
train_dataset = dataset('tmp',
train=True,
transform=transform,
download=True)
test_dataset = dataset('tmp',
train=False,
transform=transform,
download=True)
train_dataloader = DataLoader(train_dataset,
batch_size=batch_size,
sampler=SubsetRandomSampler(train_range))
validation_dataloader = DataLoader(
train_dataset,
PATH = 'squeezenet_model_2.pt'
# Parametaers
batch_size = 100
learning_rate = 0.001
num_epochs = 1
momentum = 0.5
transform = transforms.Compose([
transforms.Resize(224),
transforms.ToTensor(),
transforms.Lambda(lambda x: x.expand(3, -1, -1))
])
# Data load
dataset_train = datasets.MNIST('data',
train=True,
download=True,
transform=transform)
dataset_test = datasets.MNIST('data',
train=False,
download=True,
transform=transform)
train_loader = utils.data.DataLoader(dataset_train,
batch_size=batch_size,
def get_mnist_train_valid_loader(data_dir,
batch_size,
random_seed,
valid_size=0.2,
shuffle=True,
show_sample=False,
num_workers=1,
pin_memory=True):
"""
Utility function for loading and returning train and valid
multi-process iterators over the MNIST dataset. A sample
9x9 grid of the images can be optionally displayed.
If using CUDA, num_workers should be set to 1 and pin_memory to True.
Params
------
- data_dir: path directory to the dataset.
- batch_size: how many samples per batch to load.
- augment: whether to apply the data augmentation scheme
mentioned in the paper. Only applied on the train split.
- random_seed: fix seed for reproducibility.
- valid_size: percentage split of the training set used for
the validation set. Should be a float in the range [0, 1].
- shuffle: whether to shuffle the train/validation indices.
- show_sample: plot 9x9 sample grid of the dataset.
- num_workers: number of subprocesses to use when loading the dataset.
- pin_memory: whether to copy tensors into CUDA pinned memory. Set it to
True if using GPU.
Returns
-------
- train_loader: training set iterator.
- valid_loader: validation set iterator.
"""
error_msg = "[!] valid_size should be in the range [0, 1]."
assert ((valid_size >= 0) and (valid_size <= 1)), error_msg
# define transforms
transform = transforms.Compose([
transforms.Pad(2),
transforms.ToTensor(),
transforms.Lambda(lambda x: x.repeat(3,1,1)),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
# load the dataset
train_dataset = datasets.MNIST(root=data_dir, train=True,
download=True, transform=transform)
valid_dataset = datasets.MNIST(root=data_dir, train=True,
download=True, transform=transform)
num_train = len(train_dataset)
indices = list(range(num_train))
split = int(np.floor(valid_size * num_train))
if shuffle == True:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_idx, valid_idx = indices[split:], indices[:split]
train_sampler = SubsetRandomSampler(train_idx)
valid_sampler = SubsetRandomSampler(valid_idx)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size, sampler=train_sampler,
num_workers=num_workers, pin_memory=pin_memory)
valid_loader = torch.utils.data.DataLoader(valid_dataset,
batch_size=1, sampler=valid_sampler,
num_workers=num_workers, pin_memory=pin_memory)
# visualize some images
if show_sample:
sample_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=9,
shuffle=shuffle,
num_workers=num_workers,
pin_memory=pin_memory)
data_iter = iter(sample_loader)
images, labels = data_iter.next()
X = images.numpy()
plot_images(X, labels)
return (train_loader, valid_loader)
def get_svhn_train_valid_loader(data_dir,
batch_size,
random_seed,
valid_size=0.2,
shuffle=True,
show_sample=False,
num_workers=1,
pin_memory=True):
"""
Utility function for loading and returning train and valid
multi-process iterators over the MNIST dataset. A sample
9x9 grid of the images can be optionally displayed.
If using CUDA, num_workers should be set to 1 and pin_memory to True.
Params
------
- data_dir: path directory to the dataset.
- batch_size: how many samples per batch to load.
- augment: whether to apply the data augmentation scheme
mentioned in the paper. Only applied on the train split.
- random_seed: fix seed for reproducibility.
- valid_size: percentage split of the training set used for
the validation set. Should be a float in the range [0, 1].
- shuffle: whether to shuffle the train/validation indices.
- show_sample: plot 9x9 sample grid of the dataset.
- num_workers: number of subprocesses to use when loading the dataset.
- pin_memory: whether to copy tensors into CUDA pinned memory. Set it to
True if using GPU.
Returns
-------
- train_loader: training set iterator.
- valid_loader: validation set iterator.
"""
error_msg = "[!] valid_size should be in the range [0, 1]."
assert ((valid_size >= 0) and (valid_size <= 1)), error_msg
# define transforms
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
# load the dataset
train_dataset = datasets.SVHN(root=data_dir, split='train',
download=True, transform=transform)
valid_dataset = datasets.SVHN(root=data_dir, split='test',
download=True, transform=transform)
num_train = len(train_dataset)
indices = list(range(num_train))
split = int(np.floor(valid_size * num_train))
if shuffle == True:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_idx, valid_idx = indices[split:], indices[:split]
train_sampler = SubsetRandomSampler(train_idx)
valid_sampler = SubsetRandomSampler(valid_idx)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size, sampler=train_sampler,
num_workers=num_workers, pin_memory=pin_memory)
valid_loader = torch.utils.data.DataLoader(valid_dataset,
batch_size=1, sampler=valid_sampler,
num_workers=num_workers, pin_memory=pin_memory)
# visualize some images
if show_sample:
sample_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=9,
shuffle=shuffle,
num_workers=num_workers,
pin_memory=pin_memory)
data_iter = iter(sample_loader)
images, labels = data_iter.next()
X = images.numpy()
plot_images(X, labels)
return (train_loader, valid_loader)
def main():
args = parser.parse_args()
# create model
print("Training model with backbone", args.backbone)
model = make_backbone(args.dim, args.backbone)
# SJH: uncomment to run on a selected GPU
#
b = True
if args.backbone == "mobilenet": # mobilenet
torch.cuda.set_device(args.gpu)
b = False
model = model.cuda()
# SJH: uncomment to run on multiple GPUs - works for ResNet not MobileNet
if b is True: # false for mobilenet
model = torch.nn.DataParallel(model)
# define loss function (criterion) and optimizer
criterion = loss.SoftTriple(args.la, args.gamma, args.tau, args.margin,
args.dim, args.C, args.K).cuda()
optimizer = torch.optim.Adam([{
"params": model.parameters(),
"lr": args.modellr
}, {
"params": criterion.parameters(),
"lr": args.centerlr
}],
eps=args.eps,
weight_decay=args.weight_decay)
cudnn.benchmark = True
# load data
traindir = os.path.join(args.data, 'train')
testdir = os.path.join(args.data, 'test')
# Use this with BN-Inception
if args.backbone == "BN-Inception":
normalize = transforms.Normalize(mean=[104., 117., 128.],
std=[1., 1., 1.])
# SJH for ResNet and EfficientNet
else:
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if args.backbone == "BN-Inception":
input_dim_resize = 256
input_dim_crop = 224
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.Lambda(RGB2BGR), # SJH BN-Inception is BGR
# SJH was 224 for bn-inception and mobilenet and 299 for pytorch inception
transforms.RandomResizedCrop(input_dim_crop),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
# SJH BN-Inception needs scale
transforms.Lambda(lambda x: x.mul(255)),
normalize,
]))
else:
input_dim_resize = 1024
input_dim_crop = 598
# input_dim_resize = 512
# input_dim_crop = 299
# TODO: investigate this?
# For EfficientNet with advprop
# normalize = transforms.Lambda(lambda img: img * 2.0 - 1.0)
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(input_dim_crop),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
if args.backbone == "BN-Inception":
test_loader = torch.utils.data.DataLoader(
datasets.ImageFolder(
testdir,
transforms.Compose([
transforms.Lambda(RGB2BGR),
transforms.Resize(input_dim_resize),
transforms.CenterCrop(input_dim_crop),
transforms.ToTensor(),
# SJH BN-Inception needs scale
transforms.Lambda(lambda x: x.mul(255)),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
else:
test_loader = torch.utils.data.DataLoader(datasets.ImageFolder(
testdir,
transforms.Compose([
transforms.Resize(input_dim_resize),
transforms.CenterCrop(input_dim_crop),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
best_nmi = 0
for epoch in range(args.start_epoch, args.epochs):
print('Training in Epoch[{}]'.format(epoch))
adjust_learning_rate(optimizer, epoch, args)
# train for one epoch
train(train_loader, model, criterion, optimizer, args)
# Run validation
nmi, recall = validate(test_loader, model, args)
print(
'Recall@1, 2, 4, 8: {recall[0]:.3f}, {recall[1]:.3f}, {recall[2]:.3f}, {recall[3]:.3f}; NMI: {nmi:.3f} \n'
.format(recall=recall, nmi=nmi))
# Save the best model
if nmi > best_nmi:
best_nmi = nmi
print("Saving new best model!")
fn = "{}.pth".format(f"best_model_{epoch}_eflite")
torch.save(model, fn)
print("Model saved to", fn)
# evaluate on validation set
nmi, recall = validate(test_loader, model, args)
print(
'Recall@1, 2, 4, 8: {recall[0]:.3f}, {recall[1]:.3f}, {recall[2]:.3f}, {recall[3]:.3f}; NMI: {nmi:.3f} \n'
.format(recall=recall, nmi=nmi))
# Save the model
print("Saving model!")
fn = "{}.pth".format("last_model_eflite")
torch.save(model, fn)
print("Model saved to", fn)
from torchvision import transforms
import numpy as np
import pandas as pd
from time import time
from models.mtcnn import MTCNN, prewhiten
from models.inception_resnet_v1 import InceptionResnetV1
trans = transforms.Compose([
transforms.Resize(512), np.int_,
transforms.ToTensor(), torch.Tensor.byte
])
trans_cropped = transforms.Compose([
np.array, torch.tensor, torch.Tensor.float,
transforms.Lambda(lambda x: x.permute(2, 0, 1)), prewhiten
])
def get_image(path, trans):
img = Image.open(path)
img = trans(img)
return img
mtcnn_pt = MTCNN()
resnet_pt = InceptionResnetV1(pretrained='vggface2').eval()
names = [
'bradley_cooper', 'shea_whigham', 'paul_rudd', 'kate_siegel',
'angelina_jolie'
base_save_path = '%s_%s_cellSize%d/'%(backbone_type, opt_name, S)
if not os.path.exists(base_save_path):
os.makedirs(base_save_path)
log_name = 'train'
logger = create_logger(base_save_path, log_name)
my_vis = Visual(base_save_path[:-1])
# backbone_net_p.load_state_dict(torch.load('densenet_sgd_S7_yolo.pth'))
lossLayer = YOLOLossV1(batch_size, S, B, clsN, lbd_coord, lbd_no_obj, _logger=logger, _vis=my_vis)
backbone_net_p.train()
transform = transforms.Compose([
transforms.Lambda(cv_resize),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
data_base = 'datasets/'
train_data_name = data_base + 'train.txt'
# test_data_name = '2007_train.txt'
test_data_name = data_base + '2007_test.txt'
train_dataset = yoloDataset(list_file=train_data_name, train=True, transform = transform, device=device, little_train=False, S=S)
train_loader = DataLoader(train_dataset,batch_size=batch_size,shuffle=True, num_workers=4)
test_dataset = yoloDataset(list_file=test_data_name,train=False,transform = transform, device=device, little_train=False, with_file_path=True, S=S)
test_loader = DataLoader(test_dataset,batch_size=batch_size,shuffle=False)#, num_workers=4)
import src.dataloader
import src.train3d
# -
# ## Create Dataset and DataLoader
# +
import src.dataloader
import importlib
importlib.reload(src.dataloader)
augmentor = transforms.Compose([
transforms.Lambda(lambda x: x.repeat(3, 1, 1, 1).permute(3, 0, 1, 2)),
src.dataloader.Resize(256),
])
path = '/data/larson2/RCC_dl/new/clear_cell'
train_ds = src.dataloader.RCCDataset_h5(path, mode='train', transform=augmentor)
train_loader = DataLoader(train_ds, batch_size=1, shuffle=True, num_workers=4, drop_last=False)
val_ds = src.dataloader.RCCDataset_h5(path, mode='val', transform=augmentor)
val_loader = DataLoader(val_ds, batch_size=1, shuffle=True, num_workers=4, drop_last=False)
print(train_ds[0][0].shape)
print(val_ds[0][0].shape)
########################################################################################
torch.manual_seed(args.seed)
if args.cuda and torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
device = torch.device('cuda')
else:
device = torch.device('cpu')
########################################################################################
# LOAD DATA
########################################################################################
transform = trans.Compose(
[trans.ToTensor(),
trans.Lambda(lambda x: x.view(-1, args.input_size))])
# Load data
data_path, batch_size = args.data, args.batch_size
train_data = dataset.MNIST(root=data_path,
train=True,
transform=transform,
download=True)
# Split train data into training and validation sets
N = len(train_data)
val_size = int(N * 0.2)
train_data, validation_data = torch.utils.data.random_split(
train_data, [N - val_size, val_size])
def main():
global args
args = parser.parse_args()
#fix random seeds
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
best_prec1 = 0
# load model
model = load_model(args.model)
model.cuda()
cudnn.benchmark = True
# freeze the features layers
for param in model.features.parameters():
param.requires_grad = False
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
# data loading code
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if args.tencrops:
transformations_val = [
transforms.Resize(256),
transforms.TenCrop(224),
transforms.Lambda(lambda crops: torch.stack(
[normalize(transforms.ToTensor()(crop)) for crop in crops])),
]
else:
transformations_val = [
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(), normalize
]
transformations_train = [
transforms.Resize(256),
transforms.CenterCrop(256),
transforms.RandomCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(), normalize
]
train_dataset = datasets.ImageFolder(
traindir, transform=transforms.Compose(transformations_train))
val_dataset = datasets.ImageFolder(
valdir, transform=transforms.Compose(transformations_val))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=int(args.batch_size /
2),
shuffle=False,
num_workers=args.workers)
# logistic regression
reglog = RegLog(args.conv, len(train_dataset.classes)).cuda()
optimizer = torch.optim.SGD(filter(lambda x: x.requires_grad,
reglog.parameters()),
args.lr,
momentum=args.momentum,
weight_decay=10**args.weight_decay)
# create logs
exp_log = os.path.join(args.exp, 'log')
if not os.path.isdir(exp_log):
os.makedirs(exp_log)
loss_log = Logger(os.path.join(exp_log, 'loss_log'))
prec1_log = Logger(os.path.join(exp_log, 'prec1'))
prec5_log = Logger(os.path.join(exp_log, 'prec5'))
for epoch in range(args.epochs):
end = time.time()
# train for one epoch
train(train_loader, model, reglog, criterion, optimizer, epoch)
# evaluate on validation set
prec1, prec5, loss = validate(val_loader, model, reglog, criterion)
loss_log.log(loss)
prec1_log.log(prec1)
prec5_log.log(prec5)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
if is_best:
filename = 'model_best.pth.tar'
else:
filename = 'checkpoint.pth.tar'
torch.save(
{
'epoch': epoch + 1,
'arch': 'alexnet',
'state_dict': model.state_dict(),
'prec5': prec5,
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, os.path.join(args.exp, filename))
def __init__(self,
root: str,
normal_class: int,
preproc: str,
nominal_label: int,
supervise_mode: str,
noise_mode: str,
oe_limit: int,
online_supervision: bool,
logger: Logger = None):
"""
AD dataset for Cifar-10.
:param root: root directory where data is found or is to be downloaded to
:param normal_class: the class considered nominal
:param preproc: the kind of preprocessing pipeline
:param nominal_label: the label that marks nominal samples in training. The scores in the heatmaps always
rate label 1, thus usually the nominal label is 0, s.t. the scores are anomaly scores.
:param supervise_mode: the type of generated artificial anomalies.
See :meth:`fcdd.datasets.bases.TorchvisionDataset._generate_artificial_anomalies_train_set`.
:param noise_mode: the type of noise used, see :mod:`fcdd.datasets.noise_mode`.
:param oe_limit: limits the number of different anomalies in case of Outlier Exposure (defined in noise_mode)
:param online_supervision: whether to sample anomalies online in each epoch,
or offline before training (same for all epochs in this case).
:param logger: logger
"""
super().__init__(root, logger=logger)
self.n_classes = 2 # 0: normal, 1: outlier
self.shape = (3, 32, 32)
self.raw_shape = (3, 32, 32)
self.normal_classes = tuple([normal_class])
self.outlier_classes = list(range(0, 10))
self.outlier_classes.remove(normal_class)
assert nominal_label in [0, 1]
self.nominal_label = nominal_label
self.anomalous_label = 1 if self.nominal_label == 0 else 0
if self.nominal_label != 0:
print('Swapping labels, i.e. anomalies are 0 and nominals are 1.')
# Pre-computed min and max values (after applying LCN) from train data per class
min_max_l1 = [(-28.94083453598571, 13.802961825439636),
(-6.681770233365245, 9.158067708230273),
(-34.924463588638204, 14.419298165027628),
(-10.599172931391799, 11.093187820377565),
(-11.945022995801637, 10.628045447867583),
(-9.691969487694928, 8.948326776180823),
(-9.174940012342555, 13.847014686472365),
(-6.876682005899029, 12.282371383343161),
(-15.603507135507172, 15.2464923804279),
(-6.132882973622672, 8.046098172351265)]
# mean and std of original images per class
mean = [
[0.5256516933441162, 0.5603281855583191, 0.5888723731040955],
[0.4711322784423828, 0.45446228981018066, 0.4471212327480316],
[0.48923906683921814, 0.49146366119384766, 0.423904687166214],
[0.4954785108566284, 0.45636114478111267, 0.4154069721698761],
[0.47155335545539856, 0.46515223383903503, 0.37797248363494873],
[0.49992093443870544, 0.4646056592464447, 0.4164286255836487],
[0.47001829743385315, 0.43829214572906494, 0.34500396251678467],
[0.5019531846046448, 0.47983652353286743, 0.4167139232158661],
[0.4902143180370331, 0.5253947973251343, 0.5546804070472717],
[0.4986417591571808, 0.4852965474128723, 0.4780091941356659]
]
std = [[0.2502202093601227, 0.24083486199378967, 0.2659735083580017],
[0.26806357502937317, 0.2658274173736572, 0.2749459445476532],
[0.22705480456352234, 0.2209445983171463, 0.24337927997112274],
[0.2568431496620178, 0.25227081775665283, 0.25799375772476196],
[0.21732737123966217, 0.20652702450752258, 0.21182335913181305],
[0.2504253387451172, 0.24374878406524658, 0.2489463835954666],
[0.22888341546058655, 0.21856172382831573, 0.2204199582338333],
[0.2430490106344223, 0.243973046541214, 0.25171563029289246],
[0.24962472915649414, 0.24068884551525116, 0.25149762630462646],
[0.2680525481700897, 0.26910799741744995, 0.2810165584087372]]
# different types of preprocessing pipelines, 'lcn' is for using LCN, 'aug{X}' for augmentations
# also contains options for the black center experiments
all_transform = []
if preproc == 'lcn':
test_transform = transform = transforms.Compose([
transforms.ToTensor(),
transforms.Lambda(
lambda x: local_contrast_normalization(x, scale='l1')),
transforms.Normalize([min_max_l1[normal_class][0]] * 3, [
min_max_l1[normal_class][1] - min_max_l1[normal_class][0]
] * 3)
])
elif preproc in ['', None, 'default', 'none']:
test_transform = transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean[normal_class], std[normal_class])
])
elif preproc in ['aug1']:
test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean[normal_class], std[normal_class])
])
transform = transforms.Compose([
transforms.ColorJitter(brightness=0.01,
contrast=0.01,
saturation=0.01,
hue=0.01),
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Lambda(lambda x: x + 0.001 * torch.randn_like(x)),
transforms.Normalize(mean[normal_class], std[normal_class])
])
elif preproc in ['aug1_blackcenter']:
test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean[normal_class], std[normal_class])
])
transform = transforms.Compose([
transforms.ColorJitter(brightness=0.01,
contrast=0.01,
saturation=0.01,
hue=0.01),
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Lambda(lambda x: x + 0.001 * torch.randn_like(x)),
BlackCenter(0.6),
transforms.Normalize(mean[normal_class], std[normal_class])
])
elif preproc in ['aug1_blackcenter_inverted']:
test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean[normal_class], std[normal_class])
])
transform = transforms.Compose([
transforms.ColorJitter(brightness=0.01,
contrast=0.01,
saturation=0.01,
hue=0.01),
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Lambda(lambda x: x + 0.001 * torch.randn_like(x)),
BlackCenter(0.6, inverse=True),
transforms.Normalize(mean[normal_class], std[normal_class])
])
else:
raise ValueError(
'Preprocessing pipeline {} is not known.'.format(preproc))
target_transform = transforms.Lambda(
lambda x: self.anomalous_label
if x in self.outlier_classes else self.nominal_label)
if online_supervision:
all_transform = MultiCompose([
OnlineSupervisor(self, supervise_mode, noise_mode, oe_limit),
*all_transform
])
else:
all_transform = MultiCompose(all_transform)
train_set = MYCIFAR10(root=self.root,
train=True,
download=True,
normal_classes=self.normal_classes,
transform=transform,
target_transform=target_transform,
all_transform=all_transform)
train_set.targets = torch.from_numpy(np.asarray(train_set.targets))
train_set.data = torch.from_numpy(train_set.data).transpose(
1, 3).transpose(2, 3)
self._generate_artificial_anomalies_train_set(
supervise_mode if not online_supervision else 'unsupervised',
noise_mode, oe_limit, train_set, normal_class)
self._test_set = MYCIFAR10(root=self.root,
train=False,
download=True,
normal_classes=self.normal_classes,
transform=test_transform,
target_transform=target_transform)
gpu_id = opt.gpu_id
os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu_id)
use_cuda = torch.cuda.is_available()
print("GPU device %d:" %(gpu_id), use_cuda)
model = EfficientNet.from_name(opt.arch, num_classes=opt.classes,
override_params={'dropout_rate': opt.dropout, 'drop_connect_rate': opt.dropconnect})
model.to('cuda')
cudnn.benchmark = True
best_acc = 0
data_dir = opt.source_dataset
train_dir = os.path.join(data_dir, 'train')
train_aug = transforms.Compose([
transforms.Lambda(lambda img: data_augment(img, opt)),
transforms.Resize(opt.size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
# transforms.RandomErasing(p=0.3, scale=(0.02, 0.10), ratio=(0.3, 3.3), value=0, inplace=True),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
train_loader = DataLoader(datasets.ImageFolder(train_dir, train_aug),
batch_size=opt.train_batch, shuffle=True, num_workers=opt.num_workers, pin_memory=True)
val_dir = os.path.join(data_dir, 'val')
val_aug = transforms.Compose([
transforms.Resize(opt.size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
network.add_monitor(voltages["O"], name="O_voltages")
# Directs network to GPU
if gpu:
network.to("cuda")
# Get MNIST training images and labels.
# Load MNIST data.
dataset = MNIST(
PoissonEncoder(time=time, dt=dt),
None,
root=os.path.join("..", "..", "data", "MNIST"),
download=True,
transform=transforms.Compose(
[transforms.ToTensor(),
transforms.Lambda(lambda x: x * intensity)]),
)
inpt_axes = None
inpt_ims = None
spike_axes = None
spike_ims = None
weights_im = None
weights_im2 = None
voltage_ims = None
voltage_axes = None
# Create a dataloader to iterate and batch data
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=1,
shuffle=True,
def train(args):
check_paths(args)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
kwargs = {'num_workers': 0, 'pin_memory': False}
else:
kwargs = {}
transform = transforms.Compose([
transforms.Scale(args.image_size),
transforms.CenterCrop(args.image_size),
transforms.ToTensor(),
transforms.Lambda(lambda x: x.mul(255))
])
train_dataset = datasets.ImageFolder(args.dataset, transform)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
**kwargs)
style_model = Net(ngf=args.ngf)
if args.resume is not None:
print('Resuming, initializing using weight from {}.'.format(
args.resume))
style_model.load_state_dict(torch.load(args.resume))
print(style_model)
optimizer = Adam(style_model.parameters(), args.lr)
mse_loss = torch.nn.MSELoss()
vgg = Vgg16()
utils.init_vgg16(args.vgg_model_dir)
vgg.load_state_dict(
torch.load(os.path.join(args.vgg_model_dir, "vgg16.weight")))
if args.cuda:
style_model.cuda()
vgg.cuda()
style_loader = utils.StyleLoader(args.style_folder, args.style_size)
tbar = trange(args.epochs)
for e in tbar:
style_model.train()
agg_content_loss = 0.
agg_style_loss = 0.
count = 0
for batch_id, (x, _) in enumerate(train_loader):
n_batch = len(x)
count += n_batch
optimizer.zero_grad()
x = Variable(utils.preprocess_batch(x))
if args.cuda:
x = x.cuda()
style_v = style_loader.get(batch_id)
style_model.setTarget(style_v)
style_v = utils.subtract_imagenet_mean_batch(style_v)
features_style = vgg(style_v)
gram_style = [utils.gram_matrix(y) for y in features_style]
y = style_model(x)
xc = Variable(x.data.clone())
y = utils.subtract_imagenet_mean_batch(y)
xc = utils.subtract_imagenet_mean_batch(xc)
features_y = vgg(y)
features_xc = vgg(xc)
f_xc_c = Variable(features_xc[1].data, requires_grad=False)
content_loss = args.content_weight * mse_loss(
features_y[1], f_xc_c)
style_loss = 0.
for m in range(len(features_y)):
gram_y = utils.gram_matrix(features_y[m])
gram_s = Variable(gram_style[m].data,
requires_grad=False).repeat(
args.batch_size, 1, 1, 1)
style_loss += args.style_weight * mse_loss(
gram_y, gram_s[:n_batch, :, :])
total_loss = content_loss + style_loss
total_loss.backward()
optimizer.step()
agg_content_loss += content_loss.data[0]
agg_style_loss += style_loss.data[0]
if (batch_id + 1) % args.log_interval == 0:
mesg = "{}\tEpoch {}:\t[{}/{}]\tcontent: {:.6f}\tstyle: {:.6f}\ttotal: {:.6f}".format(
time.ctime(), e + 1, count, len(train_dataset),
agg_content_loss / (batch_id + 1),
agg_style_loss / (batch_id + 1),
(agg_content_loss + agg_style_loss) / (batch_id + 1))
tbar.set_description(mesg)
if (batch_id + 1) % (4 * args.log_interval) == 0:
# save model
style_model.eval()
style_model.cpu()
save_model_filename = "Epoch_" + str(e) + "iters_" + str(count) + "_" + \
str(time.ctime()).replace(' ', '_') + "_" + str(
args.content_weight) + "_" + str(args.style_weight) + ".model"
save_model_path = os.path.join(args.save_model_dir,
save_model_filename)
torch.save(style_model.state_dict(), save_model_path)
style_model.train()
style_model.cuda()
tbar.set_description("\nCheckpoint, trained model saved at",
save_model_path)
# save model
style_model.eval()
style_model.cpu()
save_model_filename = "Final_epoch_" + str(args.epochs) + "_" + \
str(time.ctime()).replace(' ', '_') + "_" + str(
args.content_weight) + "_" + str(args.style_weight) + ".model"
save_model_path = os.path.join(args.save_model_dir, save_model_filename)
torch.save(style_model.state_dict(), save_model_path)
print("\nDone, trained model saved at", save_model_path)
def set_seed(SEED):
SEED = 0
torch.cuda.manual_seed(SEED)
torch.manual_seed(SEED)
np.random.seed(SEED)
random.seed(SEED)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
source_transform = transforms.Compose([
# 轉灰階: Canny 不吃 RGB。
transforms.Grayscale(),
# cv2 不吃 skimage.Image，因此轉成np.array後再做cv2.Canny
transforms.Lambda(lambda x: cv2.Canny(np.array(x), 170, 300)),
#transforms.Lambda(lambda x: cv2.bitwise_or(np.uint8(np.absolute(cv2.Sobel(np.array(x), cv2.CV_64F, 1, 0))), \
#np.uint8(np.absolute(cv2.Sobel(np.array(x), cv2.CV_64F, 0, 1))))),
# 重新將np.array 轉回 skimage.Image
transforms.ToPILImage(),
# 水平翻轉 (Augmentation)
transforms.RandomHorizontalFlip(),
# 旋轉15度內 (Augmentation)，旋轉後空的地方補0
transforms.RandomRotation(15),
# 最後轉成Tensor供model使用。
transforms.ToTensor(),
])
target_transform = transforms.Compose([
# 轉灰階: 將輸入3維壓成1維。
transforms.Grayscale(),
# 縮放: 因為source data是32x32，我們將target data的28x28放大成32x32。
parser.add_argument("--base_path",
type=str,
help="top level directory where all repos, etc. live",
default=None)
args = parser.parse_args()
device = 'cuda'
transform_to_faceseg = transforms.Compose([
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
inplace=False)
])
transform_uint8 = transforms.Compose([
# transforms.Resize(256),
transforms.Lambda(convert_to_uint8)
])
# Instantiate models - generation, segmentation, landmarks 2d, landmarks 3d
#
tddfa = load_3ddfa(args, args.base_path)
fa = face_alignment.FaceAlignment(face_alignment.LandmarksType._2D,
flip_input=False)
face_seg = load_facesegment(args.base_path, device)
g_ema = load_generator(args.base_path, device)
if args.save_dir is None:
path_train_data = f'{base_path}/training_data_dfr'
else:
path_train_data = args.save_dir
def test_model(pathDirData, pathFileTest, pathModel, nnArchitecture,
nnClassCount, nnIsTrained, trBatchSize, transResize,
transCrop, launchTimeStamp):
CLASS_NAMES = [
'Atelectasis', 'Cardiomegaly', 'Effusion', 'Infiltration', 'Mass',
'Nodule', 'Pneumonia', 'Pneumothorax', 'Consolidation', 'Edema',
'Emphysema', 'Fibrosis', 'Pleural_Thickening', 'Hernia'
]
cudnn.benchmark = True
model = None
#-------------------- SETTINGS: NETWORK ARCHITECTURE, MODEL LOAD
if nnArchitecture == 'DENSE-NET-121':
model = DenseNet121(nnClassCount, nnIsTrained).cuda()
elif nnArchitecture == 'DENSE-NET-169':
model = DenseNet169(nnClassCount, nnIsTrained).cuda()
elif nnArchitecture == 'RES-NET-50':
model = ResNet50(nnClassCount, nnIsTrained).cuda()
elif nnArchitecture == 'SE-RES-NET-50':
model = SE_ResNet50(nnClassCount, nnIsTrained).cuda()
elif nnArchitecture == 'SE-DENSE-NET-121':
model = SE_DenseNet121(nnClassCount, nnIsTrained).cuda()
if model is not None and DEVICE == 'cuda':
model = torch.nn.DataParallel(model).cuda()
modelCheckpoint = torch.load(pathModel)
model.load_state_dict(modelCheckpoint['state_dict'])
#-------------------- SETTINGS: DATA TRANSFORMS, TEN CROPS
normalize = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
#-------------------- SETTINGS: DATASET BUILDERS
transformList = []
transformList.append(transforms.Resize(transResize))
transformList.append(transforms.TenCrop(transCrop))
transformList.append(
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])))
transformList.append(
transforms.Lambda(lambda crops: torch.stack(
[normalize(crop) for crop in crops])))
transformSequence = transforms.Compose(transformList)
datasetTest = DatasetGenerator(pathImageDirectory=pathDirData,
pathDatasetFile=pathFileTest,
transform=transformSequence)
dataLoaderTest = DataLoader(dataset=datasetTest,
batch_size=trBatchSize,
num_workers=8,
shuffle=False,
pin_memory=True)
outGT = torch.FloatTensor()
outPRED = torch.FloatTensor()
if DEVICE == 'cuda':
outGT = cuda()
outPRED = cuda()
model.eval()
for i, (input, target) in enumerate(dataLoaderTest):
if DEVICE == 'cuda':
target = target.cuda()
outGT = torch.cat((outGT, target), 0)
bs, n_crops, c, h, w = input.size()
varInput = torch.autograd.Variable(input.view(-1, c, h, w).cuda(),
volatile=True)
out = model(varInput)
outMean = out.view(bs, n_crops, -1).mean(1)
outPRED = torch.cat((outPRED, outMean.data), 0)
aurocIndividual = ChexnetTrainer.computeAUROC(outGT, outPRED,
nnClassCount)
aurocMean = np.array(aurocIndividual).mean()
print('AUROC mean ', aurocMean)
for i in range(0, len(aurocIndividual)):
print(CLASS_NAMES[i], ' ', aurocIndividual[i])
return
import time
import os
import PIL
from google.colab import files
import matplotlib
import matplotlib.animation as animation
import matplotlib.pyplot as plt
import torch
from torchvision import transforms
# Always use html5 for animations so they can be rendered inline on Colab.
matplotlib.rcParams['animation.html'] = 'html5'
_IMAGE_UNLOADER = transforms.Compose([
transforms.Lambda(lambda x: x.cpu().clone().squeeze(0)),
transforms.ToPILImage()
])
def get_device():
"""Returns the appropriate device depending on what's available."""
return torch.device("cuda" if torch.cuda.is_available() else "cpu")
def upload_files():
"""Creates a widget to upload files from your local machine to Colab.
The files are saved in '/tmp/<file_name>'.
"""
uploaded = files.upload()
def data_loader(args):
mean_vals = [0.485, 0.456, 0.406]
std_vals = [0.229, 0.224, 0.225]
tsfm_train = transforms.Compose([
transforms.Resize((args.resize_size, args.resize_size)),
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean_vals, std_vals)
])
if args.tencrop == 'True':
func_transforms = [
transforms.Resize(args.resize_size),
transforms.TenCrop(args.crop_size),
transforms.Lambda(lambda crops: torch.stack([
transforms.Normalize(mean_vals, std_vals)
(transforms.ToTensor()(crop)) for crop in crops
])),
]
else:
func_transforms = []
# print input_size, crop_size
if args.resize_size == 0 or args.crop_size == 0:
pass
else:
func_transforms.append(
transforms.Resize((args.resize_size, args.resize_size)))
func_transforms.append(transforms.CenterCrop(args.crop_size))
func_transforms.append(transforms.ToTensor())
func_transforms.append(transforms.Normalize(mean_vals, std_vals))
tsfm_test = transforms.Compose(func_transforms)
if args.dataset == 'ILSVRC':
with_image = False
else:
with_image = True
img_train = CUBClsDataset(root=args.data,
datalist=args.train_list,
transform=tsfm_train,
with_image=with_image)
img_test = CUBCamDataset(root=args.data,
datalist=args.test_list,
transform=tsfm_test,
with_image=with_image)
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
img_train)
else:
train_sampler = None
train_loader = DataLoader(img_train,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
sampler=train_sampler,
num_workers=args.workers)
val_loader = DataLoader(img_test,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers)
return train_loader, val_loader, train_sampler
def main_worker(gpu, ngpus_per_node, args):
writer = SummaryWriter(f'{args.out_dir}/gpu-{gpu}')
global best_acc1
args.gpu = gpu
params = {'num_classes':args.num_classes}
if args.pretrained:
params.update(pretrained=args.pretrained)
if args.finetune:
params.update(finetune=args.finetune)
if args.advprop:
params.update(advprop=args.advprop)
if args.drop_fc:
params.update(load_fc=not args.drop_fc)
grayscale = 'gray' in args.arch
model_name = args.arch.split('-gray', 1)[0]
os.makedirs(args.out_dir, exist_ok=True)
checkpoint_name = f"{args.out_dir}/{args.arch}.checkpoint.pth.tar"
if args.gpu is not None:
print("Use GPU: {} for training".format(args.gpu))
if args.distributed:
if args.dist_url == "env://" and args.rank == -1:
args.rank = int(os.environ["RANK"])
if args.multiprocessing_distributed:
# For multiprocessing distributed training, rank needs to be the
# global rank among all the processes
args.rank = args.rank * ngpus_per_node + gpu
dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
world_size=args.world_size, rank=args.rank)
# create model
if 'efficientnet' in model_name: # NEW
if args.pretrained:
model = EfficientNet.from_pretrained(model_name, **params)
print("=> using pre-trained model '{}'".format(model_name))
else:
print("=> creating model '{}'".format(model_name))
model = EfficientNet.from_name(model_name, override_params=params)
else:
if args.pretrained == 'imagenet':
print("=> using pre-trained model '{}'".format(model_name))
model = models.__dict__[model_name](pretrained=True, **params)
else:
print("=> creating model '{}'".format(model_name))
model = models.__dict__[model_name](**params)
if args.finetune:
print(f'freezing layers')
for layer, param in model.named_parameters():
if 'fc' in layer :
param.requires_grad = True
else:
param.requires_grad = False
if args.distributed:
# For multiprocessing distributed, DistributedDataParallel constructor
# should always set the single device scope, otherwise,
# DistributedDataParallel will use all available devices.
if args.gpu is not None:
torch.cuda.set_device(args.gpu)
model.cuda(args.gpu)
# When using a single GPU per process and per
# DistributedDataParallel, we need to divide the batch size
# ourselves based on the total number of GPUs we have
args.batch_size = int(args.batch_size / ngpus_per_node)
#args.workers = int(args.workers / ngpus_per_node)
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], find_unused_parameters=True)
else:
model.cuda()
# DistributedDataParallel will divide and allocate batch_size to all
# available GPUs if device_ids are not set
model = torch.nn.parallel.DistributedDataParallel(model)
elif args.gpu is not None:
torch.cuda.set_device(args.gpu)
model = model.cuda(args.gpu)
else:
# DataParallel will divide and allocate batch_size to all available GPUs
if model_name.startswith('alexnet') or model_name.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
# if args.pretrained:
# load_pretrained_weights(model, model_name, args.pretrained, load_fc=not args.drop_fc)
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda(args.gpu)
optimizer = torch.optim.SGD(model.parameters(), args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_acc1 = checkpoint['best_acc1']
if args.gpu is not None:
# best_acc1 may be from a checkpoint from a different GPU
best_acc1 = best_acc1.to(args.gpu)
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})"
.format(args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# Data loading code
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val')
testdir = os.path.join(args.data, 'test')
normalize_rgb = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
normalize_gray = transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5])
normalize = normalize_gray if grayscale else normalize_rgb
convert = transforms.Grayscale(num_output_channels=3) if grayscale else transforms.Lambda(lambda x : x)
if 'efficientnet' in model_name:
image_size = image_crop = EfficientNet.get_image_size(args.arch)
else:
image_size = 256
image_crop = 224
print('Using image size', image_crop)
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
convert,
transforms.Resize([image_size, image_size]),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(),
normalize,
]))
val_transforms = transforms.Compose([
convert,
transforms.Resize([image_size, image_size], interpolation=PIL.Image.BICUBIC),
transforms.ToTensor(),
normalize,
])
val_dataset = datasets.ImageFolder(valdir, val_transforms)#, loader=cache_img)
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
val_sampler = torch.utils.data.distributed.DistributedSampler(val_dataset)
else:
train_sampler = None
val_dataset = None
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None),
num_workers=args.workers, pin_memory=True, sampler=train_sampler)
val_loader = torch.utils.data.DataLoader(
val_dataset, batch_size=args.batch_size, shuffle=(val_sampler is None),
num_workers=args.workers, pin_memory=True)
if args.evaluate:
if args.advprop:
normalize = transforms.Lambda(adv_prop_transform)
test_transforms = transforms.Compose([
convert,
transforms.Resize([image_size, image_size], interpolation=PIL.Image.BICUBIC),
transforms.ToTensor(),
normalize,
])
test_loader = torch.utils.data.DataLoader(
datasets.ImageFolder(testdir, test_transforms),
batch_size=args.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True)
res = validate(test_loader, model, criterion, args)
with open('res.txt', 'w') as f:
f.write(str(res))
return
with open(f"{args.out_dir}/{args.arch}.params.txt", 'w') as f:
f.write(f"Training {args.arch} with params {args}")
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
adjust_learning_rate(optimizer, epoch, args)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args, writer)
# evaluate on validation set
acc1 = validate(val_loader, model, criterion, args, epoch, writer)
# remember best acc@1 and save checkpoint
is_best = acc1 > best_acc1
best_acc1 = max(acc1, best_acc1)
if not args.multiprocessing_distributed or (args.multiprocessing_distributed
and args.rank % ngpus_per_node == 0):
save_checkpoint({
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'acc1': acc1,
'best_acc1': best_acc1,
'optimizer' : optimizer.state_dict(),
}, is_best, checkpoint_name)
def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch DWT-MEC OfficeHome')
parser.add_argument('--num_workers', default=2, type=int)
parser.add_argument(
'--source_batch_size',
type=int,
default=18,
help='input source batch size for training (default: 20)')
parser.add_argument(
'--target_batch_size',
type=int,
default=18,
help='input target batch size for training (default: 20)')
parser.add_argument('--test_batch_size',
type=int,
default=10,
help='input batch size for testing (default: 10)')
parser.add_argument('--s_dset_path',
type=str,
default='../data/OfficeHomeDataset_10072016/Art',
help="The source dataset path")
parser.add_argument('--t_dset_path',
type=str,
default='../data/OfficeHomeDataset_10072016/Clipart',
help="The target dataset path")
parser.add_argument('--resnet_path',
type=str,
default='../data/models/model_best_gr_4.pth.tar',
help="The pre-trained model path")
parser.add_argument('--img_resize',
type=int,
default=256,
help='size of the input image')
parser.add_argument('--img_crop_size',
type=int,
default=224,
help='size of the cropped image')
parser.add_argument('--num_iters',
type=int,
default=10000,
help='number of iterations to train (default: 10000)')
parser.add_argument(
'--check_acc_step',
type=int,
default=100,
help=
'number of iterations steps to check validation accuracy (default: 10)'
)
parser.add_argument('--lr_change_step', type=int, default=1000)
parser.add_argument('--lr',
type=float,
default=1e-2,
help='learning rate (default: 0.01)')
parser.add_argument('--num_classes',
type=int,
default=65,
help='number of classes in the dataset')
parser.add_argument('--sgd_momentum',
type=float,
default=0.5,
help='SGD momentum (default: 0.5)')
parser.add_argument(
'--running_momentum',
type=float,
default=0.1,
help='Running momentum for domain statistics(default: 0.1)')
parser.add_argument(
'--lambda_mec_loss',
type=float,
default=0.1,
help='Value of lambda for the entropy loss (default: 0.1)')
parser.add_argument(
'--log_interval',
type=int,
default=10,
help='how many batches to wait before logging training status')
parser.add_argument('--seed',
type=int,
default=1,
help='random seed (default: 1)')
args = parser.parse_args()
# set the seed
torch.manual_seed(args.seed)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# transformation on the source data during training and test data during test
data_transform = transforms.Compose([
transforms.Resize(
(args.img_resize, args.img_resize)), # spatial size of vgg-f input
transforms.RandomCrop((args.img_crop_size, args.img_crop_size)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
# transformation on the target data
data_transform_dup = transforms.Compose([
transforms.Resize((args.img_resize, args.img_resize)),
transforms.RandomCrop((args.img_crop_size, args.img_crop_size)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Lambda(lambda x: _random_affine_augmentation(x)),
transforms.Lambda(lambda x: _gaussian_blur(x)),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
# train data sets
source_dataset = folder.ImageFolder(root=args.s_dset_path,
transform=data_transform)
target_dataset = folder.ImageFolder(root=args.t_dset_path,
transform=data_transform,
transform_aug=data_transform_dup)
# test data sets
target_dataset_test = folder.ImageFolder(root=args.t_dset_path,
transform=data_transform)
# '''''''''''' Train loaders ''''''''''''''' #
source_trainloader = torch.utils.data.DataLoader(
source_dataset,
batch_size=args.source_batch_size,
shuffle=True,
num_workers=args.num_workers,
drop_last=True)
target_trainloader = torch.utils.data.DataLoader(
target_dataset,
batch_size=args.source_batch_size,
shuffle=True,
num_workers=args.num_workers,
drop_last=True)
# '''''''''''' Test loader ''''''''''''''' #
target_testloader = torch.utils.data.DataLoader(
target_dataset_test,
batch_size=args.test_batch_size,
shuffle=True,
num_workers=args.num_workers)
model = resnet50(args.resnet_path, device).to(device)
final_layer_params = []
rest_of_the_net_params = []
for name, param in model.named_parameters():
if name.startswith('fc_out'):
final_layer_params.append(param)
else:
rest_of_the_net_params.append(param)
optimizer = optim.SGD([{
'params': rest_of_the_net_params
}, {
'params': final_layer_params,
'lr': args.lr
}],
lr=args.lr * 0.1,
momentum=0.9,
weight_decay=5e-4)
train_infinite_collect_stats(args=args,
model=model,
device=device,
source_train_loader=source_trainloader,
target_train_loader=target_trainloader,
optimizer=optimizer,
lambda_mec_loss=args.lambda_mec_loss,
target_test_loader=target_testloader)
print(args)
assert args.source != args.target, "source and target can not be the same!!!"
# svhn-like transforms
dset_transforms = list()
dset_transforms.append(transforms.ToTensor())
dset_transforms.append(
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]))
dset_transforms = transforms.Compose(dset_transforms)
# mnist-like grayscale dataset transforms
gray_transforms = list()
gray_transforms.append(transforms.Resize((32, 32)))
gray_transforms.append(transforms.ToTensor())
gray_transforms.append(transforms.Lambda(lambda x: x.repeat(3, 1, 1)))
gray_transforms.append(
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]))
gray_transforms = transforms.Compose(gray_transforms)
# mnistm transforms
mnistm_transforms = list()
mnistm_transforms.append(transforms.Resize((32, 32)))
mnistm_transforms.append(transforms.ToTensor())
mnistm_transforms.append(
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]))
mnistm_transforms = transforms.Compose(mnistm_transforms)
# create the dataloaders
dataloader = {}
if args.source == 'svhn':
testset = SVHN(root='./svhn_data',
split='test',
download=True,
transform=transform_test)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=1000,
shuffle=False)
# Cifar-100 dataset
if cifar100:
normalize = transforms.Normalize(
mean=[x / 255.0 for x in [125.3, 123.0, 113.9]],
std=[x / 255.0 for x in [63.0, 62.1, 66.7]])
transform_train = transforms.Compose([
transforms.ToTensor(),
transforms.Lambda(lambda x: F.pad(Variable(
x.unsqueeze(0), requires_grad=False, volatile=True),
(4, 4, 4, 4),
mode='reflect').data.squeeze()),
transforms.ToPILImage(),
transforms.RandomCrop(32),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
# ~ data prep for test set (CIFAR100)
transform_test = transforms.Compose([transforms.ToTensor(), normalize])
# ~ load training and test set here (CIFAR100):
training_set = datasets.CIFAR100(root='./cifar100_data',
train=True,
download=True,
transform=transform_train)
train_loader = torch.utils.data.DataLoader(training_set,
def train(args):
device = torch.device("cuda" if args.cuda else "cpu")
np.random.seed(args.seed)
torch.manual_seed(args.seed)
transform = transforms.Compose([
transforms.Resize(args.image_size),
transforms.CenterCrop(args.image_size),
transforms.ToTensor(),
transforms.Lambda(lambda x: x.mul(255))
])
train_dataset = datasets.ImageFolder(args.dataset, transform)
train_loader = DataLoader(train_dataset, batch_size=args.batch_size)
transformer = TransformerNet().to(device)
optimizer = Adam(transformer.parameters(), args.lr)
mse_loss = torch.nn.MSELoss()
vgg = Vgg16(requires_grad=False).to(device)
style_transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Lambda(lambda x: x.mul(255))])
style = utils.load_image(args.style_image, size=args.style_size)
style = style_transform(style)
style = style.repeat(args.batch_size, 1, 1, 1).to(device)
features_style = vgg(utils.normalize_batch(style))
gram_style = [utils.gram_matrix(y) for y in features_style]
for e in range(args.epochs):
transformer.train()
agg_content_loss = 0.
agg_style_loss = 0.
count = 0
for batch_id, (x, _) in enumerate(train_loader):
n_batch = len(x)
count += n_batch
optimizer.zero_grad()
x = x.to(device)
y = transformer(x)
y = utils.normalize_batch(y)
x = utils.normalize_batch(x)
features_y = vgg(y)
features_x = vgg(x)
content_loss = args.content_weight * mse_loss(
features_y.relu2_2, features_x.relu2_2)
style_loss = 0.
for ft_y, gm_s in zip(features_y, gram_style):
gm_y = utils.gram_matrix(ft_y)
style_loss += mse_loss(gm_y, gm_s[:n_batch, :, :])
style_loss *= args.style_weight
total_loss = content_loss + style_loss
total_loss.backward()
optimizer.step()
agg_content_loss += content_loss.item()
agg_style_loss += style_loss.item()
if (batch_id + 1) % args.log_interval == 0:
mesg = "{}\tEpoch {}:\t[{}/{}]\tcontent: {:.6f}\tstyle: {:.6f}\ttotal: {:.6f}".format(
time.ctime(), e + 1, count, len(train_dataset),
agg_content_loss / (batch_id + 1),
agg_style_loss / (batch_id + 1),
(agg_content_loss + agg_style_loss) / (batch_id + 1))
print(mesg)
if args.checkpoint_model_dir is not None and (
batch_id + 1) % args.checkpoint_interval == 0:
transformer.eval().cpu()
ckpt_model_filename = "ckpt_epoch_" + str(
e) + "_batch_id_" + str(batch_id + 1) + ".pth"
ckpt_model_path = os.path.join(args.checkpoint_model_dir,
ckpt_model_filename)
torch.save(transformer.state_dict(), ckpt_model_path)
transformer.to(device).train()
# save model
transformer.eval().cpu()
save_model_filename = "epoch_" + str(args.epochs) + "_" + str(
time.ctime()).replace(' ', '_') + "_" + str(
args.content_weight) + "_" + str(args.style_weight) + ".model"
save_model_path = os.path.join(args.save_model_dir, save_model_filename)
torch.save(transformer.state_dict(), save_model_path)
print("\nDone, trained model saved at", save_model_path)
optimizer = torch.optim.RMSprop(mdrnn.parameters(), lr=1e-3, alpha=.9)
scheduler = ReduceLROnPlateau(optimizer, 'min', factor=0.5, patience=5)
earlystopping = EarlyStopping('min', patience=30)
if exists(rnn_file) and not args.noreload:
rnn_state = torch.load(rnn_file)
logger.info("Loading MDRNN at epoch {} "
"with test error {}".format(rnn_state["epoch"],
rnn_state["precision"]))
mdrnn.load_state_dict(rnn_state["state_dict"])
optimizer.load_state_dict(rnn_state["optimizer"])
scheduler.load_state_dict(state['scheduler'])
earlystopping.load_state_dict(state['earlystopping'])
# Data Loading
transform = transforms.Lambda(lambda x: np.transpose(x, (0, 3, 1, 2)) / 255)
train_loader = DataLoader(RolloutSequenceDataset('datasets/carracing',
SEQ_LEN,
transform,
buffer_size=30),
batch_size=BSIZE,
num_workers=8,
shuffle=True,
drop_last=True)
test_loader = DataLoader(RolloutSequenceDataset('datasets/carracing',
SEQ_LEN,
transform,
train=False,
buffer_size=10),
batch_size=BSIZE,
num_workers=8,
def __getitem__(self, index):
"""Read an image from a file and preprocesses it and returns."""
image_path = self.image_paths[index % len(self.image_paths)]
image = Image.open(image_path)
if np.array(image).shape == 4:
image = np.array(image)[:, :, :3]
image = Image.fromarray(image)
# target (high-resolution image)
# Random Crop:
#transform = transforms.RandomCrop(self.image_size * self.scale_factor)
#hr_image = transform(image)
# Face Crop:
face_width = face_height = 128
j = (image.size[0] - face_width) // 2
i = (image.size[1] - face_height) // 2
image = image.crop([j, i, j + face_width, i + face_height])
hr_image = image
# HR_image --> [0,1] --> torch
hr_image_scaled = Scaling(hr_image)
hr_image_scaled = torch.from_numpy(hr_image_scaled).float().to(
self.device) # NUMPY to TORCH
# get HR residuals --> [-1,1] --> torch
transform_HR = transforms.Compose([
#random blur
transforms.Lambda(lambda x: self.randkern.RandomBlur(x)),
#downscale BICUBIC pro tamanho LR
transforms.Resize((self.image_size, self.image_size),
Image.BICUBIC),
#upscale BICUBIC pro tamanho HR
transforms.Resize((self.image_size * self.scale_factor,
self.image_size * self.scale_factor),
Image.BICUBIC)
])
hr_image_hat = transform_HR(hr_image)
hr_residual = np.array(hr_image).astype(float) - np.array(
hr_image_hat).astype(float)
hr_residual_scaled = Scaling(hr_residual)
hr_residual_scaled = torch.from_numpy(hr_residual_scaled).float().to(
self.device) # NUMPY to TORCH
# get LR_IMAGE --> [0,1]
transform_to_lr = transforms.Compose([
transforms.Lambda(lambda x: self.randkern.RandomBlur(x)),
transforms.Resize((self.image_size, self.image_size),
Image.BICUBIC)
])
lr_image = transform_to_lr(hr_image)
lr_image_scaled = Scaling(lr_image)
# get LR_RESIDUAL --> [-1,1]
transform_to_vlr = transforms.Compose([
transforms.Lambda(
lambda x: self.randkern.RandomBlur(x)), #random blur
transforms.Lambda(lambda x: random_downscale(x, self.scale_factor)
), #random downscale
transforms.Resize((self.image_size, self.image_size),
Image.BICUBIC) #upscale pro tamanho LR
])
lr_image_hat = transform_to_vlr(lr_image)
lr_residual = np.array(lr_image).astype(
np.float32) - np.array(lr_image_hat).astype(np.float32)
lr_residual_scaled = Scaling(lr_residual)
# LR_image_scaled + kernel_proj + LR_residual_scaled (CONCAT) ---> TO TORCH
lr_image_without_kernel = lr_image_scaled #self.randkern.ConcatDegraInfo(lr_image_scaled)
lr_image_with_resid = np.concatenate(
(lr_image_without_kernel, lr_residual_scaled), axis=-1)
lr_image_with_resid = torch.from_numpy(lr_image_with_resid).float().to(
self.device) # NUMPY to TORCH
# LR_image to torch
lr_image_scaled = torch.from_numpy(lr_image_scaled).float().to(
self.device) # NUMPY to TORCH
#Transpose - Permute since for model we need input with channels first
lr_image_scaled = lr_image_scaled.permute(2, 0, 1)
hr_image_scaled = hr_image_scaled.permute(2, 0, 1)
lr_image_with_resid = lr_image_with_resid.permute(2, 0, 1)
hr_residual_scaled = hr_residual_scaled.permute(2, 0, 1)
return image_path,\
lr_image_scaled,\
hr_image_scaled, \
lr_image_with_resid, \
hr_residual_scaled
def stylize(args):
device = torch.device("cuda" if args.cuda else "cpu")
content_image = utils.load_image(args.content_image,
scale=args.content_scale)
content_transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Lambda(lambda x: x.mul(255))])
content_image = content_transform(content_image)
content_image = content_image.unsqueeze(0).to(device)
cap = cv2.VideoCapture()
#cap = cv2.VideoCapture(0)
#cap.open("rtsp://*****:*****@192.168.80.241/video1+audio1") # 싸움터
#cap.open("rtsp://*****:*****@192.168.190.241/video1+audio1") #
#cap.open("G:\\0000001.mp4")
cap.open("rtsp://172.26.19.202:554/h264")
fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter('style_transfer_001.avi', fourcc, 30.0, (1920, 1080))
if args.model.endswith(".onnx"):
output = stylize_onnx_caffe2(content_image, args)
else:
with torch.no_grad():
style_model = TransformerNet()
state_dict = torch.load(args.model)
# remove saved deprecated running_* keys in InstanceNorm from the checkpoint
for k in list(state_dict.keys()):
if re.search(r'in\d+\.running_(mean|var)$', k):
del state_dict[k]
style_model.load_state_dict(state_dict)
style_model.to(device)
if args.export_onnx:
assert args.export_onnx.endswith(
".onnx"), "Export model file should end with .onnx"
output = torch.onnx._export(style_model, content_image,
args.export_onnx).cpu()
else:
#output = style_model(content_image).cuda()
count = 0
while (cap.isOpened()):
count += 1
try:
if count % 2 == 0:
ret, frame1 = cap.read()
frame1 = cv2.resize(frame1, (640, 480))
frame1 = cv2.cvtColor(frame1, cv2.COLOR_BGR2RGB)
frame = Image.fromarray(frame1)
frame = content_transform(frame)
frame = frame.unsqueeze(0).to(device)
# output = style_model(frame).cuda()
output = style_model(frame).to(device)
img = output[0].clone().cpu().clamp(
0, 255).cpu().numpy()
img = img.transpose(1, 2, 0).astype("uint8")
# print(np.shape(img))
# Multi-Window
img = cv2.resize(img, (640, 480))
numpy_horizontal_concat = np.concatenate(
(frame1, img), axis=1)
out.write(img)
cv2.imshow('frame', numpy_horizontal_concat)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
except Exception as e:
print(str(e))
cap.release()
cv2.destroyAllWindows()
def main(model):
# print the experiment configuration
model.eval()
root_dir = '/home/xuzhang/project/Medifor/data/' + args.test_set + '/'
world_list = image_processing.load_csv(root_dir + 'indexes/' + args.test_set + '-provenancefiltering-world.csv','|')
world_file_list = world_list["WorldFileName"]
world_id_list = world_list["WorldFileID"]
subset = 'world'
kwargs = {'num_workers': args.num_workers, 'pin_memory': args.pin_memory} if args.cuda else {}
if args.cuda:
model.cuda()
transform = transforms.Compose([
transforms.Lambda(centerCrop),
#transforms.Lambda(cv2_scale),
transforms.Lambda(np_reshape),
transforms.ToTensor(),
#transforms.Normalize((args.mean_image,), (args.std_image,))
transforms.Normalize((args.mean_image,args.mean_image,args.mean_image),
(args.std_image,args.std_image,args.std_image))
])
test_batch_size = args.test_batch_size
descriptor_dim = 128
detector_name = 'DOG'
if args.resume:
if os.path.isfile(args.resume):
print('=> loading checkpoint {}'.format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
else:
print('=> no checkpoint found at {}'.format(args.resume))
for idx,(filename,fileid) in tqdm(enumerate(zip(world_file_list,world_id_list))):
meta = np.array([])
features = np.array([])
feature_file = root_dir+'/'+detector_name+'_prov_desc_'+subset+'/'+fileid+'.npz'
try:
gray_image, ratio = image_processing.read_image_from_name(root_dir,filename)
color_image, ratio = image_processing.read_color_image_from_name(root_dir,filename)
except:
print(filename)
np.savez(feature_file, meta, features)
kp_list, sift_desc = image_processing.extract_sift(gray_image)
patches_list = []
for kp_idx, kp in enumerate(kp_list):
tmp_patch = image_processing.extract_patch(color_image, kp)
patches_list.append(tmp_patch)
patches_list = np.array(patches_list)
patches_list = torch.ByteTensor(patches_list)
test_loader = torch.utils.data.DataLoader(
testDataset(train=False,
data = patches_list,
batch_size=args.test_batch_size,
root=args.dataroot,
name=args.test_set,
download=False,
transform=transform),
batch_size=args.test_batch_size,
shuffle=False, **kwargs)
patch_number = len(kp_list)
if patch_number == 0:
np.savez(feature_file, meta=meta, features = features)
continue
offset = 0
meta = kp_list#np.zeros((patch_number,4))
features = []
for data_a in test_loader:
if args.cuda:
data_a = data_a.cuda()
data_a = Variable(data_a, volatile=True)
out_a = model(data_a)
features.append(out_a.data.cpu().numpy())
features = np.vstack(features).reshape(patch_number,descriptor_dim)
try:
os.stat(feature_file)
os.remove(feature_file)
except:
feature_file = feature_file
np.savez(feature_file, meta=meta, features = features)
