def main_worker(args):
global best_acc1
# create model
argss = default_argument_parser().parse_args()
argss.config_file = 'mv_to_new_home/configs/RearrNet_50.yaml'
cfg = setup(argss)
# model = build_gtnet_backbone_pretrain(cfg, 3, 1000)
# model = build_rearrnet_backbone_pretrain(cfg, 3, 100)
# model = build_defenet_backbone_pretrain(cfg, 3, 100)
# model = build_oidnet_backbone_pretrain(cfg, 3, 100)
# model = build_rpnet_backbone_pretrain(cfg, 3, 100)
# model = build_realnet_backbone_pretrain(cfg, 3, 100)
model = build_oinet_backbone_pretrain(cfg, 3, 100)
# model = build_deformnet_backbone_pretrain(cfg, 3, 100)
model = torch.nn.DataParallel(model.cuda())
# args.evaluate = True
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
# 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']
best_acc1 = best_acc1.to()
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))
# Data loading code
data_path = '/ws/data/imagenet'
traindir = os.path.join(data_path, 'train')
valdir = os.path.join(data_path, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
input_size = 128
cifar_data_path = '/ws/data/open_datasets/classification/cifar100'
train_dataset = datasets.CIFAR100(
cifar_data_path,
train=True,
download=True,
transform=transforms.Compose([
transforms.RandomHorizontalFlip(),
# transforms.RandomVerticalFlip(),
transforms.RandomRotation(30),
transforms.Resize((int(input_size * 1.4), int(input_size * 1.4))),
transforms.CenterCrop((input_size, input_size)),
transforms.ToTensor(),
transforms.RandomErasing(),
transforms.Normalize((0.5, ), (0.5, ))
]))
val_dataset = datasets.CIFAR100(
cifar_data_path,
train=False,
download=True,
transform=transforms.Compose([
# transforms.RandomRotation(90),
# transforms.RandomHorizontalFlip(),
transforms.Resize((int(input_size * 1.4), int(input_size * 1.4))),
transforms.CenterCrop((input_size, input_size)),
transforms.Resize((input_size, input_size)),
transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, )),
]))
# train_dataset = datasets.ImageFolder(
# traindir,
# transforms.Compose([
# transforms.RandomResizedCrop(size=299, scale=(0.08, 1), ratio=(0.75, 4/3)),
# transforms.RandomHorizontalFlip(p=0.5),
# transforms.RandomVerticalFlip(p=0.5),
# transforms.ColorJitter(brightness=[0.5, 1.5], contrast=[0.5, 1.5], saturation=[0.5, 1.5], hue=[-0.1, 0.1]),
# transforms.RandomRotation(degrees=(-45, 45)),
# transforms.ToTensor(),
# normalize,
# ]))
# val_dataset = datasets.ImageFolder(valdir, transforms.Compose([
# transforms.Resize(324),
# transforms.CenterCrop(299),
# transforms.ToTensor(),
# normalize,
# ]))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=1)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=1)
if args.evaluate:
validate(val_loader, model, criterion, args)
return
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch, args)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args)
# evaluate on validation set
acc1 = validate(val_loader, model, criterion, args)
# remember best acc@1 and save checkpoint
is_best = acc1 > best_acc1
best_acc1 = max(acc1, best_acc1)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
},
is_best,
filename='/ws/data/deformed/rp_all_ckpt.pt')
import warnings
warnings.filterwarnings('ignore', category=FutureWarning)
# Data science tools
import numpy as np
import pandas as pd
import os
from timeit import default_timer as timer
# Image transformations
image_transforms = {
# Train uses data augmentation
'train':
transforms.Compose([
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=15),
transforms.ColorJitter(),
transforms.RandomHorizontalFlip(),
transforms.CenterCrop(size=224), # Image net standards
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225]) # Imagenet standards
]),
# Validation does not use augmentation
'val':
transforms.Compose([
transforms.Resize(size=256),
transforms.CenterCrop(size=224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
def main():
#I must know that atleast something is running
print("Please wait while I train")
args = parse_args()
#path of data directories
data_dir = 'flowers'
train_dir = data_dir + '/train'
val_dir = data_dir + '/valid'
test_dir = data_dir + '/test'
#transformations to be applied on dataset
train_transforms = transforms.Compose([transforms.RandomRotation(30),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])])
test_transforms = transforms.Compose([transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])])
val_transforms = transforms.Compose([transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])])
# TODO: Load the datasets with ImageFolder
train_datasets = datasets.ImageFolder(train_dir, transform=train_transforms)
test_datasets = datasets.ImageFolder(test_dir, transform=test_transforms)
val_datasets = datasets.ImageFolder(val_dir, transform=val_transforms)
# TODO: Using the image datasets and the trainforms, define the dataloaders
trainloader = torch.utils.data.DataLoader(train_datasets, batch_size = 64, shuffle=True)
valloader = torch.utils.data.DataLoader(val_datasets, batch_size = 64, shuffle=True)
testloader = torch.utils.data.DataLoader(test_datasets, batch_size = 64, shuffle=True)
#print(summary(trainloaders))
#image, label = next(iter(trainloader))
#helper.imshow(image[0,:]);
#defining parameters that will be passed as default to the model under training
model = getattr(models, args.arch)(pretrained=True)
#choose out of two models
if args.arch == 'vgg13':
# TODO: Build and train your network
model = models.vgg13(pretrained=True)
print(model)
for param in model.parameters():
param.requires_grad = False
classifier = nn.Sequential(nn.Linear(25088, 4096),
nn.Dropout(p=0.2),
nn.ReLU(),
nn.Linear(4096, 4096),
nn.ReLU(),
nn.Dropout(p=0.2),
nn.Linear(4096,102),
nn.LogSoftmax(dim=1))
model.classifier= classifier
elif args.arch == 'densenet121':
model = models.densenet121(pretrained=True)
print(model)
for param in model.parameters():
param.requires_grad = False
classifier = nn.Sequential(nn.Linear(1024, 512),
nn.Dropout(p=0.6),
nn.ReLU(),
nn.Linear(512, 256),
nn.ReLU(),
nn.Dropout(p=0.6),
nn.Linear(256,102),
nn.LogSoftmax(dim=1))
model.classifier = classifier
model.classifier = classifier
criterion = nn.NLLLoss()
epochs = int(args.epochs)
learning_rate = float(args.learning_rate)
print_every = int(args.print_every)
optimizer = optim.Adam(model.classifier.parameters(), lr=learning_rate)
train(model, criterion, epochs, optimizer, print_every, trainloader, valloader)
model.class_to_idx = train_datasets.class_to_idx
path = args.save_dir
save_checkpoint(args, model, optimizer, learning_rate, epochs, path)
data_transform = transforms.Compose([
transforms.Resize((args.img_h, args.img_w)),
#transforms.RandomHorizontalFlip(),
transforms.ToTensor(), # range [0, 255] -> [0.0,1.0]
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
data_transform_resize = transforms.Compose([
#transforms.Resize((args.img_bi_h, args.img_bi_w)),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ColorJitter(brightness=0.2,
contrast=0.2,
saturation=0.2,
hue=0.1),
transforms.RandomRotation(10),
#transforms.RandomCrop(size=(384,128)),
my_transforms.RandomCrop(range=(0.70, 0.95)),
transforms.Resize((args.img_bi_h, args.img_bi_w)),
transforms.ToTensor(), # range [0, 255] -> [0.0,1.0]
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
data_transform_resize2 = transforms.Compose([
#transforms.Resize((args.img_tri_h, args.img_tri_w)),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ColorJitter(brightness=0.2,
contrast=0.2,
saturation=0.2,
hue=0.1),
transforms.RandomRotation(10),
def build_transforms(self):
padding = transforms.Compose([
transforms.Pad(self.padding, padding_mode='reflect'),
transforms.RandomRotation((-self.rotation, self.rotation)),
transforms.RandomApply(
[transforms.RandomAffine(0, shear=self.shear)]),
transforms.RandomCrop(self.resize)
])
rescaling = transforms.Compose([
transforms.Resize((self.resize, self.resize)),
transforms.RandomApply(
[transforms.RandomAffine(0, shear=self.shear)]),
transforms.RandomRotation((-self.rotation, self.rotation))
])
crop_rescaling = transforms.Compose([
transforms.RandomCrop(self.resize * 0.8),
transforms.Resize((self.resize, self.resize)),
transforms.RandomRotation((-self.rotation, self.rotation))
])
basic_augmentation_transform = transforms.Compose([
transforms.RandomChoice([padding, rescaling]),
transforms.RandomHorizontalFlip(p=self.flip),
transforms.ToTensor()
])
strong_augmentation_transform = transforms.Compose([
transforms.RandomChoice([padding, rescaling, crop_rescaling]),
transforms.RandomHorizontalFlip(p=self.flip),
transforms.RandomApply([
transforms.ColorJitter(brightness=0.05,
contrast=0.1,
saturation=0.05,
hue=0.1)
]),
transforms.ToTensor(),
])
val_test_transform = transforms.Compose([
transforms.Resize((128, 128)),
transforms.ToTensor(),
])
if self.normalize == 'imagenet':
normalization = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
strong_augmentation_transform = transforms.Compose(
[strong_augmentation_transform, normalization])
basic_augmentation_transform = transforms.Compose(
[basic_augmentation_transform, normalization])
val_test_transform = transforms.Compose(
[val_test_transform, normalization])
elif self.normalize is None:
pass
else:
raise Exception('Currently only support `normalize=\'imagenet\'`!')
if self.strong:
return strong_augmentation_transform, val_test_transform
else:
return basic_augmentation_transform, val_test_transform
def digit_load(args):
train_bs = args.batch_size
if args.dset == 's2m':
train_source = svhn.SVHN('./data/svhn/',
split='train',
download=True,
transform=transforms.Compose([
transforms.Resize(32),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5),
(0.5, 0.5, 0.5))
]))
test_source = svhn.SVHN('./data/svhn/',
split='test',
download=True,
transform=transforms.Compose([
transforms.Resize(32),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5),
(0.5, 0.5, 0.5))
]))
train_target = mnist.MNIST_idx(
'./data/mnist/',
train=True,
download=True,
transform=transforms.Compose([
transforms.Resize(32),
transforms.Lambda(lambda x: x.convert("RGB")),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]))
test_target = mnist.MNIST(
'./data/mnist/',
train=False,
download=True,
transform=transforms.Compose([
transforms.Resize(32),
transforms.Lambda(lambda x: x.convert("RGB")),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]))
elif args.dset == 'u2m':
train_source = usps.USPS('./data/usps/',
train=True,
download=True,
transform=transforms.Compose([
transforms.RandomCrop(28, padding=4),
transforms.RandomRotation(10),
transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, ))
]))
test_source = usps.USPS('./data/usps/',
train=False,
download=True,
transform=transforms.Compose([
transforms.RandomCrop(28, padding=4),
transforms.RandomRotation(10),
transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, ))
]))
train_target = mnist.MNIST_idx('./data/mnist/',
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, ),
(0.5, ))
]))
test_target = mnist.MNIST('./data/mnist/',
train=False,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, ))
]))
elif args.dset == 'm2u':
train_source = mnist.MNIST('./data/mnist/',
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, ))
]))
test_source = mnist.MNIST('./data/mnist/',
train=False,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, ))
]))
train_target = usps.USPS_idx('./data/usps/',
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, ))
]))
test_target = usps.USPS('./data/usps/',
train=False,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, ))
]))
dset_loaders = {}
dset_loaders["source_tr"] = DataLoader(train_source,
batch_size=train_bs,
shuffle=True,
num_workers=args.worker,
drop_last=False)
dset_loaders["source_te"] = DataLoader(test_source,
batch_size=train_bs * 2,
shuffle=True,
num_workers=args.worker,
drop_last=False)
dset_loaders["target"] = DataLoader(train_target,
batch_size=train_bs,
shuffle=True,
num_workers=args.worker,
drop_last=False)
dset_loaders["target_te"] = DataLoader(train_target,
batch_size=train_bs,
shuffle=False,
num_workers=args.worker,
drop_last=False)
dset_loaders["test"] = DataLoader(test_target,
batch_size=train_bs * 2,
shuffle=False,
num_workers=args.worker,
drop_last=False)
return dset_loaders
def get_dataset(args):
""" return given network
"""
if args.dataset == 'mnist':
train_dataset = datasets.MNIST(MNIST_PATH,
download=True,
transform=transforms.Compose([
transforms.Resize((32, 32)),
transforms.ToTensor(),
]))
val_dataset = datasets.MNIST(MNIST_PATH,
train=False,
download=True,
transform=transforms.Compose([
transforms.Resize((32, 32)),
transforms.ToTensor(),
]))
elif args.dataset == 'cifar10':
train_dataset = datasets.CIFAR10(
CIFAR10_PATH,
download=True,
transform=transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(15),
transforms.ToTensor(),
transforms.Normalize(mean=[0.49139968, 0.48215827, 0.44653124],
std=[0.24703233, 0.24348505, 0.26158768]),
]))
val_dataset = datasets.CIFAR10(
CIFAR10_PATH,
train=False,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.49139968, 0.48215827, 0.44653124],
std=[0.24703233, 0.24348505, 0.26158768]),
]))
elif args.dataset == 'cifar100':
train_dataset = datasets.CIFAR100(
CIFAR100_PATH,
download=True,
transform=transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(15),
transforms.ToTensor(),
transforms.Normalize(mean=[
0.5070751592371323, 0.48654887331495095, 0.4409178433670343
],
std=[
0.2673342858792401,
0.2564384629170883,
0.27615047132568404
]),
]))
val_dataset = datasets.CIFAR100(
CIFAR100_PATH,
train=False,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[
0.5070751592371323, 0.48654887331495095, 0.4409178433670343
],
std=[
0.2673342858792401,
0.2564384629170883,
0.27615047132568404
]),
]))
elif args.dataset == 'imagenet':
input_image_size = 224
scale = 256 / 224
train_dataset_path = os.path.join(ILSVRC2012_path, 'train')
val_dataset_path = os.path.join(ILSVRC2012_path, 'val')
train_dataset = datasets.ImageFolder(
train_dataset_path,
transforms.Compose([
transforms.RandomResizedCrop(input_image_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
]))
val_dataset = datasets.ImageFolder(
val_dataset_path,
transforms.Compose([
transforms.Resize(int(input_image_size * scale)),
transforms.CenterCrop(input_image_size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
]))
elif args.dataset == 'imagenet_32_noise':
input_image_size = 224
scale = 256 / 224
train_dataset_path = os.path.join(IMAGENET_32_NOISE_PATH, 'train')
val_dataset_path = os.path.join(IMAGENET_32_NOISE_PATH, 'val')
train_dataset = datasets.ImageFolder(
train_dataset_path,
transforms.Compose([
transforms.RandomResizedCrop(input_image_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
]))
val_dataset = datasets.ImageFolder(
val_dataset_path,
transforms.Compose([
transforms.Resize(int(input_image_size * scale)),
transforms.CenterCrop(input_image_size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
]))
elif args.dataset == 'imagenet_32_noise_2':
input_image_size = 224
scale = 256 / 224
train_dataset_path = os.path.join(IMAGENET_32_NOISE_PATH_2, 'train')
val_dataset_path = os.path.join(IMAGENET_32_NOISE_PATH_2, 'val')
train_dataset = datasets.ImageFolder(
train_dataset_path,
transforms.Compose([
transforms.RandomResizedCrop(input_image_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
]))
val_dataset = datasets.ImageFolder(
val_dataset_path,
transforms.Compose([
transforms.Resize(int(input_image_size * scale)),
transforms.CenterCrop(input_image_size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
]))
else:
print('the dataset name you have entered is not supported yet')
sys.exit()
return train_dataset, val_dataset
import torch
from torchvision import datasets, transforms
import PIL
data_root = './data/'
train_root = data_root + 'train'
val_root = data_root + 'val'
test_root = data_root + 'test'
base_transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize([0.5] * 3, [0.5] * 3)])
aug_transform = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(20, resample=PIL.Image.BILINEAR),
transforms.ColorJitter(hue=.05, saturation=.05),
transforms.ToTensor(),
transforms.Normalize([0.5] * 3, [0.5] * 3)
])
train_dataset = datasets.ImageFolder(root=train_root, transform=aug_transform)
val_dataset = datasets.ImageFolder(root=val_root, transform=base_transform)
test_dataset = datasets.ImageFolder(root=test_root, transform=base_transform)
def get_data_loaders(batch_size):
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
def main():
args = parse_args()
data_dir = args.data_dir
train_dir = data_dir + '/train'
valid_dir = data_dir + '/valid'
test_dir = data_dir + '/test'
training_transforms = transforms.Compose([
transforms.RandomRotation(30),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
validataion_transforms = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
testing_transforms = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# TODO: Load the datasets with ImageFolder
image_datasets = [
datasets.ImageFolder(train_dir, transform=training_transforms),
datasets.ImageFolder(valid_dir, transform=validataion_transforms),
datasets.ImageFolder(test_dir, transform=testing_transforms)
]
# TODO: Using the image datasets and the trainforms, define the dataloaders
dataloaders = [
torch.utils.data.DataLoader(image_datasets[0],
batch_size=64,
shuffle=True),
torch.utils.data.DataLoader(image_datasets[1],
batch_size=64,
shuffle=True),
torch.utils.data.DataLoader(image_datasets[2],
batch_size=64,
shuffle=True)
]
model = getattr(models, args.arch)(pretrained=True)
hidden_units = int(args.hidden_units) #added before or it won't work
for param in model.parameters():
param.requires_grad = False
if args.arch == "vgg13":
feature_num = model.classifier[0].in_features
classifier = nn.Sequential(
OrderedDict([('fc1', nn.Linear(feature_num, hidden_units)),
('drop', nn.Dropout(p=0.5)), ('relu', nn.ReLU()),
('fc2', nn.Linear(hidden_units, 102)),
('output', nn.LogSoftmax(dim=1))]))
elif args.arch == "densenet121":
classifier = nn.Sequential(
OrderedDict([('fc1', nn.Linear(1024, hidden_units)),
('drop', nn.Dropout(p=0.6)), ('relu', nn.ReLU()),
('fc2', nn.Linear(hidden_units, 102)),
('output', nn.LogSoftmax(dim=1))]))
model.classifier = classifier
criterion = nn.NLLLoss()
optimizer = optim.Adam(model.classifier.parameters(),
lr=float(args.learning_rate))
epochs = int(args.epochs)
#hidden_units = int(args.hidden_units) #added for hidden units user def
class_index = image_datasets[0].class_to_idx
gpu = args.gpu # get gpu set
train(model, criterion, optimizer, dataloaders, epochs, gpu)
model.class_to_idx = class_index
path = args.save_dir # new save location
save_checkpoint(path, model, optimizer, args, classifier)
def main():
parser = train_args.get_args()
parser.add_argument('--version',
action='version',
version='%(prog)s ' + __version__ + ' by ' + __author__)
cli_args = parser.parse_args()
# directory
#First check
if not os.path.isdir(cli_args.data_directory):
print(f'Data directory {cli_args.data_directory} not found.')
exit(1)
# Then save directory
if not os.path.isdir(cli_args.save_dir):
print(f'Directory {cli_args.save_dir} does not exist. Creating...')
os.makedirs(cli_args.save_dir)
with open(cli_args.categories_json, 'r') as f:
cat_to_name = json.load(f)
output_size = len(cat_to_name)
expected_means = [0.485, 0.456, 0.406]
expected_std = [0.229, 0.224, 0.225]
max_image_size = 224
batch_size = 32
#train_transform
tr_transform = transforms.Compose([transforms.RandomHorizontalFlip(p=0.25),
transforms.RandomRotation(25),
transforms.RandomGrayscale(p=0.02),
transforms.RandomResizedCrop(max_image_size),
transforms.ToTensor(),
transforms.Normalize(expected_means, expected_std)])
#train_dataset
tr_dataset = datasets.ImageFolder(cli_args.data_directory, transform=tr_transform)
#tr_dataloader
tr_dataloader = torch.utils.data.DataLoader(tr_dataset, batch_size=batch_size, shuffle=True)
# model
if not cli_args.arch.startswith("vgg") and not cli_args.arch.startswith("densenet"):
print("Only supporting VGG and DenseNet")
exit(1)
print(f"Using a pre-trained {cli_args.arch} network.")
my_model = models.__dict__[cli_args.arch](pretrained=True)
densenet_input = {
'densenet121': 1024,
'densenet169': 1664,
'densenet161': 2208,
'densenet201': 1920
}
input_size = 0
if cli_args.arch.startswith("vgg"):
input_size = my_model.classifier[0].in_features
if cli_args.arch.startswith("densenet"):
input_size = densenet_input[cli_args.arch]
for param in my_model.parameters():
param.requires_grad = False
od = OrderedDict()
hidden_sizes = cli_args.hidden_units
hidden_sizes.insert(0, input_size)
print(f"Building a {len(cli_args.hidden_units)} hidden layer classifier with inputs {cli_args.hidden_units}")
for i in range(len(hidden_sizes) - 1):
od['fc' + str(i + 1)] = nn.Linear(hidden_sizes[i], hidden_sizes[i + 1])
od['relu' + str(i + 1)] = nn.ReLU()
od['dropout' + str(i + 1)] = nn.Dropout(p=0.15)
od['output'] = nn.Linear(hidden_sizes[i + 1], output_size)
od['softmax'] = nn.LogSoftmax(dim=1)
classifier = nn.Sequential(od)
# Replace the classifier
my_model.classifier = classifier
my_model.zero_grad()
def create_cifar_experiment(num_targets: int, num_reps: int, target_dir: str, sleep: float = 0.0):
# Converting data to torch.FloatTensor
transform = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomRotation(degrees=45),
transforms.RandomHorizontalFlip(),
transforms.ToTensor()
])
# Download the training and test datasets
train_data = torchvision.datasets.CIFAR10(root='data', train=True, download=True, transform=transform)
val_data = torchvision.datasets.CIFAR10(root='data', train=False, download=True, transform=transform)
# Prepare data loaders
train_loader = torch.utils.data.DataLoader(train_data, batch_size=32, num_workers=0, shuffle=True)
val_loader = torch.utils.data.DataLoader(val_data, batch_size=32, num_workers=0, shuffle=True)
parameter_dict = {
"lr": [0.0005, 0.001, 0.005, 0.01],
"num_filters": [4, 6, 8, 10, 12]
}
grid = ParameterGrid(parameter_dict)
grid = list(grid)[:num_targets]
grid = grid[:num_targets]
iterations = 1
baseline_iterations = [1, 3, 8]
burn_in_phase_length = 3
m_max = 10000
strategies = []
j = 0
for it in baseline_iterations:
algorithms = [
ConvolutionalAEAlg(num_channels=3, num_filters=params["num_filters"], learning_rate=params["lr"])
for params in grid
]
strategies.append(Baseline("Baseline (round robin, m={})".format(it),
algorithms=algorithms,
iterations=it,
burn_in_phase_length=burn_in_phase_length,
sleep=0.0))
j += 1
algorithms = [
ConvolutionalAEAlg(num_channels=3, num_filters=params["num_filters"], learning_rate=params["lr"])
for params in grid
]
strategies.append(AnygradSelectAll("Anygrad (no target selection)",
algorithms=algorithms,
iterations=iterations,
burn_in_phase_length=burn_in_phase_length,
sleep=0.0))
j += 1
algorithms = [
ConvolutionalAEAlg(num_channels=3, num_filters=params["num_filters"], learning_rate=params["lr"])
for params in grid
]
strategies.append(AnygradOnlySelection("Anygrad (m={})".format(150),
algorithms=algorithms,
iterations=3,
burn_in_phase_length=burn_in_phase_length,
sleep=0.0))
j += 1
algorithms = [
ConvolutionalAEAlg(num_channels=3, num_filters=params["num_filters"], learning_rate=params["lr"])
for params in grid
]
strategies.append(Anygrad("Anygrad (full)", algorithms=algorithms,
iterations=iterations,
burn_in_phase_length=burn_in_phase_length,
sleep=0.0))
return Experiment(name="Convolutional on Cifar", strategies=strategies,
train_data=[train_loader], val_data=[val_loader],
targets=[i for i in range(num_targets)],
num_reps=num_reps, parallel=False,
target_dir=target_dir, m_max=m_max)
def log(self, strdata):
outstr = strdata + '\n'
outstr = outstr.encode("utf-8")
self.file.write(outstr)
def __del__(self):
self.file.close()
jigsaw_image_transform = t.Compose([
# t.Resize(96, Image.BILINEAR),
# t.CenterCrop(90),
t.Resize(600, Image.BILINEAR),
t.RandomHorizontalFlip(0.5),
t.RandomRotation([0, 360]),
t.RandomCrop(300),
t.ColorJitter(hue=0.01, saturation=0.01, brightness=0.01, contrast=0.01),
t.ToTensor(),
])
rotation_image_transform = t.Compose([
t.Resize(150, Image.BILINEAR),
t.RandomCrop(96),
t.ColorJitter(hue=0.01, saturation=0.01, brightness=0.01, contrast=0.01),
t.ToTensor(),
])
tile_transform = t.Compose([
t.Resize((100, 100)),
t.RandomCrop(96),
# Just normalization for validation
data_transforms = {
# 'train': transforms.Compose([
# transforms.RandomResizedCrop(size=input_size, scale=(0.8, 1.0)),
# transforms.RandomRotation(degrees=15),
# transforms.ColorJitter(),
# transforms.RandomHorizontalFlip(),
# transforms.CenterCrop(size=input_size), # Image net standards
# transforms.ToTensor(),
# transforms.Normalize([0.485, 0.456, 0.406],
# [0.229, 0.224, 0.225]) # Imagenet standards
# ]),
'train':
transforms.Compose([
transforms.Resize(input_size + 20),
transforms.RandomRotation(15, expand=True),
transforms.RandomCrop(input_size),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=0.4,
contrast=0.4,
saturation=0.4,
hue=0.2),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'val':
transforms.Compose([
transforms.Resize(input_size),
transforms.CenterCrop(input_size),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
def get_transforms():
# Keep the same
t0 = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=(0.4054, 0.3780, 0.3547),
std=(0.2221, 0.2151, 0.2112))
])
# Scale brightness between the range (1.5,3.5)
t1 = transforms.Compose([
transforms.ColorJitter(brightness=2.5),
transforms.ToTensor(),
transforms.Normalize(mean=(0.4054, 0.3780, 0.3547),
std=(0.2221, 0.2151, 0.2112))
])
# Scale saturation between (1,2)
t2 = transforms.Compose([
transforms.ColorJitter(saturation=2),
transforms.ToTensor(),
transforms.Normalize(mean=(0.4054, 0.3780, 0.3547),
std=(0.2221, 0.2151, 0.2112))
])
# Scale contrast between (1,1.5)
t3 = transforms.Compose([
transforms.ColorJitter(contrast=1.5),
transforms.ToTensor(),
transforms.Normalize(mean=(0.4054, 0.3780, 0.3547),
std=(0.2221, 0.2151, 0.2112))
])
# Scale hue
t4 = transforms.Compose([
transforms.ColorJitter(hue=0.2),
transforms.ToTensor(),
transforms.Normalize(mean=(0.4054, 0.3780, 0.3547),
std=(0.2221, 0.2151, 0.2112))
])
# Random horizontal flips
t5 = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=(0.4054, 0.3780, 0.3547),
std=(0.2221, 0.2151, 0.2112))
])
# Random shearing
t6 = transforms.Compose([
transforms.RandomAffine(degrees=20, shear=3),
transforms.ToTensor(),
transforms.Normalize(mean=(0.4054, 0.3780, 0.3547),
std=(0.2221, 0.2151, 0.2112))
])
# Random Translation
t7 = transforms.Compose([
transforms.RandomAffine(degrees=10, translate=(0.2, 0.2)),
transforms.ToTensor(),
transforms.Normalize(mean=(0.4054, 0.3780, 0.3547),
std=(0.2221, 0.2151, 0.2112))
])
# Random perspective change
t8 = transforms.Compose([
transforms.RandomPerspective(),
transforms.ToTensor(),
transforms.Normalize(mean=(0.4054, 0.3780, 0.3547),
std=(0.2221, 0.2151, 0.2112))
])
# Random rotation
t9 = transforms.Compose([
transforms.RandomRotation(20),
transforms.ToTensor(),
transforms.Normalize(mean=(0.4054, 0.3780, 0.3547),
std=(0.2221, 0.2151, 0.2112))
])
return t0, t1, t2, t3, t4, t5, t6, t7, t8, t9
ha="center",
va="center",
color="white" if cm[i, j] > thresh else "black")
fig.tight_layout()
return ax
# # DataLoader
imgSize = 512
rotateAngle = 15
preprocess = transforms.Compose([
transforms.Scale(imgSize),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.RandomRotation(rotateAngle),
transforms.ToTensor(),
transforms.Normalize((0.3749, 0.2601, 0.1856), (0.2526, 0.1780, 0.1291)),
])
def getData(mode):
if mode == 'train':
img = pd.read_csv('data\\csv\\train_img.csv')
label = pd.read_csv('data\\csv\\train_label.csv')
return np.squeeze(img.values), np.squeeze(label.values)
elif mode == 'test':
img = pd.read_csv('data\\csv\\test_img.csv')
label = pd.read_csv('data\\csv\\test_label.csv')
return np.squeeze(img.values), np.squeeze(label.values)
# 中心裁剪
for name in os.listdir(path):
img = Image.open(os.path.join(path, name))
size_scale = (int(img.size[1] * 1.2), int(img.size[0] * 1.2))
img_resize = tfs.Resize(size_scale, interpolation=2)(img)
img_crop = tfs.RandomCrop((img.size[1], img.size[0]),
padding=0,
pad_if_needed=False)(img_resize)
img_crop.save(os.path.join(path, 'C_' + name))
# 旋转
for name in os.listdir(path):
img = Image.open(os.path.join(path, name))
img_rot_1 = tfs.RandomRotation(30,
resample=False,
expand=False,
center=None)(img)
img_rot_2 = tfs.RandomRotation(30,
resample=False,
expand=False,
center=None)(img)
img_rot_1.save(os.path.join(path, 'R0_' + name))
img_rot_2.save(os.path.join(path, 'R1_' + name))
# 亮度
for name in os.listdir(path):
img = Image.open(os.path.join(path, name))
img_clj_1 = tfs.ColorJitter(brightness=0.8,
contrast=0,
saturation=0,
hue=0)(img)
def initialize_dataloaders(data_dir):
""" Initializes the dataloaders for train, valid, and test image sets
Parameters:
data_dir -- root directory with train, valid, and test subdirectories
Returns: -- data_loaders, image_datasets
"""
data_dirs = {
'train': data_dir + '/train',
'valid': data_dir + '/valid',
'test': data_dir + '/test'
}
# Special transforms for each set
data_transforms = {
'train':
transforms.Compose([
transforms.RandomRotation(30),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'valid':
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'test':
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
}
# Load the datasets
image_datasets = {
'train':
datasets.ImageFolder(data_dirs['train'],
transform=data_transforms['train']),
'valid':
datasets.ImageFolder(data_dirs['valid'],
transform=data_transforms['valid']),
'test':
datasets.ImageFolder(data_dirs['test'],
transform=data_transforms['test'])
}
# Initialize the dataloaders
data_loaders = {
'train':
torch.utils.data.DataLoader(image_datasets['train'],
batch_size=64,
shuffle=True),
'valid':
torch.utils.data.DataLoader(image_datasets['valid'], batch_size=32),
'test':
torch.utils.data.DataLoader(image_datasets['test'], batch_size=32)
}
return data_loaders, image_datasets
if self.transform:
image = self.transform(image)
return image, label, p_path
def listToJson(data, json_save):
jsonData = json.dumps(data)
fileObject = open(json_save, 'w')
fileObject.write(jsonData)
fileObject.close()
train_transform = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomRotation((-120, 120)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
train_df, val_df = train_test_split(train_csv,
test_size=0.3,
random_state=2018,
stratify=train_csv.SCORE)
#train_df = train_csv
#val_df = val_csv
# Random sampling
#train_labels=train_df.values[:,1]
#sampler_count=[len(np.where(train_labels==i)[0]) for i in range(num_classes)]
#weight = np.array(1./np.array(sampler_count))
#weights = [weight[train_label[1]] for train_label in train_df.values ]
def main():
### define transformations for the data
train_transforms = transforms.Compose([
transforms.RandomRotation(60),
transforms.Resize(255),
transforms.CenterCrop(224),
transforms.RandomHorizontalFlip(30),
transforms.ColorJitter(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
valid_transforms = transforms.Compose([
transforms.Resize(255),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
test_transforms = transforms.Compose([
transforms.Resize(255),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
### define paths to the train, validation, and test data sets
train_dir = data_dir + '/train'
valid_dir = data_dir + '/valid'
test_dir = data_dir + '/test'
### load in the datasets
train_data = datasets.ImageFolder(train_dir, transform=train_transforms)
valid_data = datasets.ImageFolder(valid_dir, transform=valid_transforms)
test_data = datasets.ImageFolder(test_dir, transform=test_transforms)
### set up dataloaders
trainloader = torch.utils.data.DataLoader(train_data,
batch_size=64,
shuffle=True)
validloader = torch.utils.data.DataLoader(valid_data,
batch_size=64,
shuffle=True)
testloader = torch.utils.data.DataLoader(test_data, batch_size=64)
### define processor
device = torch.device(dev)
print("using device '{}'".format(device))
### define model architecture and optimizer
if arch == 'vgg16':
model = models.vgg16(pretrained=True)
#class_in = 25088
else:
model = models.densenet161(pretrained=True)
#class_in = 2208
class_in = model.classifier.in_features
for param in model.parameters():
param.requires_grad = False
model.classifier = nn.Sequential(nn.Linear(class_in, 2000), nn.ReLU(),
nn.Dropout(p=0.2), nn.Linear(2000, 512),
nn.ReLU(), nn.Dropout(p=0.2),
nn.Linear(512, 102), nn.LogSoftmax(dim=1))
criterion = nn.NLLLoss()
optimizer = optim.Adam(model.classifier.parameters(), lr=learnrate)
model = model.to(device)
### train the network
epochs = epoch_count
training_losses = []
validation_losses = []
model.train()
for e in range(epochs):
running_loss = 0
for images, labels in trainloader:
images = images.to(device)
labels = labels.to(device)
#print("image shape: '{}'".format(images.shape))
optimizer.zero_grad()
log_ps = model.forward(images)
loss = criterion(log_ps, labels)
loss.backward()
optimizer.step()
#print("loss: {}".format(loss.item()))
running_loss += loss.item()
else:
valid_loss = 0
accuracy = 0
with torch.no_grad():
model.eval()
for images, labels in validloader:
images, labels = images.to(device), labels.to(device)
logps = model.forward(images)
valid_loss += criterion(logps, labels)
#print("step: {}, valid_loss: {}".format(e, valid_loss))
ps = torch.exp(logps)
top_p, top_class = ps.topk(1, dim=1)
equals = top_class == labels.view(*top_class.shape)
accuracy += torch.mean(equals.type(
torch.FloatTensor)).item()
model.train()
training_losses.append(running_loss / len(trainloader))
validation_losses.append(valid_loss / len(validloader))
print("Epoch: {}/{}.. ".format(e + 1, epochs),
"Training Loss: {:.3f}.. ".format(training_losses[-1]),
"Test Loss: {:.3f}.. ".format(validation_losses[-1]),
"Test Accuracy: {:.3f}".format(accuracy / len(validloader)))
### map from integer values to flower names
with open('cat_to_name.json', 'r') as f:
cat_to_name = json.load(f)
### attach map as a parameter to the model
model.class_to_idx = train_data.class_to_idx
### save model parameters
checkpoint = {
'input size':
25088,
'output size':
102,
'epochs':
epochs,
'model':
model,
'classifier':
nn.Sequential(nn.Linear(class_in, 2000), nn.ReLU(), nn.Dropout(p=0.2),
nn.Linear(2000, 512), nn.ReLU(), nn.Dropout(p=0.2),
nn.Linear(512, 102), nn.LogSoftmax(dim=1)),
#'classifier': model.classifier(),
'optimizer':
optimizer.state_dict(),
'class_to_idx':
model.class_to_idx
}
torch.save(checkpoint, 'checkpoint.pth')
# save the state dict
torch.save(model.state_dict(), 'state_dict.pth')
from torchvision import transforms
#一般随机裁剪 + 旋转用的比较多
#torch无加噪声接口
#Data augmentation helps, but not much(特征都是一类，方差小)
transform = transforms.Compose([
transforms.Resize([32, 32]),
transforms.Scale([32, 32]),
transforms.RandomCrop([28, 28]),
transforms.RandomHorizontalFlip(), #随机水平翻转
transforms.RandomVerticalFlip(), #随机垂直翻转
transforms.RandomRotation(15), #随机旋转(-15度 < x < 15度)
transforms.RandomRotation([0, 90, 180, 270]) #随机旋转
])
if self.transform_target is not None:
target=self.transform_target(target)
target=np.array(target)
target=target.reshape(1,128,128)
return data,data2,target
def __len__(self):
return len(self.data)
transform=transforms.Compose([
# transforms.RandomVerticalFlip(p=1),
transforms.RandomHorizontalFlip(p=1),
# transforms.ToTensor(),
])
transform_target=transforms.Compose([
transforms.RandomRotation(degrees=(90,90)),
transforms.RandomHorizontalFlip(p=1),
# transforms.ToTensor()
])
#使用dataloader处理dataset
train_data=MyDataset(data,data2,target,transform=None,transform_target=None)
valid_data=MyDataset(val_data,val_data2,val_target,transform=None,transform_target=None)
# train_data2=MyDataset(data,target,transform=transform,transform_target=transform_target)
# valid_data2=MyDataset(val_data,val_target,transform=transform,transform_target=transform_target)
BATCH_SIZE=32
train_loader=DataLoader(train_data,BATCH_SIZE,True)
valid_loader=DataLoader(valid_data,BATCH_SIZE,True)
# train_loader=DataLoader(train_data + train_data2,BATCH_SIZE,True)
# valid_loader=DataLoader(valid_data + valid_data2,BATCH_SIZE,True)
import copy
import PIL
from utils import UpperAndLowerCenterCrop, TargetCenterCrop, CircleToRectangle
epochs = 12
batch_size = 16
lr = 0.001
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
data_transforms = {
'train':
transforms.Compose([
# transforms.CenterCrop(1200),
TargetCenterCrop(),
transforms.RandomRotation(180),
transforms.RandomVerticalFlip(p=0.5),
transforms.ColorJitter(0.1, 0.1, 0.1, 0.1),
transforms.Resize(448),
transforms.ToTensor(),
]),
'val':
transforms.Compose([
# transforms.CenterCrop(1200),
TargetCenterCrop(),
transforms.Resize(448),
transforms.ToTensor()
]),
}
data_dir = os.path.join('data', 'upper')
'gpu': args.gpu,
'input_layers': args.input_layers,
'hidden_layers': args.hidden_layers,
'output_layers': args.output_layers,
'drop': args.drop_rate,
'topk':args.topk
}
########## Step 2: Get Data
train_dir = h_params['data_dir'] + '/train'
valid_dir = h_params['data_dir'] + '/valid'
test_dir = h_params['data_dir'] + '/test'
data_transforms = {
'train': transforms.Compose([transforms.RandomRotation(30),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])]),
'test': transforms.Compose([transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])]),
'valid': transforms.Compose([transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])])}
if dataset_name == 'DeepFashion':
path_to_data = 'dataset_files/df_train/V_list_eval_partition.txt'
#############################
#
# Dataloader
#
#############################
# net.preprocess
# {'mean': [0.485, 0.456, 0.406], 'std': [0.229, 0.224, 0.225], 'input_size': 224}
output_size_ = 224
transform_train_1 = transforms.Compose([
transforms.RandomRotation(45),
transforms.Resize(256),
transforms.RandomCrop(output_size_),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
transform_train_2 = transforms.Compose([
transforms.RandomRotation(45),
transforms.Resize(256),
transforms.RandomCrop(output_size_),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
def main(args):
device = torch.device(args.device if torch.cuda.is_available() else "cpu")
torch.cuda.set_device(0)
train_batch_size = args.batch_size
input_transform = Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(10),
#transforms.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5, hue=0),
ToTensor(),
#Normalize([0.35676643, 0.33378336, 0.31191254], [0.24681774, 0.23830362, 0.2326341 ]),
])
target_transform = Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(10),
])
datadir = args.datadir
dataset = carla(datadir,
input_transform=input_transform,
target_transform=target_transform)
val_dataset = carla(datadir,
input_transform=ToTensor(),
target_transform=None)
dataset_len = len(dataset)
dataset_idx = list(range(dataset_len))
#split into training & validation set
train_ratio = 0.8
val_ratio = 1 - train_ratio
split = int(np.floor(train_ratio * dataset_len))
train_idx = np.random.choice(dataset_idx, size=split, replace=False)
val_idx = list(set(dataset_idx) - set(train_idx))
train_sampler = SubsetRandomSampler(train_idx)
val_sampler = SubsetRandomSampler(val_idx)
train_loader = torch.utils.data.DataLoader(dataset,
batch_size=train_batch_size,
sampler=train_sampler)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=1,
sampler=val_sampler)
print('Total images = ', dataset_len)
print('Number of images in train set = ',
train_batch_size * len(train_loader))
print('Number of images in validation set = ', len(val_loader))
net = Net(num_classes=3)
net = net.to(device)
weights = [0.1, 0.5, 2.0]
weights = torch.FloatTensor(weights).to(device)
criterion = nn.CrossEntropyLoss(weight=weights)
optimizer = optim.Adam(net.parameters(), lr=args.lr)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
factor=0.5,
patience=3,
verbose=True)
if (args.loadfile != None):
net.load_state_dict(
torch.load(args.loadfile, map_location={'cuda:1': 'cuda:0'}))
print("Loaded saved model: ", args.loadfile)
train(train_loader, val_loader, optimizer, scheduler, criterion, net, args,
device)
def load_data(where="./flowers"):
'''
Arguments : the datas' path
Returns : The loaders for the train, validation and test datasets
This function receives the location of the image files, applies the necessery transformations (rotations,flips,normalizations and crops) and converts the images to tensor in order to be able to be fed into the neural network
'''
data_dir = where
train_dir = data_dir + '/train'
valid_dir = data_dir + '/valid'
test_dir = data_dir + '/test'
#Apply the required transfomations to the test dataset in order to maximize the efficiency of the learning
#process
train_transforms = transforms.Compose([
transforms.RandomRotation(50),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# Crop and Resize the data and validation images in order to be able to be fed into the network
test_transforms = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
validation_transforms = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# TODO: Load the datasets with ImageFolder
train_data = datasets.ImageFolder(train_dir, transform=train_transforms)
validation_data = datasets.ImageFolder(valid_dir,
transform=validation_transforms)
test_data = datasets.ImageFolder(test_dir, transform=test_transforms)
image_datasets = [train_data, validation_data, test_data]
# TODO: Using the image datasets and the trainforms, define the dataloaders
# The data loaders are going to use to load the data to the NN(no shit Sherlock)
trainloader = torch.utils.data.DataLoader(train_data,
batch_size=64,
shuffle=True)
vloader = torch.utils.data.DataLoader(validation_data,
batch_size=32,
shuffle=True)
testloader = torch.utils.data.DataLoader(test_data,
batch_size=20,
shuffle=True)
dataloaders = [trainloader, vloader, testloader]
return image_datasets, dataloaders
list(classifier.parameters()),
lr=0.0001)
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[35], gamma=0.1)
#herb std-mean
#tensor([0.0808, 0.0895, 0.1141])
#tensor([0.7410, 0.7141, 0.6500])
#photo std-mean
#tensor([0.1399, 0.1464, 0.1392])
#tensor([0.2974, 0.3233, 0.2370])
data_transforms = {
'train':
transforms.Compose([
#transforms.Resize((img_size, img_size)),
transforms.RandomRotation(15),
#transforms.RandomCrop((img_size, img_size)),
#transforms.RandomResizedCrop((img_size, img_size)),
#transforms.CenterCrop((img_size, img_size)),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(),
transformations.TileCircle(),
#transformations.ScaleChange(),
transforms.CenterCrop((img_size, img_size)),
transforms.ToTensor(),
#transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
transforms.Normalize([0.7410, 0.7141, 0.6500],
[0.0808, 0.0895, 0.1141])
#transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
]),
'val':
#transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
])
test_transforms = transforms.Compose([
transforms.Resize((IMG_SIZE, IMG_SIZE)),
#transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
])
# Fine tuning augmentation(comment the above code if using this one)
train_transforms = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(10, resample=Image.BILINEAR),
transforms.RandomAffine(8, translate=(.15, .15)),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
val_transforms = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(10, resample=Image.BILINEAR),
transforms.RandomAffine(8, translate=(.15, .15)),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
test_transforms = transforms.Compose([
transforms.RandomHorizontalFlip(),
valid_dir = os.path.join(data_dir, 'valid/')
test_dir = os.path.join(data_dir, 'test/')
standard_normalization = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
#standard_normalization = transforms.Normalize(mean=[0.5, 0.5, 0.5],
# std=[0.5, 0.5, 0.5])
data_transforms = {
'train':
transforms.Compose([ #transforms.RandomResizedCrop(224),
transforms.Resize(256),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(30),
transforms.ToTensor(), standard_normalization
]),
'val':
transforms.Compose([
transforms.Resize(size=(224, 224)),
transforms.ToTensor(), standard_normalization
]),
'test':
transforms.Compose([
transforms.Resize(size=(224, 224)),
transforms.ToTensor(), standard_normalization
])
}
train_data = datasets.ImageFolder(train_dir,
def construct_transforms(n_in: int,
mode: str,
mean: tuple = (0.0, 0.0, 0.0),
std: tuple = (1.0, 1.0, 1.0),
augment: bool = False,
rotation: bool = False,
num_channels: int = 3,
jitter: float = 0.0):
"""
:param n_in:
:param mode:
:param augment:
:param rotation:
:param jitter:
:return:
"""
assert mode in ['train', 'eval', 'ood']
assert not jitter < 0.0
transf_list = []
# TODO Make automatic. This is here temporaricly...
#mean = (0.4914, 0.4823, 0.4465)
#std = (0.247, 0.243, 0.261)
if augment:
if mode == 'eval':
transf_list.extend([
transforms.Resize(n_in, Image.BICUBIC),
transforms.CenterCrop(n_in)
])
elif mode == 'train':
transf_list.extend([
transforms.Resize(n_in, Image.BICUBIC),
transforms.Pad(4, padding_mode='reflect')
])
if rotation:
transf_list.append(
transforms.RandomRotation(degrees=15,
resample=Image.BICUBIC))
transf_list.extend([
torchvision.transforms.ColorJitter(jitter, jitter, jitter,
jitter),
transforms.RandomHorizontalFlip(),
transforms.RandomCrop(n_in)
])
else:
transf_list.extend([
transforms.Resize(n_in, Image.BICUBIC),
transforms.Pad(4, padding_mode='reflect')
])
if rotation:
transf_list.append(
transforms.RandomRotation(degrees=15,
resample=Image.BICUBIC))
transf_list.extend([
torchvision.transforms.ColorJitter(jitter, jitter, jitter,
jitter),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.RandomCrop(n_in)
])
else:
transf_list.extend([
transforms.Resize(n_in, Image.BICUBIC),
transforms.CenterCrop(n_in)
])
if num_channels < 3:
transf_list.extend([transforms.Grayscale(num_output_channels=3)])
transf_list.extend(
[transforms.ToTensor(),
transforms.Normalize(mean, std)])
return transforms.Compose(transf_list)