def main():
global args, best_prec1
args = parser.parse_args()
args.distributed = args.world_size > 1
if args.distributed:
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size)
# create model
if args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
model = models.__dict__[args.arch](pretrained=True)
else:
print("=> creating model '{}'".format(args.arch))
model = models.__dict__[args.arch](low_dim=args.low_dim)
if not args.distributed:
if args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
else:
model.cuda()
model = torch.nn.parallel.DistributedDataParallel(model)
# 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])
train_dataset = datasets.ImageFolderInstance(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224, scale=(0.2, 1.)),
transforms.RandomGrayscale(p=0.2),
transforms.ColorJitter(0.4, 0.4, 0.4, 0.4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
train_sampler = 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(datasets.ImageFolderInstance(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
# define lemniscate and loss function (criterion)
ndata = train_dataset.__len__()
if args.nce_k > 0:
lemniscate = NCEAverage(args.low_dim, ndata, args.nce_k, args.nce_t,
args.nce_m).cuda()
criterion = NCECriterion(ndata).cuda()
else:
lemniscate = LinearAverage(args.low_dim, ndata, args.nce_t,
args.nce_m).cuda()
criterion = nn.CrossEntropyLoss().cuda()
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_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
lemniscate = checkpoint['lemniscate']
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
if args.evaluate:
kNN(0, model, lemniscate, train_loader, val_loader, 200, args.nce_t)
return
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, lemniscate, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = NN(epoch, model, lemniscate, train_loader, val_loader)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'lemniscate': lemniscate,
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best)
# evaluate KNN after last epoch
kNN(0, model, lemniscate, train_loader, val_loader, 200, args.nce_t)
def main():
# so other functions can access these variables
global args, dataloaders, data_sizes, image_sets
args = get_user_args()
# defining processing device, if cuda is available then GPU else CPU
device = torch.device("cuda:0" if (
torch.cuda.is_available() and args.gpu) else "cpu")
print('=> beginning training using {}'.format(str(device).upper()))
# lets the user know which model is being trained
print('=> creating model: {}'.format(args.arch))
model = models.__dict__[args.arch](pretrained=True)
print('* ' * 20)
model.to(device) # send device to processor
# image location with child folders of train, valid, test
data_dir = Path(args.data)
train_dir = data_dir / 'train'
valid_dir = data_dir / 'valid'
test_dir = data_dir / 'test'
# variable for various iterations later
states = ['train', 'valid', 'test']
# for easy iteration later
dirs_dict = {'train': train_dir, 'valid': valid_dir, 'test': test_dir}
# image normalization parameters, predefined
normalize = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
# transforms for valid and test data, use same parameters
valid_test_transforms = [
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize]
data_transforms = {
'train': # vector manipulation for generalized learning
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomRotation(30),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(),
normalize
]),
'valid':
transforms.Compose(valid_test_transforms),
'test':
transforms.Compose(valid_test_transforms)
}
image_sets = {
i_set: datasets.ImageFolder(
dirs_dict[i_set], transform=data_transforms[i_set])
for i_set in states
}
dataloaders = {
'train': torch.utils.data.DataLoader(
image_sets['train'], batch_size=args.batch_size, shuffle=True),
'valid': torch.utils.data.DataLoader(image_sets['valid'], batch_size=args.batch_size),
'test': torch.utils.data.DataLoader(image_sets['test'], batch_size=args.batch_size)
}
classes = image_sets['train'].classes
data_sizes = {x: len(image_sets[x]) for x in states}
for p in model.parameters():
p.requires_grad = False # ensures gradients aren't calculated for parameters
classifier = nn.Sequential(
OrderedDict([
('fc1', nn.Linear(
model.classifier[0].in_features, args.hidden_units)),
('relu1,', nn.ReLU()),
('dropout', nn.Dropout(args.dropout)),
('fc2', nn.Linear(args.hidden_units, len(classes))),
('output', nn.LogSoftmax(dim=1)),
]))
model.classifier = classifier
criterion = nn.NLLLoss()
optimizer = optim.Adam(model.classifier.parameters(),
lr=args.learning_rate)
scheduler = lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.0125)
model_trained = train(model, optimizer, criterion,
scheduler, args.epochs, device)
save_checkpoint(model_trained, args.epochs, args.save_dir,
args.arch, args.learning_rate, optimizer, args.hidden_units)
# import matplotlib.pyplot as plt
from tqdm import tqdm # Displays a progress bar
import torch
from torch import nn
from torch import optim
import torch.nn.functional as F
from torchvision import datasets, transforms
from torch.utils.data import Dataset, Subset, DataLoader, random_split
# Load the dataset and train, val, test splits
# TODO: currently only experiment with FashionMNIST. Need to process your own dataset
print("Loading datasets...")
DATA_transform= transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.ToTensor(), # Transform from [0,255] uint8 to [0,1] float
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) # Normalize to zero mean and unit variance
])
training_data_dir='./data/Training'
testing_data_dir='./data/Testing'
TUMOR_train=datasets.ImageFolder(training_data_dir, DATA_transform)
TUMOR_test = datasets.ImageFolder(testing_data_dir, DATA_transform)
print(TUMOR_test.classes)
# FASHION_trainval = datasets.FashionMNIST('.', download=True, train=True, transform=FASHION_transform)
# FASHION_train = Subset(FASHION_trainval, range(50000))
# FASHION_val = Subset(FASHION_trainval, range(50000,60000))
# FASHION_test = datasets.FashionMNIST('.', download=True, train=False, transform=FASHION_transform)
print("Done!")
def main_worker(gpu, ngpus_per_node, args):
global best_acc1
args.gpu = gpu
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)
if args.gen_map:
args.qw = -1
args.qa = -1
# create model
if args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
else:
print("=> creating model '{}'".format(args.arch))
try:
model = models.__dict__[args.arch](pretrained=args.pretrained)
args.qw = -1
args.qa = -1
except KeyError:
if 'bp' in args.arch:
model = imagenet_extra_models.__dict__[args.arch](
pretrained=args.pretrained,
nbits_w=args.qw,
log=args.quan_log,
increase_factor=args.increase_factor)
else:
model = imagenet_extra_models.__dict__[args.arch](
pretrained=args.pretrained)
print('model:\n=========\n{}\n=========='.format(model))
if args.gen_map:
try:
ori_model = models.__dict__[args.original_model]()
except KeyError:
ori_model = imagenet_extra_models.__dict__[args.original_model]()
print('original model:\n=========\n{}\n=========='.format(ori_model))
key_map = gen_key_map(model.state_dict(), ori_model.state_dict())
with open('models/imagenet/{}_map.json'.format(args.arch), 'w') as wf:
json.dump(key_map, wf)
print('Generate key map done')
return
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])
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 args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda(args.gpu)
params = add_weight_decay(model,
weight_decay=args.weight_decay,
skip_keys=['expand_'])
optimizer = torch.optim.SGD(params, args.lr, momentum=args.momentum)
# 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, map_location='cpu')
model.load_state_dict(checkpoint['state_dict'])
# if not args.quant_bias_scale:
# 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)
# try:
# model.load_state_dict(checkpoint['state_dict'])
# # ValueError: loaded state dict has a different number of parameter groups
# # different version
# # optimizer.load_state_dict(checkpoint['optimizer'])
# except RuntimeError:
# print('Fine-tune qfi_wide model using qfn_relaxed weights.')
# key_map = gen_key_map(model.state_dict(), checkpoint['state_dict'])
# load_fake_quantized_state_dict(model, checkpoint['state_dict'], key_map)
# args.start_epoch = 0
# best_acc1 = 0
# optimizer = torch.optim.SGD(params, args.lr,
# momentum=args.momentum)
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')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
train_sampler = 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(datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion, args)
return
args.log_name = 'logger/{}_{}'.format(args.arch, args.log_name)
writer = SummaryWriter(args.log_name)
with open('{}/{}.txt'.format(args.log_name, args.arch), 'w') as wf:
wf.write(str(model))
scheduler_cosine = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, T_max=args.epochs)
scheduler_warmup = GradualWarmupScheduler(optimizer,
multiplier=10,
total_epoch=3,
after_scheduler=scheduler_cosine)
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
# adjust_learning_rate(optimizer, epoch, args)
scheduler_warmup.step()
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args, writer)
# evaluate on validation set
acc1, acc5 = validate(val_loader, model, criterion, args)
writer.add_scalar('val/acc1', acc1, epoch)
writer.add_scalar('val/acc5', acc5, epoch)
# writer.add_scalar('val/bs', bs, epoch)
# 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(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
},
is_best,
prefix='{}/{}'.format(args.log_name, args.arch))
if epoch % 10 == 0:
save_checkpoint_backup(model.state_dict(),
prefix='{}/{}_{}'.format(
args.log_name, args.arch, epoch))
writer.close()
def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size',
type=int,
default=64,
metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size',
type=int,
default=1000,
metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--world_size',
type=int,
default=1,
help='number of GPUs to use')
parser.add_argument('--epochs',
type=int,
default=10,
metavar='N',
help='number of epochs to train (default: 10)')
parser.add_argument('--lr',
type=float,
default=0.01,
metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--wd',
type=float,
default=1e-4,
help='weight decay (default: 5e-4)')
parser.add_argument('--lr-decay-every',
type=int,
default=100,
help='learning rate decay by 10 every X epochs')
parser.add_argument('--lr-decay-scalar',
type=float,
default=0.1,
help='--')
parser.add_argument('--momentum',
type=float,
default=0.5,
metavar='M',
help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='disables CUDA training')
parser.add_argument('--seed',
type=int,
default=1,
metavar='S',
help='random seed (default: 1)')
parser.add_argument(
'--log-interval',
type=int,
default=10,
metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--run_test',
default=False,
type=str2bool,
nargs='?',
help='run test only')
parser.add_argument(
'--limit_training_batches',
type=int,
default=-1,
help='how many batches to do per training, -1 means as many as possible'
)
parser.add_argument('--no_grad_clip',
default=False,
type=str2bool,
nargs='?',
help='turn off gradient clipping')
parser.add_argument('--get_flops',
default=False,
type=str2bool,
nargs='?',
help='add hooks to compute flops')
parser.add_argument(
'--get_inference_time',
default=False,
type=str2bool,
nargs='?',
help='runs valid multiple times and reports the result')
parser.add_argument('--mgpu',
default=False,
type=str2bool,
nargs='?',
help='use data paralization via multiple GPUs')
parser.add_argument('--dataset',
default="MNIST",
type=str,
help='dataset for experiment, choice: MNIST, CIFAR10')
parser.add_argument('--data',
metavar='DIR',
default='/imagenet',
help='path to imagenet dataset')
parser.add_argument(
'--model',
default="lenet3",
type=str,
help='model selection, choices: lenet3, vgg, mobilenetv2, resnet18',
choices=[
"lenet3", "vgg", "mobilenetv2", "resnet18", "resnet152",
"resnet50", "resnet50_noskip", "resnet20", "resnet34", "resnet101",
"resnet101_noskip", "densenet201_imagenet", 'densenet121_imagenet',
"multprun_gate5_gpu_0316_1", "mult_prun8_gpu", "multnas5_gpu"
])
parser.add_argument('--tensorboard',
type=str2bool,
nargs='?',
help='Log progress to TensorBoard')
parser.add_argument(
'--save_models',
default=True,
type=str2bool,
nargs='?',
help='if True, models will be saved to the local folder')
parser.add_argument('--fineturn_model',
type=str2bool,
nargs='?',
help='Log progress to TensorBoard')
# ============================PRUNING added
parser.add_argument(
'--pruning_config',
default=None,
type=str,
help=
'path to pruning configuration file, will overwrite all pruning parameters in arguments'
)
parser.add_argument('--group_wd_coeff',
type=float,
default=0.0,
help='group weight decay')
parser.add_argument('--name',
default='test',
type=str,
help='experiment name(folder) to store logs')
parser.add_argument(
'--augment',
default=False,
type=str2bool,
nargs='?',
help=
'enable or not augmentation of training dataset, only for CIFAR, def False'
)
parser.add_argument('--load_model',
default='',
type=str,
help='path to model weights')
parser.add_argument('--pruning',
default=False,
type=str2bool,
nargs='?',
help='enable or not pruning, def False')
parser.add_argument(
'--pruning-threshold',
'--pt',
default=100.0,
type=float,
help=
'Max error perc on validation set while pruning (default: 100.0 means always prune)'
)
parser.add_argument(
'--pruning-momentum',
default=0.0,
type=float,
help=
'Use momentum on criteria between pruning iterations, def 0.0 means no momentum'
)
parser.add_argument('--pruning-step',
default=15,
type=int,
help='How often to check loss and do pruning step')
parser.add_argument('--prune_per_iteration',
default=10,
type=int,
help='How many neurons to remove at each iteration')
parser.add_argument(
'--fixed_layer',
default=-1,
type=int,
help='Prune only a given layer with index, use -1 to prune all')
parser.add_argument('--start_pruning_after_n_iterations',
default=0,
type=int,
help='from which iteration to start pruning')
parser.add_argument('--maximum_pruning_iterations',
default=1e8,
type=int,
help='maximum pruning iterations')
parser.add_argument('--starting_neuron',
default=0,
type=int,
help='starting position for oracle pruning')
parser.add_argument('--prune_neurons_max',
default=-1,
type=int,
help='prune_neurons_max')
parser.add_argument('--pruning-method',
default=0,
type=int,
help='pruning method to be used, see readme.md')
parser.add_argument('--pruning_fixed_criteria',
default=False,
type=str2bool,
nargs='?',
help='enable or not criteria reevaluation, def False')
parser.add_argument('--fixed_network',
default=False,
type=str2bool,
nargs='?',
help='fix network for oracle or criteria computation')
parser.add_argument(
'--zero_lr_for_epochs',
default=-1,
type=int,
help='Learning rate will be set to 0 for given number of updates')
parser.add_argument(
'--dynamic_network',
default=False,
type=str2bool,
nargs='?',
help=
'Creates a new network graph from pruned model, works with ResNet-101 only'
)
parser.add_argument('--use_test_as_train',
default=False,
type=str2bool,
nargs='?',
help='use testing dataset instead of training')
parser.add_argument('--pruning_mask_from',
default='',
type=str,
help='path to mask file precomputed')
parser.add_argument(
'--compute_flops',
default=True,
type=str2bool,
nargs='?',
help=
'if True, will run dummy inference of batch 1 before training to get conv sizes'
)
# ============================END pruning added
best_prec1 = 0
global global_iteration
global group_wd_optimizer
global_iteration = 0
args = parser.parse_args()
use_cuda = not args.no_cuda and torch.cuda.is_available()
torch.manual_seed(args.seed)
args.distributed = args.world_size > 1
if args.distributed:
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=0)
device = torch.device("cuda" if use_cuda else "cpu")
# dataset loading section
if args.dataset == "MNIST":
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(datasets.MNIST(
'../data',
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])),
batch_size=args.batch_size,
shuffle=True,
**kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data',
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])),
batch_size=args.test_batch_size,
shuffle=True,
**kwargs)
elif args.dataset == "CIFAR10":
# Data loading code
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]])
if args.augment:
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
else:
transform_train = transforms.Compose([
transforms.ToTensor(),
normalize,
])
transform_test = transforms.Compose([transforms.ToTensor(), normalize])
kwargs = {'num_workers': 8, 'pin_memory': True}
train_loader = torch.utils.data.DataLoader(datasets.CIFAR10(
'../data', train=True, download=True, transform=transform_train),
batch_size=args.batch_size,
shuffle=True,
drop_last=True,
**kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.CIFAR10('../data', train=False, transform=transform_test),
batch_size=args.test_batch_size,
shuffle=True,
**kwargs)
elif args.dataset == "Imagenet":
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])
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
train_sampler = None
kwargs = {'num_workers': 16}
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
sampler=train_sampler,
pin_memory=True,
**kwargs)
if args.use_test_as_train:
train_loader = torch.utils.data.DataLoader(
datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=(train_sampler is None),
**kwargs)
test_loader = torch.utils.data.DataLoader(datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
pin_memory=True,
**kwargs)
#wm
elif args.dataset == "mult_5T":
args.data_root = ['/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/CX_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TK_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/ZR_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TX_20200616',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/WM_20200709']
args.data_root_val = ['/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/CX_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TK_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/ZR_20200709',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TX_20200616',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/WM_20200709']
args.train_data_list = ['/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/CX_20200709/txt/cx_train.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TK_20200709/txt/tk_train.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/ZR_20200709/txt/zr_train.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TX_20200616/txt/tx_train.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/WM_20200709/txt/wm_train.txt']
args.val_data_list = ['/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/CX_20200709/txt/cx_val.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TK_20200709/txt/tk_val.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/ZR_20200709/txt/zr_val.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/TX_20200616/txt/tx_val.txt',\
'/workspace/mnt/storage/yangdecheng/yangdecheng/data/TR-NMA-07/WM_20200709/txt/wm_val.txt']
num_tasks = len(args.data_root)
args.ngpu = 8
args.workers = 8
args.train_batch_size = [40, 40, 40, 40, 40] #36
args.val_batch_size = [100, 100, 100, 100, 100]
args.loss_weight = [1.0, 1.0, 1.0, 1.0, 1.0]
args.val_num_classes = [[0, 1, 2, 3, 4], [0, 1, 2], [0, 1], [0, 1],
[0, 1, 2, 3, 4, 5, 6]]
args.mixup_alpha = None #None
for i in range(num_tasks):
args.train_batch_size[i] *= args.ngpu
args.val_batch_size[i] *= args.ngpu
pixel_mean = [0.406, 0.456, 0.485]
pixel_std = [0.225, 0.224, 0.229]
#私人定制：
train_dataset = []
for i in range(num_tasks):
if i == 1:
train_dataset.append(
FileListLabeledDataset(
args.train_data_list[i],
args.data_root[i],
Compose([
RandomResizedCrop(
112,
scale=(0.94, 1.),
ratio=(1. / 4., 4. / 1.)
), #scale=(0.7, 1.2), ratio=(1. / 1., 4. / 1.)
RandomHorizontalFlip(),
ColorJitter(brightness=[0.5, 1.5],
contrast=[0.5, 1.5],
saturation=[0.5, 1.5],
hue=0),
ToTensor(),
Lighting(1, [0.2175, 0.0188, 0.0045],
[[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]]),
Normalize(pixel_mean, pixel_std),
])))
else:
train_dataset.append(
FileListLabeledDataset(
args.train_data_list[i], args.data_root[i],
Compose([
RandomResizedCrop(112,
scale=(0.7, 1.2),
ratio=(1. / 1., 4. / 1.)),
RandomHorizontalFlip(),
ColorJitter(brightness=[0.5, 1.5],
contrast=[0.5, 1.5],
saturation=[0.5, 1.5],
hue=0),
ToTensor(),
Lighting(1, [0.2175, 0.0188, 0.0045],
[[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]]),
Normalize(pixel_mean, pixel_std),
])))
#原来的
# train_dataset = [FileListLabeledDataset(
# args.train_data_list[i], args.data_root[i],
# Compose([
# RandomResizedCrop(112,scale=(0.7, 1.2), ratio=(1. / 1., 4. / 1.)),
# RandomHorizontalFlip(),
# ColorJitter(brightness=[0.5,1.5], contrast=[0.5,1.5], saturation=[0.5,1.5], hue= 0),
# ToTensor(),
# Lighting(1, [0.2175, 0.0188, 0.0045], [[-0.5675, 0.7192, 0.4009], [-0.5808, -0.0045, -0.8140], [-0.5836, -0.6948, 0.4203]]),
# Normalize(pixel_mean, pixel_std),]),
# memcached=False,
# memcached_client="") for i in range(num_tasks)]
args.num_classes = [td.num_class for td in train_dataset]
train_longest_size = max([
int(np.ceil(len(td) / float(bs)))
for td, bs in zip(train_dataset, args.train_batch_size)
])
train_sampler = [
GivenSizeSampler(td,
total_size=train_longest_size * bs,
rand_seed=0)
for td, bs in zip(train_dataset, args.train_batch_size)
]
train_loader = [
DataLoader(train_dataset[k],
batch_size=args.train_batch_size[k],
shuffle=False,
num_workers=args.workers,
pin_memory=False,
sampler=train_sampler[k]) for k in range(num_tasks)
]
val_dataset = [
FileListLabeledDataset(
args.val_data_list[i],
args.data_root_val[i],
Compose([
Resize((112, 112)),
# CenterCrop(112),
ToTensor(),
Normalize(pixel_mean, pixel_std),
]),
memcached=False,
memcached_client="") for i in range(num_tasks)
]
val_longest_size = max([
int(np.ceil(len(vd) / float(bs)))
for vd, bs in zip(val_dataset, args.val_batch_size)
])
test_loader = [
DataLoader(val_dataset[k],
batch_size=args.val_batch_size[k],
shuffle=False,
num_workers=args.workers,
pin_memory=False) for k in range(num_tasks)
]
if args.model == "lenet3":
model = LeNet(dataset=args.dataset)
elif args.model == "vgg":
model = vgg11_bn(pretrained=True)
elif args.model == "resnet18":
model = PreActResNet18()
elif (args.model == "resnet50") or (args.model == "resnet50_noskip"):
if args.dataset == "CIFAR10":
model = PreActResNet50(dataset=args.dataset)
else:
from models.resnet import resnet50
skip_gate = True
if "noskip" in args.model:
skip_gate = False
if args.pruning_method not in [22, 40]:
skip_gate = False
model = resnet50(skip_gate=skip_gate)
elif args.model == "resnet34":
if not (args.dataset == "CIFAR10"):
from models.resnet import resnet34
model = resnet34()
elif args.model == "multprun_gate5_gpu_0316_1":
from models.multitask import MultiTaskWithLoss
model = MultiTaskWithLoss(backbone=args.model,
num_classes=args.num_classes,
feature_dim=2560,
spatial_size=112,
arc_fc=False,
feat_bn=False)
print(model)
elif args.model == "mult_prun8_gpu":
from models.multitask import MultiTaskWithLoss
model = MultiTaskWithLoss(backbone=args.model,
num_classes=args.num_classes,
feature_dim=18,
spatial_size=112,
arc_fc=False,
feat_bn=False)
print(model)
elif args.model == "multnas5_gpu": #作为修改项
from models.multitask import MultiTaskWithLoss
model = MultiTaskWithLoss(backbone=args.model,
num_classes=args.num_classes,
feature_dim=512,
spatial_size=112,
arc_fc=False,
feat_bn=False)
print(model)
elif "resnet101" in args.model:
if not (args.dataset == "CIFAR10"):
from models.resnet import resnet101
if args.dataset == "Imagenet":
classes = 1000
if "noskip" in args.model:
model = resnet101(num_classes=classes, skip_gate=False)
else:
model = resnet101(num_classes=classes)
elif args.model == "resnet20":
if args.dataset == "CIFAR10":
NotImplementedError(
"resnet20 is not implemented in the current project")
# from models.resnet_cifar import resnet20
# model = resnet20()
elif args.model == "resnet152":
model = PreActResNet152()
elif args.model == "densenet201_imagenet":
from models.densenet_imagenet import DenseNet201
model = DenseNet201(gate_types=['output_bn'], pretrained=True)
elif args.model == "densenet121_imagenet":
from models.densenet_imagenet import DenseNet121
model = DenseNet121(gate_types=['output_bn'], pretrained=True)
else:
print(args.model, "model is not supported")
####end dataset preparation
if args.dynamic_network:
# attempts to load pruned model and modify it be removing pruned channels
# works for resnet101 only
if (len(args.load_model) > 0) and (args.dynamic_network):
if os.path.isfile(args.load_model):
load_model_pytorch(model, args.load_model, args.model)
else:
print("=> no checkpoint found at '{}'".format(args.load_model))
exit()
dynamic_network_change_local(model)
# save the model
log_save_folder = "%s" % args.name
if not os.path.exists(log_save_folder):
os.makedirs(log_save_folder)
if not os.path.exists("%s/models" % (log_save_folder)):
os.makedirs("%s/models" % (log_save_folder))
model_save_path = "%s/models/pruned.weights" % (log_save_folder)
model_state_dict = model.state_dict()
if args.save_models:
save_checkpoint({'state_dict': model_state_dict},
False,
filename=model_save_path)
print("model is defined")
# aux function to get size of feature maps
# First it adds hooks for each conv layer
# Then runs inference with 1 image
output_sizes = get_conv_sizes(args, model)
if use_cuda and not args.mgpu:
model = model.to(device)
elif args.distributed:
model.cuda()
print(
"\n\n WARNING: distributed pruning was not verified and might not work correctly"
)
model = torch.nn.parallel.DistributedDataParallel(model)
elif args.mgpu:
model = torch.nn.DataParallel(model).cuda()
else:
model = model.to(device)
print(
"model is set to device: use_cuda {}, args.mgpu {}, agrs.distributed {}"
.format(use_cuda, args.mgpu, args.distributed))
weight_decay = args.wd
if args.fixed_network:
weight_decay = 0.0
# remove updates from gate layers, because we want them to be 0 or 1 constantly
if 1:
parameters_for_update = []
parameters_for_update_named = []
for name, m in model.named_parameters():
if "gate" not in name:
parameters_for_update.append(m)
parameters_for_update_named.append((name, m))
else:
print("skipping parameter", name, "shape:", m.shape)
total_size_params = sum(
[np.prod(par.shape) for par in parameters_for_update])
print("Total number of parameters, w/o usage of bn consts: ",
total_size_params)
optimizer = optim.SGD(parameters_for_update,
lr=args.lr,
momentum=args.momentum,
weight_decay=weight_decay)
if 1:
# helping optimizer to implement group lasso (with very small weight that doesn't affect training)
# will be used to calculate number of remaining flops and parameters in the network
group_wd_optimizer = group_lasso_decay(
parameters_for_update,
group_lasso_weight=args.group_wd_coeff,
named_parameters=parameters_for_update_named,
output_sizes=output_sizes)
cudnn.benchmark = True
# define objective
criterion = nn.CrossEntropyLoss()
###=======================added for pruning
# logging part
log_save_folder = "%s" % args.name
if not os.path.exists(log_save_folder):
os.makedirs(log_save_folder)
if not os.path.exists("%s/models" % (log_save_folder)):
os.makedirs("%s/models" % (log_save_folder))
train_writer = None
if args.tensorboard:
try:
# tensorboardX v1.6
train_writer = SummaryWriter(log_dir="%s" % (log_save_folder))
except:
# tensorboardX v1.7
train_writer = SummaryWriter(logdir="%s" % (log_save_folder))
time_point = time.strftime("%Y_%m_%d_%H_%M_%S", time.localtime())
textfile = "%s/log_%s.txt" % (log_save_folder, time_point)
stdout = Logger(textfile)
sys.stdout = stdout
print(" ".join(sys.argv))
# initializing parameters for pruning
# we can add weights of different layers or we can add gates (multiplies output with 1, useful only for gradient computation)
pruning_engine = None
if args.pruning:
pruning_settings = dict()
if not (args.pruning_config is None):
pruning_settings_reader = PruningConfigReader()
pruning_settings_reader.read_config(args.pruning_config)
pruning_settings = pruning_settings_reader.get_parameters()
# overwrite parameters from config file with those from command line
# needs manual entry here
# user_specified = [key for key in vars(default_args).keys() if not (vars(default_args)[key]==vars(args)[key])]
# argv_of_interest = ['pruning_threshold', 'pruning-momentum', 'pruning_step', 'prune_per_iteration',
# 'fixed_layer', 'start_pruning_after_n_iterations', 'maximum_pruning_iterations',
# 'starting_neuron', 'prune_neurons_max', 'pruning_method']
has_attribute = lambda x: any([x in a for a in sys.argv])
if has_attribute('pruning-momentum'):
pruning_settings['pruning_momentum'] = vars(
args)['pruning_momentum']
if has_attribute('pruning-method'):
pruning_settings['method'] = vars(args)['pruning_method']
pruning_parameters_list = prepare_pruning_list(
pruning_settings,
model,
model_name=args.model,
pruning_mask_from=args.pruning_mask_from,
name=args.name)
print("Total pruning layers:", len(pruning_parameters_list))
folder_to_write = "%s" % log_save_folder + "/"
log_folder = folder_to_write
pruning_engine = pytorch_pruning(pruning_parameters_list,
pruning_settings=pruning_settings,
log_folder=log_folder)
pruning_engine.connect_tensorboard(train_writer)
pruning_engine.dataset = args.dataset
pruning_engine.model = args.model
pruning_engine.pruning_mask_from = args.pruning_mask_from
pruning_engine.load_mask()
gates_to_params = connect_gates_with_parameters_for_flops(
args.model, parameters_for_update_named)
pruning_engine.gates_to_params = gates_to_params
###=======================end for pruning
# loading model file
if (len(args.load_model) > 0) and (not args.dynamic_network):
if os.path.isfile(args.load_model):
if args.fineturn_model:
checkpoint = torch.load(args.load_model)
state_dict = checkpoint['state_dict']
model = load_module_state_dict_checkpoint(model, state_dict)
else:
load_model_pytorch(model, args.load_model, args.model)
else:
print("=> no checkpoint found at '{}'".format(args.load_model))
exit()
if args.tensorboard and 0:
if args.dataset == "CIFAR10":
dummy_input = torch.rand(1, 3, 32, 32).to(device)
elif args.dataset == "Imagenet":
dummy_input = torch.rand(1, 3, 224, 224).to(device)
train_writer.add_graph(model, dummy_input)
for epoch in range(1, args.epochs + 1):
if args.distributed:
train_sampler.set_epoch(epoch)
adjust_learning_rate(args, optimizer, epoch, args.zero_lr_for_epochs,
train_writer)
if not args.run_test and not args.get_inference_time:
train(args,
model,
device,
train_loader,
optimizer,
epoch,
criterion,
train_writer=train_writer,
pruning_engine=pruning_engine)
if args.pruning:
# skip validation error calculation and model saving
if pruning_engine.method == 50: continue
# evaluate on validation set
prec1 = validate(args,
test_loader,
model,
device,
criterion,
epoch,
train_writer=train_writer)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
model_save_path = "%s/models/checkpoint.weights" % (log_save_folder)
paths = "%s/models" % (log_save_folder)
model_state_dict = model.state_dict()
if args.save_models:
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model_state_dict,
'best_prec1': best_prec1,
},
is_best,
filename=model_save_path)
states = {
'epoch': epoch + 1,
'state_dict': model_state_dict,
}
torch.save(states, '{}/{}.pth.tar'.format(paths, epoch + 1))
import pandas as pd
from PIL import Image
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
import os
import torch
import json
import time
import argparse
import pickle
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 __init__(self, setname, args, augment=False):
im_size = args.orig_imsize
csv_path = osp.join(SPLIT_PATH, setname + '.csv')
cache_path = osp.join( CACHE_PATH, "{}.{}.{}.pt".format(self.__class__.__name__, setname, im_size) )
self.use_im_cache = ( im_size != -1 ) # not using cache
if self.use_im_cache:
if not osp.exists(cache_path):
print('* Cache miss... Preprocessing {}...'.format(setname))
resize_ = identity if im_size < 0 else transforms.Resize(im_size)
data, label = self.parse_csv(csv_path, setname)
self.data = [ resize_(Image.open(path).convert('RGB')) for path in data ]
self.label = label
print('* Dump cache from {}'.format(cache_path))
torch.save({'data': self.data, 'label': self.label }, cache_path)
else:
print('* Load cache from {}'.format(cache_path))
cache = torch.load(cache_path)
self.data = cache['data']
self.label = cache['label']
else:
self.data, self.label = self.parse_csv(csv_path, setname)
self.num_class = len(set(self.label))
image_size = 84
if augment and setname == 'train':
transforms_list = [
transforms.RandomResizedCrop(image_size),
transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
]
else:
transforms_list = [
transforms.Resize(92),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
]
# Transformation
if args.backbone_class == 'ConvNet':
self.transform = transforms.Compose(
transforms_list + [
transforms.Normalize(np.array([0.485, 0.456, 0.406]),
np.array([0.229, 0.224, 0.225]))
])
elif args.backbone_class == 'Res12':
self.transform = transforms.Compose(
transforms_list + [
transforms.Normalize(np.array([x / 255.0 for x in [120.39586422, 115.59361427, 104.54012653]]),
np.array([x / 255.0 for x in [70.68188272, 68.27635443, 72.54505529]]))
])
elif args.backbone_class == 'Res18':
self.transform = transforms.Compose(
transforms_list + [
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
elif args.backbone_class == 'WRN':
self.transform = transforms.Compose(
transforms_list + [
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
else:
raise ValueError('Non-supported Network Types. Please Revise Data Pre-Processing Scripts.')
def load_data(train=True,
data_dir='dataset/imagenet',
batch_size=128,
subset_len=None,
sample_method='random',
distributed=False,
model_name='resnet18',
**kwargs):
#prepare data
# random.seed(12345)
traindir = data_dir + '/train'
valdir = data_dir + '/val'
train_sampler = None
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if model_name == 'inception_v3':
size = 299
resize = 299
else:
size = 224
resize = 256
if train:
dataset = torchvision.datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
if subset_len:
assert subset_len <= len(dataset)
if sample_method == 'random':
dataset = torch.utils.data.Subset(
dataset, random.sample(range(0, len(dataset)), subset_len))
else:
dataset = torch.utils.data.Subset(dataset,
list(range(subset_len)))
if distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
dataset)
data_loader = torch.utils.data.DataLoader(
dataset,
batch_size=batch_size,
shuffle=(train_sampler is None),
sampler=train_sampler,
**kwargs)
else:
dataset = torchvision.datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(resize),
transforms.CenterCrop(size),
transforms.ToTensor(),
normalize,
]))
if subset_len:
assert subset_len <= len(dataset)
if sample_method == 'random':
dataset = torch.utils.data.Subset(
dataset, random.sample(range(0, len(dataset)), subset_len))
else:
dataset = torch.utils.data.Subset(dataset,
list(range(subset_len)))
data_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
shuffle=False,
**kwargs)
return data_loader, train_sampler
def train(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Save the arguments.
with open(os.path.join(args.model_path, 'args.json'), 'w') as args_file:
json.dump(args.__dict__, args_file)
# Config logging.
log_format = '%(levelname)-8s %(message)s'
logfile = os.path.join(args.model_path, 'train.log')
logging.basicConfig(filename=logfile, level=logging.INFO, format=log_format)
logging.getLogger().addHandler(logging.StreamHandler())
logging.info(json.dumps(args.__dict__))
# Image preprocessing
transform = transforms.Compose([
transforms.ToTensor(),
transforms.ToPILImage(),
transforms.RandomResizedCrop(args.crop_size,
scale=(1.00, 1.2),
ratio=(0.75, 1.3333333333333333)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])])##0.5
# Load vocabulary wrapper.
vocab = load_vocab(args.vocab_path)
# Build data loader
logging.info("Building data loader...")
train_sampler = None
val_sampler = None
data_loader = get_loader(args.dataset, transform,
args.batch_size, shuffle=False,
num_workers=args.num_workers,
max_examples=args.max_examples,
sampler=train_sampler)
val_data_loader = get_loader(args.val_dataset, transform,
args.batch_size, shuffle=False,
num_workers=args.num_workers,
max_examples=args.max_examples,
sampler=val_sampler)
logging.info("Done")
vqg = create_model(args, vocab)
print(vqg)
"""
graph = hl.build_graph(vqg.modules())
graph = graph.build_dot()
graph.render("weights/tf1/", view=True, format='png')
g = make_dot(vqg.modules(),vqg.state_dict())
g.view()"""
if args.load_model is not None:
vqg.load_state_dict(torch.load(args.load_model))
logging.info("Done")
# Loss criterion.
pad = vocab(vocab.SYM_PAD) # Set loss weight for 'pad' symbol to 0
criterion = nn.CrossEntropyLoss(ignore_index=pad)
l2_criterion = nn.MSELoss()
# Setup GPUs.
if torch.cuda.is_available():
logging.info("Using available GPU...")
vqg.cuda()
criterion.cuda()
l2_criterion.cuda()
# Parameters to train.
gen_params = vqg.generator_parameters()
info_params = vqg.info_parameters()
learning_rate = args.learning_rate
info_learning_rate = args.learning_rate
gen_optimizer = torch.optim.Adam(gen_params, lr=learning_rate)
info_optimizer = torch.optim.Adam(info_params, lr=info_learning_rate)
scheduler = ReduceLROnPlateau(optimizer=gen_optimizer, mode='min',
factor=0.1, patience=args.patience,
verbose=True, min_lr=1e-7)
info_scheduler = ReduceLROnPlateau(optimizer=info_optimizer, mode='min',
factor=0.1, patience=args.patience,
verbose=True, min_lr=1e-7)
# Train the model.
total_steps = len(data_loader)
start_time = time.time()
n_steps = 0
for epoch in range(args.num_epochs):
for i, (images, questions, qindices) in enumerate(data_loader):
n_steps += 1
# Set mini-batch dataset.
if torch.cuda.is_available():
images = images.cuda()
questions = questions.cuda()
qindices = qindices.cuda()
# Eval now.
if (args.eval_every_n_steps is not None and
n_steps >= args.eval_every_n_steps and
n_steps % args.eval_every_n_steps == 0):
run_eval(vqg, val_data_loader, criterion, l2_criterion,
args, epoch, scheduler,info_scheduler)
compare_outputs(images, questions, vqg, vocab, logging, args)
# Forward.
vqg.train()
gen_optimizer.zero_grad()
info_optimizer.zero_grad()
image_features = vqg.encode_images(images)
# Question generation.
mus, logvars = vqg.encode_into_z(image_features)
zs = vqg.reparameterize(mus, logvars)
(outputs, _, _) = vqg.decode_questions(
image_features, zs, questions=questions,
teacher_forcing_ratio=1.0)
# Reorder the questions based on length.
questions = torch.index_select(questions, 0, qindices)
# Ignoring the start token.
questions = questions[:, 1:]
qlengths = process_lengths(questions)
# Convert the output from MAX_LEN list of (BATCH x VOCAB) ->
# (BATCH x MAX_LEN x VOCAB).
outputs = [o.unsqueeze(1) for o in outputs]
outputs = torch.cat(outputs, dim=1)
outputs = torch.index_select(outputs, 0, qindices)
# Calculate the generation loss.
targets = pack_padded_sequence(questions, qlengths,
batch_first=True)[0]
outputs = pack_padded_sequence(outputs, qlengths,
batch_first=True)[0]
gen_loss = criterion(outputs, targets)
total_loss = 0.0
total_loss += args.lambda_gen * gen_loss
gen_loss = gen_loss.item()
# Variational loss.
kl_loss = gaussian_KL_loss(mus, logvars)
total_loss += args.lambda_z * kl_loss
kl_loss = kl_loss.item()
total_loss.backward()
gen_optimizer.step()
# Print log info
if i % args.log_step == 0:
delta_time = time.time() - start_time
start_time = time.time()
logging.info('Time: %.4f, Epoch [%d/%d], Step [%d/%d], '
'LR: %f, gen: %.4f, KL: %.4f '
% (delta_time, epoch, args.num_epochs, i,
total_steps, gen_optimizer.param_groups[0]['lr'],
gen_loss, kl_loss
))
# Save the models
if args.save_step is not None and (i+1) % args.save_step == 0:
torch.save(vqg.state_dict(),
os.path.join(args.model_path,
'vqg-tf-%d-%d.pkl'
% (epoch + 1, i + 1)))
torch.save(vqg.state_dict(),
os.path.join(args.model_path,
'vqg-tf-%d.pkl' % (epoch+1)))
# Evaluation and learning rate updates.
run_eval(vqg, val_data_loader, criterion, l2_criterion,
args, epoch, scheduler,info_scheduler)
def main():
global args, best_prec1
args = parser.parse_args()
if not os.path.isdir(args.checkpoint):
mkdir_p(args.checkpoint)
print('dataset:', args.dataset)
if args.dataset == 'EPIC':
with open(args.foptions, 'rb') as handle:
options = pickle.load(handle)
if options['root'][0] == '.':
# update relative path
options['root'] = 'basemodel' + options['root'][1:]
options['num_crops'] = 1
else:
options = None
model = models.Net(options=options).cuda()
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
extractor_params = list(map(id, model.extractor.parameters()))
classifier_params = filter(lambda p: id(p) not in extractor_params,
model.parameters())
optimizer = torch.optim.SGD([{
'params': model.extractor.parameters()
}, {
'params': classifier_params,
'lr': args.lr * 10
}],
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=30,
gamma=0.1)
# optionally resume from a checkpoint
# title = 'CUB'
title = args.dataset
if args.resume:
print('==> Resuming from checkpoint..')
assert os.path.isfile(
args.resume), 'Error: no checkpoint directory found!'
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
logger = Logger(os.path.join(args.checkpoint, 'log.txt'),
title=title,
resume=True)
else:
logger = Logger(os.path.join(args.checkpoint, 'log.txt'), title=title)
logger.set_names([
'Learning Rate', 'Train Loss', 'Valid Loss', 'Train Acc.',
'Valid Acc.'
])
cudnn.benchmark = True
print(' Total params: %.2fM' %
(sum(p.numel() for p in model.parameters()) / 1000000.0))
# Data loading code
normalize = transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
if args.dataset == 'CUB':
print('Using dataset CUB')
train_dataset = loader.ImageLoader(
args.data,
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]),
train=True)
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(loader.ImageLoader(
args.data,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
elif args.dataset == 'EPIC':
print('Using dataset EPIC')
train_dataset, val_dataset, train_loader, val_loader = get_datasets_and_dataloaders(
options, device=options['device'])
if args.evaluate:
validate(val_loader, model, criterion)
return
print('# train_loader:', len(train_loader))
print('# val_loader:', len(val_loader))
print('Training epoch from {} to {}'.format(args.start_epoch, args.epochs))
for epoch in range(args.start_epoch, args.epochs):
scheduler.step()
lr = optimizer.param_groups[1]['lr']
print('\nEpoch: [%d | %d] LR: %f' % (epoch + 1, args.epochs, lr))
# train for one epoch
train_loss, train_acc = train(train_loader, model, criterion,
optimizer, epoch)
# evaluate on validation set
test_loss, test_acc = validate(val_loader, model, criterion)
# append logger file
logger.append([lr, train_loss, test_loss, train_acc, test_acc])
# remember best prec@1 and save checkpoint
is_best = test_acc > best_prec1
best_prec1 = max(test_acc, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
},
is_best,
checkpoint=args.checkpoint)
logger.close()
logger.plot()
savefig(os.path.join(args.checkpoint, 'log.eps'))
print('Best acc:')
print(best_prec1)
def main_worker(gpu, ngpus_per_node, args):
global best_acc1
# global aim_sess = aim.Session()
# aim_sess.set_params({
# 'num_epochs': args.epochs,
# 'num_classes': 1000,
# 'batch_size': args.batch_size,
# }, name='hparams')
args.gpu = gpu
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)
repvgg_build_func = get_RepVGG_func_by_name(args.arch)
model = repvgg_build_func(deploy=False)
if not torch.cuda.is_available():
print('using CPU, this will be slow')
elif 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 - 1) / ngpus_per_node)
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
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
model = torch.nn.DataParallel(model).cuda()
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda(args.gpu)
optimizer = sgd_optimizer(model, args.lr, args.momentum, args.weight_decay)
lr_scheduler = CosineAnnealingLR(
optimizer=optimizer,
T_max=args.epochs * IMAGENET_TRAINSET_SIZE // args.batch_size //
ngpus_per_node)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
if args.gpu is None:
checkpoint = torch.load(args.resume)
else:
# Map model to be loaded to specified single gpu.
loc = 'cuda:{}'.format(args.gpu)
checkpoint = torch.load(args.resume, map_location=loc)
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'])
lr_scheduler.load_state_dict(checkpoint['scheduler'])
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')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
train_sampler = 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_dataset = datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
]))
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion, args)
return
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,
lr_scheduler)
# 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)
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(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
'scheduler': lr_scheduler.state_dict(),
}, is_best)
def build_imagenet(model_state_dict, optimizer_state_dict, **kwargs):
ratio = kwargs.pop('ratio')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_transform = transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=0.4,
contrast=0.4,
saturation=0.4,
hue=0.2),
transforms.ToTensor(),
normalize,
])
valid_transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])
if args.zip_file:
logging.info('Loading data from zip file')
traindir = os.path.join(args.data, 'train.zip')
if args.lazy_load:
train_data = utils.ZipDataset(traindir, train_transform)
else:
logging.info('Loading data into memory')
train_data = utils.InMemoryZipDataset(traindir,
train_transform,
num_workers=32)
else:
logging.info('Loading data from directory')
traindir = os.path.join(args.data, 'train')
if args.lazy_load:
train_data = dset.ImageFolder(traindir, train_transform)
else:
logging.info('Loading data into memory')
train_data = utils.InMemoryDataset(traindir,
train_transform,
num_workers=32)
num_train = len(train_data)
indices = list(range(num_train))
np.random.shuffle(indices)
split = int(np.floor(ratio * num_train))
train_indices = sorted(indices[:split])
valid_indices = sorted(indices[split:])
train_queue = torch.utils.data.DataLoader(
train_data,
batch_size=args.child_batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(train_indices),
pin_memory=True,
num_workers=16)
valid_queue = torch.utils.data.DataLoader(
train_data,
batch_size=args.child_eval_batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(valid_indices),
pin_memory=True,
num_workers=16)
model = NASWSNetworkImageNet(args, 1000, args.child_layers,
args.child_nodes, args.child_channels,
args.child_keep_prob,
args.child_drop_path_keep_prob,
args.child_use_aux_head, args.steps)
model = model.cuda()
train_criterion = CrossEntropyLabelSmooth(1000,
args.child_label_smooth).cuda()
eval_criterion = nn.CrossEntropyLoss().cuda()
logging.info("param size = %fMB", utils.count_parameters_in_MB(model))
optimizer = torch.optim.SGD(
model.parameters(),
args.child_lr,
momentum=0.9,
weight_decay=args.child_l2_reg,
)
if model_state_dict is not None:
model.load_state_dict(model_state_dict)
if optimizer_state_dict is not None:
optimizer.load_state_dict(optimizer_state_dict)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
args.child_decay_period,
gamma=args.child_gamma)
return train_queue, valid_queue, model, train_criterion, eval_criterion, optimizer, scheduler
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--dataroot',
default='/mnt/Data1',
help='path to images')
parser.add_argument('--workers',
default=4,
type=int,
help='number of data loading workers (default: 4)')
parser.add_argument('--batch-size',
default=2,
type=int,
help='mini-batch size')
parser.add_argument('--outroot',
default='./results',
help='path to save the results')
parser.add_argument('--exp-name',
default='test',
help='name of expirement')
parser.add_argument('--load',
default='',
help='name of pth to load weights from')
parser.add_argument('--freeze-cc-net',
dest='freeze_cc_net',
action='store_true',
help='dont train the color corrector net')
parser.add_argument('--freeze-warp-net',
dest='freeze_warp_net',
action='store_true',
help='dont train the warp net')
parser.add_argument('--test',
dest='test',
action='store_true',
help='only test the network')
parser.add_argument(
'--synth-data',
dest='synth_data',
action='store_true',
help='use synthetic data instead of tank data for training')
parser.add_argument('--epochs',
default=3,
type=int,
help='number of epochs to train for')
parser.add_argument('--no-warp-net',
dest='warp_net',
action='store_false',
help='do not include warp net in the model')
parser.add_argument('--warp-net-downsample',
default=3,
type=int,
help='number of downsampling layers in warp net')
parser.add_argument('--no-color-net',
dest='color_net',
action='store_false',
help='do not include color net in the model')
parser.add_argument('--color-net-downsample',
default=3,
type=int,
help='number of downsampling layers in color net')
parser.add_argument(
'--no-color-net-skip',
dest='color_net_skip',
action='store_false',
help='dont use u-net skip connections in the color net')
parser.add_argument(
'--dim',
default=32,
type=int,
help='initial feature dimension (doubled at each downsampling layer)')
parser.add_argument('--n-res',
default=8,
type=int,
help='number of residual blocks')
parser.add_argument('--norm',
default='gn',
type=str,
help='type of normalization layer')
parser.add_argument(
'--denormalize',
dest='denormalize',
action='store_true',
help='denormalize output image by input image mean/var')
parser.add_argument(
'--weight-X-L1',
default=1.,
type=float,
help='weight of L1 reconstruction loss after color corrector net')
parser.add_argument('--weight-Y-L1',
default=1.,
type=float,
help='weight of L1 reconstruction loss after warp net')
parser.add_argument('--weight-Y-VGG',
default=1.,
type=float,
help='weight of perceptual loss after warp net')
parser.add_argument(
'--weight-Z-L1',
default=1.,
type=float,
help='weight of L1 reconstruction loss after color net')
parser.add_argument('--weight-Z-VGG',
default=.5,
type=float,
help='weight of perceptual loss after color net')
parser.add_argument('--weight-Z-Adv',
default=0.2,
type=float,
help='weight of adversarial loss after color net')
args = parser.parse_args()
# set random seed for consistent fixed batch
torch.manual_seed(8)
# set weights of losses of intermediate outputs to zero if not necessary
if not args.warp_net:
args.weight_Y_L1 = 0
args.weight_Y_VGG = 0
if not args.color_net:
args.weight_Z_L1 = 0
args.weight_Z_VGG = 0
args.weight_Z_Adv = 0
# datasets
train_dir_1 = os.path.join(args.dataroot, 'Water', 'train')
train_dir_2 = os.path.join(args.dataroot, 'ImageNet', 'train')
val_dir_1 = os.path.join(args.dataroot, 'Water', 'test')
val_dir_2 = os.path.join(args.dataroot, 'ImageNet', 'test')
test_dir = os.path.join(args.dataroot, 'Water_Real')
if args.synth_data:
train_data = ImageFolder(train_dir_2,
transform=transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5]),
synthdata.SynthData(224,
n=args.batch_size),
]))
else:
train_data = PairedImageFolder(
train_dir_1,
train_dir_2,
transform=transforms.Compose([
pairedtransforms.RandomResizedCrop(224),
pairedtransforms.RandomHorizontalFlip(),
pairedtransforms.ToTensor(),
pairedtransforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5]),
]))
val_data = PairedImageFolder(val_dir_1,
val_dir_2,
transform=transforms.Compose([
pairedtransforms.Resize(256),
pairedtransforms.CenterCrop(256),
pairedtransforms.ToTensor(),
pairedtransforms.Normalize(
mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5]),
]))
test_data = ImageFolder(test_dir,
transform=transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(256),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5]),
]),
return_path=True)
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=args.batch_size,
num_workers=args.workers,
pin_memory=True,
shuffle=True)
val_loader = torch.utils.data.DataLoader(val_data,
batch_size=args.batch_size,
num_workers=args.workers,
pin_memory=True,
shuffle=True)
test_loader = torch.utils.data.DataLoader(test_data,
batch_size=args.batch_size,
num_workers=args.workers,
pin_memory=True,
shuffle=False)
# fixed test batch for visualization during training
fixed_batch = iter(val_loader).next()[0]
# model
model = networks.Model(args)
model.cuda()
# load weights from checkpoint
if args.test and not args.load:
args.load = args.exp_name
if args.load:
model.load_state_dict(torch.load(
os.path.join(args.outroot, '%s_net.pth' % args.load)),
strict=args.test)
# create outroot if necessary
if not os.path.exists(args.outroot):
os.makedirs(args.outroot)
# if args.test only run test script
if args.test:
test(test_loader, model, args)
return
# main training loop
for epoch in range(args.epochs):
train(train_loader, model, fixed_batch, epoch, args)
torch.save(model.state_dict(),
os.path.join(args.outroot, '%s_net.pth' % args.exp_name))
test(test_loader, model, args)
def transforming(path):
# Transform image data into tensors to be used in PyTorch
# Getting the specific paths to each of the training, validation and testing folder
data_dir = path
train_dir = data_dir + '/train'
valid_dir = data_dir + '/valid'
test_dir = data_dir + '/test'
# Preparing the transformations needed to convert images into tensors
data_transforms = {
'training':
transforms.Compose([
transforms.RandomRotation(30),
transforms.RandomHorizontalFlip(30),
transforms.RandomResizedCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'validation':
transforms.Compose([
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'testing':
transforms.Compose([
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
}
# Loading the data from the directories
image_datasets = {
'training':
datasets.ImageFolder(train_dir, transform=data_transforms['training']),
'validation':
datasets.ImageFolder(valid_dir,
transform=data_transforms['validation']),
'testing':
datasets.ImageFolder(test_dir, transform=data_transforms['testing'])
}
# Batch iterators to feed the network
dataloaders = {
'training':
torch.utils.data.DataLoader(image_datasets['training'],
batch_size=100,
shuffle=True),
'validation':
torch.utils.data.DataLoader(image_datasets['validation'],
batch_size=100,
shuffle=True),
'testing':
torch.utils.data.DataLoader(image_datasets['testing'],
batch_size=100,
shuffle=True)
}
return dataloaders, image_datasets
sample = np.array(sample)
# blur the image with a 50% chance
prob = np.random.random_sample()
if prob < 0.5:
sigma = (self.max -
self.min) * np.random.random_sample() + self.min
sample = cv2.GaussianBlur(sample,
(self.kernel_size, self.kernel_size),
sigma)
return sample
train_transform = transforms.Compose([
transforms.RandomResizedCrop(32),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomApply([transforms.ColorJitter(0.4, 0.4, 0.4, 0.1)],
p=0.8),
transforms.RandomGrayscale(p=0.2),
GaussianBlur(kernel_size=int(0.1 * 32)),
transforms.ToTensor(),
transforms.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010])
])
test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010])
])
def __init__(self,
normalize=None,
size_crops: List[int] = [96, 36],
nmb_crops: List[int] = [2, 4],
min_scale_crops: List[float] = [0.33, 0.10],
max_scale_crops: List[float] = [1, 0.33],
gaussian_blur: bool = True,
jitter_strength: float = 1.):
self.jitter_strength = jitter_strength
self.gaussian_blur = gaussian_blur
assert len(size_crops) == len(nmb_crops)
assert len(min_scale_crops) == len(nmb_crops)
assert len(max_scale_crops) == len(nmb_crops)
self.size_crops = size_crops
self.nmb_crops = nmb_crops
self.min_scale_crops = min_scale_crops
self.max_scale_crops = max_scale_crops
self.color_jitter = transforms.ColorJitter(0.8 * self.jitter_strength,
0.8 * self.jitter_strength,
0.8 * self.jitter_strength,
0.2 * self.jitter_strength)
transform = []
color_transform = [
transforms.RandomApply([self.color_jitter], p=0.8),
transforms.RandomGrayscale(p=0.2)
]
if self.gaussian_blur:
kernel_size = int(0.1 * self.size_crops[0])
if kernel_size % 2 == 0:
kernel_size += 1
color_transform.append(GaussianBlur(kernel_size=kernel_size,
p=0.5))
self.color_transform = transforms.Compose(color_transform)
if normalize is None:
self.final_transform = transforms.ToTensor()
else:
self.final_transform = transforms.Compose(
[transforms.ToTensor(), normalize])
for i in range(len(self.size_crops)):
random_resized_crop = transforms.RandomResizedCrop(
self.size_crops[i],
scale=(self.min_scale_crops[i], self.max_scale_crops[i]),
)
transform.extend([
transforms.Compose([
random_resized_crop,
transforms.RandomHorizontalFlip(p=0.5),
self.color_transform, self.final_transform
])
] * self.nmb_crops[i])
self.transform = transform
# add online train transform of the size of global view
online_train_transform = transforms.Compose([
transforms.RandomResizedCrop(self.size_crops[0]),
transforms.RandomHorizontalFlip(), self.final_transform
])
self.transform.append(online_train_transform)
def main_worker(gpu, ngpus_per_node, args):
global best_acc1
args.gpu = gpu
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 args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
model = models.__dict__[args.arch](pretrained=True)
else:
print("=> creating model '{}'".format(args.arch))
model = models.__dict__[args.arch]()
n_ftrs = model.fc.in_features
model.fc = nn.Linear(n_ftrs, 3)
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])
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 args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
# 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')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_dataset = Synthetic('../annotation.json',
subset='training',
transform=transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
val_dataset = Synthetic('../annotation.json',
subset='validation',
transform=transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
train_sampler = 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=False,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion, args)
return
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)
# 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)
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(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
}, is_best)
def load_data(
prop_noPV_training: float,
min_rescale_images: float,
batch_size: int,
dir_data_training: str = "",
dir_data_validation: str = "",
dir_data_test: str = "",
):
"""
Create the DataLoader objects that will generate the training, validation and test sets.
Parameters
----------
prop_noPV_training : float
Proportion of noPV images to add for the training of the model.
min_rescale_images : float
Minimum proportion of the image to keep for the RandomResizedCrop transform.
batch_size : int
Number of samples per batch in the DataLoaders.
dir_data_training : str, optional
Directory where the folders "images/", "labels/" and "noPV/" are for the training set.
If empty, the data is not generated. The default is "".
dir_data_validation : str, optional
Directory where the folders "images/", "labels/" and "noPV/" are for the validation set.
If empty, the data is not generated. The default is "".
dir_data_test : str, optional
Directory where the folders "images/", "labels/" and "noPV/" are for the test set.
If empty, the data is not generated. The default is "".
Returns
-------
train_dl : torch.utils.data.DataLoader
Training DataLoader, if data directory is provided, otherwise None.
validation_dl : torch.utils.data.DataLoader
Validation DataLoader, if data directory is provided, otherwise None.
test_dl : torch.utils.data.DataLoader
Test DataLoader, if data directory is provided, otherwise None.
"""
roof_train_dataset = None
if dir_data_training:
# Transforms to augment the data (for training set)
transform_aug = transforms.Compose([
transforms.ToPILImage(),
transforms.RandomResizedCrop(250,
scale=(min_rescale_images, 1.0),
ratio=(1.0, 1.0)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
])
# Instantiate the training dataset
roof_train_dataset = AvailableRooftopDataset(
dir_PV=os.path.join(dir_data_training, "PV"),
dir_noPV=os.path.join(dir_data_training, "noPV"),
dir_labels=os.path.join(dir_data_training, "labels"),
transform=transform_aug,
prop_noPV=prop_noPV_training,
)
# No transform applied to validation and train images (the model should not need
# any preprocessing)
transform_id = transforms.Compose([
transforms.ToPILImage(),
transforms.ToTensor(),
])
# Instantiate the validation and test datasets
roof_validation_dataset, roof_test_dataset = (
AvailableRooftopDataset(
dir_PV=os.path.join(dir_data, "PV"),
dir_noPV=os.path.join(dir_data, "noPV"),
dir_labels=os.path.join(dir_data, "labels"),
transform=transform_id,
prop_noPV=-1, # All of them
) if dir_data else None
for dir_data in (dir_data_validation, dir_data_test))
# Instantiate the DataLoaders
roof_train_dl, roof_validation_dl, roof_test_dl = (DataLoader(
roof_dataset, batch_size=batch_size, shuffle=True,
num_workers=0) if roof_dataset else None for roof_dataset in (
roof_train_dataset,
roof_validation_dataset,
roof_test_dataset,
))
return roof_train_dl, roof_validation_dl, roof_test_dl
def get_dataloader(args, add_erasing=False, aug_plus=False):
if 'cifar' in args.dataset or 'kitchen' in args.dataset:
if aug_plus:
# MoCo v2's aug: similar to SimCLR https://arxiv.org/abs/2002.05709
transform_train_list = [
transforms.RandomResizedCrop(size=32, scale=(0.2, 1.)),
transforms.RandomApply(
[
transforms.ColorJitter(0.4, 0.4, 0.4,
0.1) # not strengthened
],
p=0.8),
transforms.RandomGrayscale(p=0.2),
transforms.RandomApply([GaussianBlur([.1, 2.])], p=0.5),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010)),
]
else:
transform_train_list = [
transforms.RandomResizedCrop(size=32, scale=(0.2, 1.)),
transforms.ColorJitter(0.4, 0.4, 0.4, 0.4),
transforms.RandomGrayscale(p=0.2),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010)),
]
if add_erasing:
transform_train_list.append(transforms.RandomErasing(p=1.0))
transform_train = transforms.Compose(transform_train_list)
if 'kitchen' in args.dataset:
transform_test = transforms.Compose([
transforms.Resize((32, 32), interpolation=2),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010)),
])
else:
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010)),
])
elif 'stl' in args.dataset:
transform_train = transforms.Compose([
transforms.RandomResizedCrop(size=96, scale=(0.2, 1.)),
transforms.ColorJitter(0.4, 0.4, 0.4, 0.4),
transforms.RandomGrayscale(p=0.2),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
if args.dataset == 'cifar10':
trainset = datasets.CIFAR10Instance(root='./data/CIFAR-10',
train=True,
download=True,
transform=transform_train,
two_imgs=args.two_imgs,
three_imgs=args.three_imgs)
train_sampler = torch.utils.data.distributed.DistributedSampler(
trainset)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=False,
sampler=train_sampler)
testset = datasets.CIFAR10Instance(root='./data/CIFAR-10',
train=False,
download=True,
transform=transform_test)
testloader = torch.utils.data.DataLoader(testset,
batch_size=100,
shuffle=False,
num_workers=2,
pin_memory=False)
args.pool_len = 4
ndata = trainset.__len__()
elif args.dataset == 'cifar100':
trainset = datasets.CIFAR100Instance(root='./data/CIFAR-100',
train=True,
download=True,
transform=transform_train,
two_imgs=args.two_imgs,
three_imgs=args.three_imgs)
train_sampler = torch.utils.data.distributed.DistributedSampler(
trainset)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=False,
sampler=train_sampler)
testset = datasets.CIFAR100Instance(root='./data/CIFAR-100',
train=False,
download=True,
transform=transform_test)
testloader = torch.utils.data.DataLoader(testset,
batch_size=100,
shuffle=False,
num_workers=2,
pin_memory=False)
args.pool_len = 4
ndata = trainset.__len__()
elif args.dataset == 'stl10':
trainset = datasets.STL10(root='./data/STL10',
split='train',
download=True,
transform=transform_train,
two_imgs=args.two_imgs,
three_imgs=args.three_imgs)
train_sampler = torch.utils.data.distributed.DistributedSampler(
trainset)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=False,
sampler=train_sampler)
testset = datasets.STL10(root='./data/STL10',
split='test',
download=True,
transform=transform_test)
testloader = torch.utils.data.DataLoader(testset,
batch_size=100,
shuffle=False,
num_workers=2,
pin_memory=False)
args.pool_len = 7
ndata = trainset.__len__()
elif args.dataset == 'stl10-full':
trainset = datasets.STL10(root='./data/STL10',
split='train+unlabeled',
download=True,
transform=transform_train,
two_imgs=args.two_imgs,
three_imgs=args.three_imgs)
train_sampler = torch.utils.data.distributed.DistributedSampler(
trainset)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=False,
sampler=train_sampler)
labeledTrainset = datasets.STL10(root='./data/STL10',
split='train',
download=True,
transform=transform_train,
two_imgs=args.two_imgs)
labeledTrain_sampler = torch.utils.data.distributed.DistributedSampler(
labeledTrainset)
labeledTrainloader = torch.utils.data.DataLoader(
labeledTrainset,
batch_size=args.batch_size,
shuffle=False,
num_workers=2,
pin_memory=False,
sampler=labeledTrain_sampler)
testset = datasets.STL10(root='./data/STL10',
split='test',
download=True,
transform=transform_test)
testloader = torch.utils.data.DataLoader(testset,
batch_size=100,
shuffle=False,
num_workers=2,
pin_memory=False)
args.pool_len = 7
ndata = labeledTrainset.__len__()
elif args.dataset == 'kitchen':
trainset = datasets.CIFARImageFolder(root='./data/Kitchen-HC/train',
train=True,
transform=transform_train,
two_imgs=args.two_imgs,
three_imgs=args.three_imgs)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=False)
testset = datasets.CIFARImageFolder(root='./data/Kitchen-HC/test',
train=False,
transform=transform_test)
testloader = torch.utils.data.DataLoader(testset,
batch_size=100,
shuffle=False,
num_workers=2,
pin_memory=False)
args.pool_len = 4
ndata = trainset.__len__()
return trainloader, testloader, ndata
def main():
global best_acc1
best_acc1 = 0
args = parse_option()
if args.gpu is not None:
print("Use GPU: {} for training".format(args.gpu))
# set the data loader
train_folder = os.path.join(args.data_folder, 'train')
val_folder = os.path.join(args.data_folder, 'val')
image_size = 224
crop_padding = 32
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
normalize = transforms.Normalize(mean=mean, std=std)
if args.aug == 'NULL':
train_transform = transforms.Compose([
transforms.RandomResizedCrop(image_size, scale=(args.crop, 1.)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
elif args.aug == 'CJ':
train_transform = transforms.Compose([
transforms.RandomResizedCrop(image_size, scale=(args.crop, 1.)),
transforms.RandomGrayscale(p=0.2),
transforms.ColorJitter(0.4, 0.4, 0.4, 0.4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
else:
raise NotImplemented('augmentation not supported: {}'.format(args.aug))
train_dataset = datasets.ImageFolder(train_folder, train_transform)
val_dataset = datasets.ImageFolder(
val_folder,
transforms.Compose([
transforms.Resize(image_size + crop_padding),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
normalize,
]))
print(len(train_dataset))
train_sampler = None
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.num_workers,
pin_memory=True,
sampler=train_sampler)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=True)
# create model and optimizer
if args.model == 'resnet50':
model = InsResNet50()
classifier = LinearClassifierResNet(args.layer, args.n_label, 'avg', 1)
elif args.model == 'resnet50x2':
model = InsResNet50(width=2)
classifier = LinearClassifierResNet(args.layer, args.n_label, 'avg', 2)
elif args.model == 'resnet50x4':
model = InsResNet50(width=4)
classifier = LinearClassifierResNet(args.layer, args.n_label, 'avg', 4)
else:
raise NotImplementedError('model not supported {}'.format(args.model))
print('==> loading pre-trained model')
ckpt = torch.load(args.model_path)
model.load_state_dict(ckpt['model'])
print("==> loaded checkpoint '{}' (epoch {})".format(
args.model_path, ckpt['epoch']))
print('==> done')
model = model.cuda()
classifier = classifier.cuda()
criterion = torch.nn.CrossEntropyLoss().cuda(args.gpu)
if not args.adam:
optimizer = torch.optim.SGD(classifier.parameters(),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
else:
optimizer = torch.optim.Adam(classifier.parameters(),
lr=args.learning_rate,
betas=(args.beta1, args.beta2),
weight_decay=args.weight_decay,
eps=1e-8)
model.eval()
cudnn.benchmark = True
# set mixed precision training
# if args.amp:
# model = amp.initialize(model, opt_level=args.opt_level)
# classifier, optimizer = amp.initialize(classifier, optimizer, opt_level=args.opt_level)
# optionally resume from a checkpoint
args.start_epoch = 1
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume, map_location='cpu')
# checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch'] + 1
classifier.load_state_dict(checkpoint['classifier'])
optimizer.load_state_dict(checkpoint['optimizer'])
best_acc1 = checkpoint['best_acc1']
best_acc1 = best_acc1.cuda()
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
if 'opt' in checkpoint.keys():
# resume optimization hyper-parameters
print('=> resume hyper parameters')
if 'bn' in vars(checkpoint['opt']):
print('using bn: ', checkpoint['opt'].bn)
if 'adam' in vars(checkpoint['opt']):
print('using adam: ', checkpoint['opt'].adam)
if 'cosine' in vars(checkpoint['opt']):
print('using cosine: ', checkpoint['opt'].cosine)
args.learning_rate = checkpoint['opt'].learning_rate
# args.lr_decay_epochs = checkpoint['opt'].lr_decay_epochs
args.lr_decay_rate = checkpoint['opt'].lr_decay_rate
args.momentum = checkpoint['opt'].momentum
args.weight_decay = checkpoint['opt'].weight_decay
args.beta1 = checkpoint['opt'].beta1
args.beta2 = checkpoint['opt'].beta2
del checkpoint
torch.cuda.empty_cache()
else:
print("=> no checkpoint found at '{}'".format(args.resume))
# set cosine annealing scheduler
if args.cosine:
# last_epoch = args.start_epoch - 2
# eta_min = args.learning_rate * (args.lr_decay_rate ** 3) * 0.1
# scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs, eta_min, last_epoch)
eta_min = args.learning_rate * (args.lr_decay_rate**3) * 0.1
scheduler = optim.lr_scheduler.CosineAnnealingLR(
optimizer, args.epochs, eta_min, -1)
# dummy loop to catch up with current epoch
for i in range(1, args.start_epoch):
scheduler.step()
# tensorboard
logger = tb_logger.Logger(logdir=args.tb_folder, flush_secs=2)
# routine
for epoch in range(args.start_epoch, args.epochs + 1):
if args.cosine:
scheduler.step()
else:
adjust_learning_rate(epoch, args, optimizer)
print("==> training...")
time1 = time.time()
train_acc, train_acc5, train_loss = train(epoch, train_loader, model,
classifier, criterion,
optimizer, args)
time2 = time.time()
print('train epoch {}, total time {:.2f}'.format(epoch, time2 - time1))
logger.log_value('train_acc', train_acc, epoch)
logger.log_value('train_acc5', train_acc5, epoch)
logger.log_value('train_loss', train_loss, epoch)
logger.log_value('learning_rate', optimizer.param_groups[0]['lr'],
epoch)
print("==> testing...")
test_acc, test_acc5, test_loss = validate(val_loader, model,
classifier, criterion, args)
logger.log_value('test_acc', test_acc, epoch)
logger.log_value('test_acc5', test_acc5, epoch)
logger.log_value('test_loss', test_loss, epoch)
# save the best model
if test_acc > best_acc1:
best_acc1 = test_acc
state = {
'opt': args,
'epoch': epoch,
'classifier': classifier.state_dict(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
}
save_name = '{}_layer{}.pth'.format(args.model, args.layer)
save_name = os.path.join(args.save_folder, save_name)
print('saving best model!')
torch.save(state, save_name)
# save model
if epoch % args.save_freq == 0:
print('==> Saving...')
state = {
'opt': args,
'epoch': epoch,
'classifier': classifier.state_dict(),
'best_acc1': test_acc,
'optimizer': optimizer.state_dict(),
}
save_name = 'ckpt_epoch_{epoch}.pth'.format(epoch=epoch)
save_name = os.path.join(args.save_folder, save_name)
print('saving regular model!')
torch.save(state, save_name)
# tensorboard logger
pass
def main_worker(gpu, ngpus_per_node, args):
global best_acc1
args.gpu = gpu
lr_decay_epoch = [int(i) for i in args.lr_decay_epoch.split(',')]
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 args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
model = models.__dict__[args.arch](pretrained=True)
else:
print("=> creating model '{}'".format(args.arch))
model = models.__dict__[args.arch]()
if not torch.cuda.is_available():
print('using CPU, this will be slow')
elif 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 - 1) / ngpus_per_node)
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
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 args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
criterion = nn.CrossEntropyLoss().cuda(args.gpu)
logname = args.prefix
if args.optimizer == 'sgd':
logname += "SGD_"
print("sgd")
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
elif args.optimizer == 'nero':
logname += "Nero_"
print("Nero")
optimizer = Nero(model.parameters(), lr=args.lr, constraints=True)
cos_sch = False
scheduler = None
T_max = math.ceil(1281167.0 / float(args.batch_size)) * (args.epochs)
if args.sch == 'cos' or args.sch == 'cosine':
print("cosine scheduler")
cos_sch = True
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer,
T_max=T_max,
eta_min=0.0)
logname += args.arch + "_sch_" +str(args.sch)+ "_lr" +str(args.lr) + \
'_epoch' + str(args.epochs) + \
"_opt_" + args.optimizer + \
"_b" + str(args.batch_size) + \
'_momentum' + str(args.momentum) + "_beta" + str(args.beta) + \
'_wd' + str(args.weight_decay)
writer = SummaryWriter(args.logdir + '/' + logname)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
if args.gpu is None:
checkpoint = torch.load(args.resume)
else:
# Map model to be loaded to specified single gpu.
loc = 'cuda:{}'.format(args.gpu)
checkpoint = torch.load(args.resume, map_location=loc)
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'], strict=False)
optimizer.load_state_dict(checkpoint['optimizer'])
for group in optimizer.param_groups:
group["lr"] = args.lr
if args.sch == 'cos':
for i in range(checkpoint['epoch'] *
math.ceil(1281167.0 / float(args.batch_size))):
scheduler.step()
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')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
train_sampler = 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(datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(), normalize
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion, args)
return
lr = args.lr
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
current_lr = optimizer.param_groups[0]['lr']
print("current learning rate: {}".format(current_lr))
writer.add_scalar('lr', current_lr, epoch)
# train for one epoch
top1_train, top5_train, losses_train, batch_time_train, scheduler = train(
train_loader,
model,
criterion,
optimizer,
epoch,
args,
writer,
scheduler=scheduler)
if not cos_sch:
lr = adjust_learning_rate(optimizer, epoch, lr, lr_decay_epoch,
args.lr_decay)
writer.add_scalar('train/batch_time_mean', batch_time_train, epoch)
writer.add_scalar('train/loss_mean', losses_train, epoch)
writer.add_scalar('train/top1_mean', top1_train, epoch)
writer.add_scalar('train/top5_mean', top5_train, epoch)
# evaluate on validation set
top1_val, top5_val, losses_val, batch_time_val = validate(
val_loader, model, criterion, args, epoch, writer)
writer.add_scalar('val/batch_time_mean', batch_time_val, epoch)
writer.add_scalar('val/loss_mean', losses_val, epoch)
writer.add_scalar('val/top1_mean', top1_val, epoch)
writer.add_scalar('val/top5_mean', top5_val, epoch)
# remember best acc@1 and save checkpoint
is_best = top1_val > best_acc1
best_acc1 = max(top1_val, best_acc1)
if not args.multiprocessing_distributed or (
args.multiprocessing_distributed
and args.rank % ngpus_per_node == 0):
if ((epoch + 1) % 5 == 0):
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
}, is_best, args.logdir + '/' + logname + '/epoch' +
str(epoch + 1) + '_checkpoint.pth.tar')
writer.close()
import torch
import torchvision.transforms as transforms
from PIL import Image
from torch.autograd import Variable
# Defines where the model is located.
model_path = 'models/classify_resnet_152_80.pth'
# Defines where the test data is located.
test_folder = './data/test'
# Defines how often to print progress.
print_batch_size = 50
# Defines how to pre-process the image data.
transform = transforms.Compose([
transforms.RandomResizedCrop(200),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
])
# Loads the model and sets to eval mode.
model = torch.load(model_path)
model.eval()
# Iterates over all images in test folder.
output = []
for index, row in pd.read_csv('./data/test.csv').iterrows():
if index % print_batch_size == 0:
print('Progress: #{:5d} | time: {}'.format(index, time.ctime()))
def get_dataloader(args):
# Define data files path.
root_img_folder = "/ais/gobi4/fashion/edge_detection/data_aug"
root_label_folder = "/ais/gobi4/fashion/edge_detection/data_aug"
train_anno_txt = "/ais/gobi4/fashion/edge_detection/data_aug/list_train_aug.txt"
val_anno_txt = "/ais/gobi4/fashion/edge_detection/data_aug/list_test.txt"
train_hdf5_file = "/ais/gobi6/jiaman/github/CASENet/utils/train_aug_label_binary_np.h5"
val_hdf5_file = "/ais/gobi6/jiaman/github/CASENet/utils/test_label_binary_np.h5"
input_size = 472
normalize = transforms.Normalize(mean=[104.008, 116.669, 122.675],
std=[1, 1, 1])
train_augmentation = transforms.Compose([
transforms.RandomResizedCrop(input_size,
scale=(0.75, 1.0),
ratio=(0.75, 1.0)),
transforms.RandomHorizontalFlip()
])
train_label_augmentation = transforms.Compose([transforms.RandomResizedCrop(input_size, scale=(0.75,1.0), ratio=(0.75,1.0), interpolation=PIL.Image.NEAREST), \
transforms.RandomHorizontalFlip()])
train_dataset = CityscapesData(root_img_folder,
root_label_folder,
train_anno_txt,
train_hdf5_file,
input_size,
cls_num=args.cls_num,
img_transform=transforms.Compose([
train_augmentation,
RGB2BGR(roll=True),
ToTorchFormatTensor(div=False),
normalize,
]),
label_transform=transforms.Compose([
transforms.ToPILImage(),
train_label_augmentation,
transforms.ToTensor(),
]))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_dataset = CityscapesData(
root_img_folder,
root_label_folder,
val_anno_txt,
val_hdf5_file,
input_size,
cls_num=args.cls_num,
img_transform=transforms.Compose([
transforms.Resize([input_size, input_size]),
RGB2BGR(roll=True),
ToTorchFormatTensor(div=False),
normalize,
]),
label_transform=transforms.Compose([
transforms.ToPILImage(),
transforms.Resize([input_size, input_size],
interpolation=PIL.Image.NEAREST),
transforms.ToTensor(),
]))
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size / 2,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
return train_loader, val_loader
# Initialize the model for this run
model_ft, input_size = initialize_model(model_name,
num_classes,
feature_extract,
use_pretrained=True)
# Print the model we just instantiated
print(model_ft)
# -------------------------- LOADING THE DATA --------------------------
# Data augmentation and normalization for training
# Just normalization for validation
data_transforms = {
'train':
transforms.Compose([
transforms.RandomResizedCrop(input_size),
transforms.RandomHorizontalFlip(),
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])
]),
}
print("Initializing Datasets and Dataloaders...")
import torch
import torch.nn.functional as F
from torch import nn
from torch import optim
from torchvision import datasets, transforms, models
from collections import OrderedDict
import json
data_dir = 'flowers'
train_dir = data_dir + '/train'
valid_dir = data_dir + '/valid'
test_dir = data_dir + '/test'
device = "cuda"
transform_train = transforms.Compose([transforms.RandomResizedCrop(224),
transforms.RandomRotation(30),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
transform_test = transforms.Compose([transforms.Resize(256),
transforms.RandomResizedCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
transform_valid = transform_test
trainset = datasets.ImageFolder(train_dir, transform = transform_train)
testset = datasets.ImageFolder(test_dir, transform = transform_test)
validset = datasets.ImageFolder(valid_dir, transform = transform_valid)
def main_worker(gpu, ngpus_per_node, args):
args.gpu = gpu
# suppress printing if not master
if args.multiprocessing_distributed and args.gpu != 0:
def print_pass(*args):
pass
builtins.print = print_pass
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)
logger = setup_logger(output='output/{}/'.format(args.output_dir), distributed_rank=dist.get_rank(), name="moco")
# create model
logger.info("=> creating model '{}'".format(args.arch))
model = moco.builder.MoCo(
models.__dict__[args.arch],
args.moco_dim, args.moco_k, args.moco_m, args.moco_t, args.mlp)
logger.info(model)
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 - 1) / ngpus_per_node)
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
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)
# comment out the following line for debugging
raise NotImplementedError("Only DistributedDataParallel is supported.")
else:
# AllGather implementation (batch shuffle, queue update, etc.) in
# this code only supports DistributedDataParallel.
raise NotImplementedError("Only DistributedDataParallel is supported.")
# 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):
logger.info("=> loading checkpoint '{}'".format(args.resume))
if args.gpu is None:
checkpoint = torch.load(args.resume)
else:
# Map model to be loaded to specified single gpu.
loc = 'cuda:{}'.format(args.gpu)
checkpoint = torch.load(args.resume, map_location=loc)
args.start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
logger.info("=> loaded checkpoint '{}' (epoch {})"
.format(args.resume, checkpoint['epoch']))
else:
logger.info("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# Data loading code
traindir = os.path.join(args.data, 'train')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if args.aug_plus:
# MoCo v2's aug: similar to SimCLR https://arxiv.org/abs/2002.05709
augmentation = [
transforms.RandomResizedCrop(224, scale=(0.2, 1.)),
transforms.RandomApply([
transforms.ColorJitter(0.4, 0.4, 0.4, 0.1) # not strengthened
], p=0.8),
transforms.RandomGrayscale(p=0.2),
transforms.RandomApply([moco.loader.GaussianBlur([.1, 2.])], p=0.5),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize
]
else:
# MoCo v1's aug: the same as InstDisc https://arxiv.org/abs/1805.01978
augmentation = [
transforms.RandomResizedCrop(224, scale=(0.2, 1.)),
transforms.RandomGrayscale(p=0.2),
transforms.ColorJitter(0.4, 0.4, 0.4, 0.4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize
]
train_dataset = datasets.ImageFolder(
traindir,
moco.loader.TwoCropsTransform(transforms.Compose(augmentation)))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
else:
train_sampler = 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, drop_last=True)
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, logger, args)
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(),
'optimizer' : optimizer.state_dict(),
}, is_best=False, filename='output/{}/checkpoint_current.pth.tar'.format(args.output_dir))
if epoch % args.save_freq == 0:
save_checkpoint({
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'optimizer' : optimizer.state_dict(),
}, is_best=False, filename='output/{}/checkpoint_{:04d}.pth.tar'.format(args.output_dir,epoch))
def make_dataset():
# Small noise is added, following SN-GAN
def noise(x):
return x + torch.FloatTensor(x.size()).uniform_(0, 1.0 / 128)
if opt.dataset == "cifar10":
trans = tfs.Compose([
tfs.RandomCrop(opt.img_width, padding=4),
tfs.RandomHorizontalFlip(),
tfs.ToTensor(),
tfs.Normalize(mean=[.5, .5, .5], std=[.5, .5, .5]),
tfs.Lambda(noise)
])
data = CIFAR10(root=opt.root,
train=True,
download=True,
transform=trans)
loader = DataLoader(data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.workers)
elif opt.dataset == "dog_and_cat_64":
trans = tfs.Compose([
tfs.RandomResizedCrop(opt.img_width,
scale=(0.8, 0.9),
ratio=(1.0, 1.0)),
tfs.RandomHorizontalFlip(),
tfs.ToTensor(),
tfs.Normalize(mean=[.5, .5, .5], std=[.5, .5, .5]),
tfs.Lambda(noise)
])
data = ImageFolder(opt.root, transform=trans)
loader = DataLoader(data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.workers)
elif opt.dataset == "dog_and_cat_128":
trans = tfs.Compose([
tfs.RandomResizedCrop(opt.img_width,
scale=(0.8, 0.9),
ratio=(1.0, 1.0)),
tfs.RandomHorizontalFlip(),
tfs.ToTensor(),
tfs.Normalize(mean=[.5, .5, .5], std=[.5, .5, .5]),
tfs.Lambda(noise)
])
data = ImageFolder(opt.root, transform=trans)
loader = DataLoader(data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.workers)
elif opt.dataset == "imagenet":
trans = tfs.Compose([
tfs.RandomResizedCrop(opt.img_width,
scale=(0.8, 0.9),
ratio=(1.0, 1.0)),
tfs.RandomHorizontalFlip(),
tfs.ToTensor(),
tfs.Normalize(mean=[.5, .5, .5], std=[.5, .5, .5]),
tfs.Lambda(noise)
])
data = ImageFolder(opt.root, transform=trans)
loader = DataLoader(data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.workers)
else:
raise ValueError(f"Unknown dataset: {opt.dataset}")
return loader
SIZE = 224
transform_test = transforms.Compose([
transforms.Resize(SIZE),
transforms.CenterCrop(SIZE),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
transform_train = transforms.Compose([
transforms.ColorJitter(brightness=0.4,
contrast=0.2,
saturation=0.2,
hue=0.2),
transforms.RandomRotation(15),
transforms.RandomResizedCrop(SIZE, scale=(0.8, 1.0), ratio=(0.90, 1.10)),
transform_test,
])
def imshow(img):
img = img / 2 + 0.5
img = img.numpy()
plt.imshow(img.transpose((1, 2, 0)))
plt.show()
def imload(path):
return Image.open(path).convert("RGB")
def main():
ia.seed(1)
train_datapath = "./food11re/skewed_training"
valid_datapath = "./food11re/validation"
test_datapath = "./food11re/evaluation"
transform = transforms.Compose([
transforms.RandomRotation(30),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
ImgAugTransform(), lambda x: PIL.Image.fromarray(x),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
train_dataset = Food11Dataset(train_datapath, is_train=True)
train_dataset_folder = torchvision.datasets.ImageFolder(
root='./food11re/skewed_training', transform=transform)
valid_dataset = Food11Dataset(valid_datapath, is_train=False)
test_dataset = Food11Dataset(test_datapath, is_train=False)
#wts = [100, 781, 67, 169, 196, 75, 757, 1190, 194, 67, 2857]
#train_dataset.augmentation(wts)
weight = []
for i in range(11):
class_count = train_dataset_folder.targets.count(i)
weight.append(1. / (class_count / len(train_dataset_folder.targets)))
samples_weight = np.array([weight[t] for _, t in train_dataset_folder])
weighted_sampler = data.WeightedRandomSampler(samples_weight,
num_samples=15000,
replacement=True)
randon_sampler = data.RandomSampler(train_dataset,
replacement=True,
num_samples=9000,
generator=None)
print(
"----------------------------------------------------------------------------------"
)
print("Dataset bf. loading - ", train_datapath)
print(train_dataset.show_details())
print(
"----------------------------------------------------------------------------------"
)
print("Dataset bf. loading - ", valid_datapath)
print(valid_dataset.show_details())
print(
"----------------------------------------------------------------------------------"
)
print("Dataset bf. loading - ", test_datapath)
print(test_dataset.show_details())
train_folder_loader = DataLoader(dataset=train_dataset_folder,
num_workers=0,
batch_size=100,
sampler=weighted_sampler)
train_loader = DataLoader(dataset=train_dataset,
num_workers=0,
batch_size=100,
sampler=randon_sampler)
valid_loader = DataLoader(dataset=valid_dataset,
num_workers=0,
batch_size=100,
shuffle=False)
test_loader = DataLoader(dataset=test_dataset,
num_workers=0,
batch_size=100,
shuffle=False)
data_loading(train_folder_loader, train_dataset)
data_loading(train_loader, train_dataset)
data_loading(valid_loader, valid_dataset)
data_loading(test_loader, test_dataset)
import torch.optim as optim
import pickle
# In[73]:
num_classes = 120
batch_size = 13
epochs = 15
sample_submission = pd.read_csv('./data/sample_submission.csv')
# In[74]:
normalize = transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
train_transform = transforms.Compose([
transforms.RandomResizedCrop(299),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(), normalize
])
valid_transform = transforms.Compose([
transforms.Resize(360),
transforms.CenterCrop(299),
transforms.ToTensor(), normalize
])
trainset = torchvision.datasets.ImageFolder("./data/train/", train_transform)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=2)
