def main():
# Data loading
train_dataset = CustomDataset(root=config.root_train,
annFile=config.annFile_train,
transforms=config.train_transforms,
catagory=config.CATEGORY_FILTER)
val_dataset = CustomDataset(root=config.root_train,
annFile=config.annFile_train,
transforms=config.val_transforms,
catagory=config.CATEGORY_FILTER)
train_loader = DataLoader(dataset=train_dataset,
batch_size=16,
num_workers=4,
pin_memory=True,
shuffle=False,
drop_last=True)
val_loader = DataLoader(dataset=val_dataset,
batch_size=16,
num_workers=4,
pin_memory=True,
shuffle=False,
drop_last=True)
# Model
model = YoloV3(num_classes=config.C).to(device=config.DEVICE)
optimizer = optim.Adam(model.parameters(),
lr=config.LEARNING_RATE,
weight_decay=config.WEIGHT_DECAY)
loss_function = YoloLoss()
scalar = torch.cuda.amp.GradScaler()
# Miscellaneous
scaled_anchors = (torch.tensor(config.anchors) * torch.tensor(
config.Scale).unsqueeze(1).unsqueeze(1).repeat(1, 3, 2)).to(
config.DEVICE)
#writer = SummaryWriter()
current_time = time.time()
print("Train loader length:", len(train_loader))
# Training loop
model.train()
for cycle, (x, y) in enumerate(train_loader):
print("Current cycle:", cycle)
delta_time, current_time = time_function(current_time)
def main_worker(gpu, args):
torch.cuda.set_device(gpu)
device = torch.device('cuda')
basic_transforms = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
unlabeled_dataset = CustomDataset(root=args.data,
split='unlabeled',
transform=basic_transforms)
labeled_dataset = CustomDataset(root=args.data,
split='train',
transform=basic_transforms)
unlabeled_dataset = torch.utils.data.DataLoader(unlabeled_dataset,
batch_size=args.batch_size,
num_workers=args.workers)
labeled_dataset = torch.utils.data.DataLoader(labeled_dataset,
batch_size=args.batch_size,
num_workers=args.workers)
# load pre-trained model from checkpoint
model = ft_model()
model.load_state_dict(
torch.load(args.checkpoint_dir / args.checkpoint_file))
model.eval()
model.to(device)
unlabeled_entropy = []
labeled_entropy = []
since = time.time()
steps = 100
with torch.no_grad():
for i, batch in enumerate(unlabeled_dataset):
entropy = get_entropy(model, batch, device).tolist()
unlabeled_entropy.extend(entropy)
if i % steps == 0:
print(i, sum(entropy) / len(entropy))
for i, batch in enumerate(labeled_dataset):
entropy = get_entropy(model, batch, device).tolist()
labeled_entropy.extend(entropy)
if i % steps == 0:
print(i, sum(entropy) / len(entropy))
return unlabeled_entropy, labeled_entropy
def __init__(self, FLAGS):
self.use_cuda = FLAGS.cuda and torch.cuda.is_available()
self.gpu = FLAGS.gpu
self.train = FLAGS.train
self.batch_size = FLAGS.batch_size
self.data_path = FLAGS.data_path
self.learning_rate = FLAGS.learning_rate
self.beta1 = FLAGS.beta_1
self.beta2 = FLAGS.beta_2
self.load_path = FLAGS.load_path
self.epoch = 0
self.end_epoch = FLAGS.end_epoch
self.model = SRModel().double()
if not self.train:
if self.use_cuda:
self.model.load_state_dict(
torch.load(os.path.join('checkpoints', self.load_path)))
else:
self.model.load_state_dict(
torch.load(os.path.join('checkpoints', self.load_path),
map_location='cpu'))
if self.use_cuda:
torch.cuda.set_device(self.gpu)
self.model.cuda()
self.optim = optim.Adam(list(self.model.parameters()),
lr=self.learning_rate,
betas=(self.beta1, self.beta2))
date_time = datetime.now().strftime("%b%d_%H-%M-%S")
self.save_folder = os.path.join('checkpoints', date_time)
if (not os.path.exists(self.save_folder)) and (self.train == True):
os.makedirs(self.save_folder, exist_ok=True)
if self.train:
self.writer = SummaryWriter()
settings = ''
for arg in vars(FLAGS):
settings += str(arg) + '=' + str(getattr(FLAGS, arg)) + ' '
self.writer.add_text('Settings', settings)
print("Loading data...")
self.train_set = CustomDataset(root=self.data_path, train=True)
# self.test_set = CustomDataset(root=self.data_path,train=False)
self.loader = cycle(
DataLoader(self.train_set,
batch_size=self.batch_size,
shuffle=True,
num_workers=0,
drop_last=True))
def main():
train_dataset = CustomDataset(root=config.root_train, annFile=config.annFile_train, transforms=config.transforms, catagory=config.CATEGORY_FILTER)
val_dataset = CustomDataset(root=config.root_train, annFile=config.annFile_train, transforms=config.transforms, catagory=config.CATEGORY_FILTER)
train_loader = DataLoader(dataset=train_dataset, batch_size=16, num_workers=4, pin_memory=True, shuffle=False, drop_last=True)
val_loader = DataLoader(dataset=val_dataset, batch_size=16, num_workers=4, pin_memory=True, shuffle=False, drop_last=True)
for index, (x, y) in enumerate(train_loader):
x = draw_y_on_x(x, y)
grid = torchvision.utils.make_grid(x, nrow=4)
# Save batch grid as image
image_dir = "./batch_dir"
image_dir_exists = os.path.exists(image_dir)
if not image_dir_exists:
os.makedirs(image_dir)
img_name = str(image_dir) + "/batch_" + str(index) + ".png"
save_image(grid.float() / 255, img_name)
print(index)
print(x.shape)
print(y.shape)
for index, (x, y) in enumerate(val_loader):
x = draw_y_on_x(x, y)
grid = torchvision.utils.make_grid(x, nrow=4)
# Save batch grid as image
image_dir = "./batch_dir"
image_dir_exists = os.path.exists(image_dir)
if not image_dir_exists:
os.makedirs(image_dir)
img_name = str(image_dir) + "/batch_" + str(index) + ".png"
save_image(grid.float() / 255, img_name)
print(index)
print(x.shape)
print(y.shape)
def main():
# Data loading
# train_dataset = CustomDataset(root=config.root_train, annFile=config.annFile_train, transforms=config.train_transforms, catagory=config.CATEGORY_FILTER)
val_dataset = CustomDataset(root=config.root_val,
annFile=config.annFile_val,
transforms=config.val_transforms,
catagory=config.CATEGORY_FILTER)
# train_loader = DataLoader(dataset=train_dataset, batch_size=16, num_workers=2, pin_memory=True, shuffle=True, drop_last=True)
val_loader = DataLoader(dataset=val_dataset,
batch_size=16,
num_workers=2,
pin_memory=True,
shuffle=False,
drop_last=True)
# Model
model = YoloV3(num_classes=config.C).to(device=config.DEVICE)
optimizer = optim.Adam(model.parameters(),
lr=config.LEARNING_RATE,
weight_decay=config.WEIGHT_DECAY)
# Miscellaneous
scaled_anchors = (torch.tensor(config.anchors) * torch.tensor(
config.Scale).unsqueeze(1).unsqueeze(1).repeat(1, 3, 2)).to(
config.DEVICE)
# Loading previously saved model weights
load_checkpoint("res50_35k.pth.tar", model, optimizer,
config.LEARNING_RATE)
# Rendering loop
model.eval()
for cycle, (x, y) in enumerate(val_loader):
with torch.no_grad():
x_gpu = x.to(config.DEVICE)
yp = model(x_gpu)
yp = [yp[0].to('cpu'), yp[1].to('cpu'), yp[2].to('cpu')]
x = denormalize(x) * 255
draw_y_on_x(x, y)
draw_yp_on_x(x, yp, probability_threshold=0.5, anchors=config.anchors)
# Save batch grid as image
image_dir = "./batch_dir"
image_dir_exists = os.path.exists(image_dir)
if not image_dir_exists:
os.makedirs(image_dir)
img_name = str(image_dir) + "/batch_" + str(cycle) + ".png"
save_image(x / 255, img_name)
def __init__(self, args):
self.args = args
# data
dataset = CustomDataset(args)
self.collate_fn = dataset.collate_fn # For zero-padding
# For K-fold
train_size = int(len(dataset) / args.cv_num)
# Randomly split a dataset into non-overlapping new datasets of given lengths
self.dataset_list = random_split(dataset,
[train_size for i in range(args.cv_num -1)] +\
[len(dataset) - (args.cv_num - 1)*train_size]) # 각 데이터 size가 들어있는 리스트
# arguments, loss
self.vocab_size = len(dataset.vocab) ## ??
self.pad_idx = dataset.vocab.word2idx['<pad>'] ## ??
self.embeddings = dataset.pretrained_embeddings
self.criterion = nn.BCEWithLogitsLoss().to(device)
# make directory if not exist data path
if not osp.isdir(args.ck_path):
os.makedirs(args.ck_path, exist_ok=True
) # If exist_ok is False, FileExistsError is raised.
def main_worker(gpu, args):
torch.cuda.set_device(gpu)
device = torch.device('cuda')
train_transforms = transforms.Compose([
transforms.RandomResizedCrop(args.image_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
validation_transforms = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
train_dataset = CustomDataset(root=args.data,
split='train',
transform=train_transforms)
validation_dataset = CustomDataset(root=args.data,
split='val',
transform=validation_transforms)
train_dataloader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers)
validation_dataloader = torch.utils.data.DataLoader(
validation_dataset,
batch_size=args.batch_size,
num_workers=args.workers)
# load pre-trained model from checkpoint
model = ft_model(args.pretrained_algo, args.model_name,
args.pretrained_dir_file, args.finetuning,
args.num_classes)
model.train()
model.to(device)
criterion = nn.CrossEntropyLoss()
criterion.to(device)
# optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
optimizer = torch.optim.SGD(model.parameters(),
lr=args.learning_rate,
momentum=0.9)
args.checkpoint_dir.mkdir(parents=True, exist_ok=True)
stats_file = open(args.checkpoint_dir /
'{}_stats.txt'.format(args.checkpoint_file[:-4]),
'a',
buffering=1)
print(' '.join(sys.argv))
print(' '.join(sys.argv), file=stats_file)
best_validation_accuracy = 0
since = time.time()
for i in range(args.epochs):
total_train_loss = 0.0
total_train_correct = 0.0
total_validation_loss = 0.0
total_validation_correct = 0.0
model.train()
for batch in train_dataloader:
loss, correct = get_loss_and_correct(model, batch, criterion,
device)
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_train_loss += loss.item()
total_train_correct += correct.item()
with torch.no_grad():
for batch in validation_dataloader:
loss, correct = get_loss_and_correct(model, batch, criterion,
device)
total_validation_loss += loss.item()
total_validation_correct += correct.item()
mean_train_loss = total_train_loss / len(train_dataset)
train_accuracy = total_train_correct / len(train_dataset)
mean_validation_loss = total_validation_loss / len(validation_dataset)
validation_accuracy = total_validation_correct / len(
validation_dataset)
# save the best model
if validation_accuracy > best_validation_accuracy:
best_validation_accuracy = validation_accuracy
torch.save(model.state_dict(),
args.checkpoint_dir / args.checkpoint_file)
time_elapsed = time.time() - since
print(
'Epoch: {}, Train Loss: {:.4f}, Val Loss: {:.4f}, Train Acc: {:.4f}, Val Acc: {:.4f}, Time: {}'
.format(i, mean_train_loss, mean_validation_loss, train_accuracy,
validation_accuracy, time_elapsed))
print(
'Epoch: {}, Train Loss: {:.4f}, Val Loss: {:.4f}, Train Acc: {:.4f}, Val Acc: {:.4f}, Time: {}'
.format(i, mean_train_loss, mean_validation_loss, train_accuracy,
validation_accuracy, time_elapsed),
file=stats_file)
def train(img_dir,classes_csv,model_fname=None,resnet_depth=50,epochs=1000,steps=100,train_split=0.8,out_dir ='',out_prefix=''):
if not os.path.exists(out_dir):
os.makedirs(out_dir)
# Create the data loaders
# Get all image fnames in folder
img_list = []
if not isinstance(img_dir, list):
img_dir = [img_dir]
for dir in img_dir:
for file in os.listdir(dir):
if file.endswith(".png"):
img_list.append(dir + file)
randomised_list = random.sample(img_list, len(img_list))
num_train = int(0.8*len(img_list))
train_imgs, val_imgs = randomised_list[:num_train], randomised_list[num_train:]
dataset_train = CustomDataset(img_list=train_imgs, class_list=classes_csv, transform=transforms.Compose([Normalizer(), Augmenter(), Resizer()]))
dataset_val = CustomDataset(img_list=val_imgs, class_list=classes_csv,transform=transforms.Compose([Normalizer(), Resizer()]))
sampler = AspectRatioBasedSampler(dataset_train, batch_size=2, drop_last=False)
dataloader_train = DataLoader(dataset_train, num_workers=3, collate_fn=collater, batch_sampler=sampler)
if dataset_val is not None:
sampler_val = AspectRatioBasedSampler(dataset_val, batch_size=1, drop_last=False)
dataloader_val = DataLoader(dataset_val, num_workers=3, collate_fn=collater, batch_sampler=sampler_val)
# Create the model
if resnet_depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(), pretrained=True)
elif resnet_depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(), pretrained=True)
elif resnet_depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(), pretrained=True)
elif resnet_depth == 101:
retinanet = model.resnet101(num_classes=dataset_train.num_classes(), pretrained=True)
elif resnet_depth == 152:
retinanet = model.resnet152(num_classes=dataset_train.num_classes(), pretrained=True)
else:
raise ValueError('Unsupported model depth, must be one of 18, 34, 50, 101, 152')
# retinanet = torch.load(model_fname)
if use_gpu:
retinanet = retinanet.cuda()
retinanet = torch.nn.DataParallel(retinanet).cuda()
if model_fname is not None:
retinanet.load_state_dict(torch.load(model_fname))
retinanet.training = True
optimizer = optim.Adam(retinanet.parameters(), lr=1e-5)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=3, verbose=True)
loss_hist = collections.deque(maxlen=500)
retinanet.train()
retinanet.module.freeze_bn()
start_time = time.clock()
print('Num training images: {}'.format(len(dataset_train)))
for epoch_num in range(epochs):
retinanet.train()
retinanet.module.freeze_bn()
epoch_loss = []
for iter_num, data in enumerate(dataloader_train):
try:
optimizer.zero_grad()
classification_loss, regression_loss = retinanet([data['img'].cuda().float(), data['annot']])
classification_loss = classification_loss.mean()
regression_loss = regression_loss.mean()
loss = classification_loss + regression_loss
if bool(loss == 0):
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(retinanet.parameters(), 0.1)
optimizer.step()
loss_hist.append(float(loss))
epoch_loss.append(float(loss))
# print('Epoch: {} | Iteration: {} | Classification loss: {:1.5f} | Regression loss: {:1.5f} | Running loss: {:1.5f}'.format(epoch_num, iter_num, float(classification_loss), float(regression_loss), np.mean(loss_hist)))
del classification_loss
del regression_loss
except Exception as e:
print(e)
continue
print('Epoch: {} | Running loss: {:1.5f} | Elapsed Time: {}'.format(epoch_num, np.mean(loss_hist),(time.clock() - start_time)/60))
mAP = csv_eval.evaluate(dataset_val, retinanet)
scheduler.step(np.mean(epoch_loss))
if (epoch_num) % steps == 0:
torch.save(retinanet.module, '{}{}_model_{}.pt'.format(out_dir, out_prefix, epoch_num))
torch.save(retinanet.state_dict(), '{}{}_state_{}.pt'.format(out_dir, out_prefix, epoch_num))
torch.save(retinanet, out_dir + '{}model_final.pt'.format(out_prefix))
torch.save(retinanet.state_dict(), out_dir + '{}state_final_.pt'.format(out_prefix))
parser.add_argument("--train", action='store_true', default=False)
parser.add_argument("--predict", action='store_true', default=False)
parser.add_argument("--zazu_timer", action='store_true', default=False)
parser.add_argument("--checkpoint_path",
type=str,
default='/root/ZazuML/best_checkpoint.pt')
parser.add_argument("--dataset_path", type=str, default='')
parser.add_argument("--output_path", type=str, default='')
args = parser.parse_args()
with open('configs.json', 'r') as fp:
configs = json.load(fp)
logger = init_logging(__name__)
from dataloader import CustomDataset
CustomDataset("../data/tiny_coco", 'coco')
zazu = ZaZu(configs['model_name'], configs['home_path'],
configs['annotation_type'])
if args.search:
zazu.search(configs['search_method'], configs['epochs'],
configs['max_trials'], configs['max_instances_at_once'],
configs['augmentation_search'])
if args.train:
adapter = TrialAdapter(0)
adapter.load(checkpoint_path=args.checkpoint_path)
adapter.train()
print('model checkpoint is saved to: ', adapter.checkpoint_path)
if args.predict:
predict(pred_on_path=args.dataset_path,
output_path=args.output_path,
checkpoint_path=args.checkpoint_path,
aug_transform3 = transforms.Compose([
transforms.RandomResizedCrop((96, 96),
scale=(0.08, 1.0),
ratio=(0.75, 1.3333333333333333)),
rnd_color_jitter3,
rnd_gray,
transforms.ToTensor(),
])
train_transform = transforms.Compose([
transforms.ToTensor(),
])
trainset = CustomDataset(root='/dataset',
split="train",
transform=train_transform)
augset1 = CustomDataset(root='/dataset',
split="train",
transform=aug_transform1)
augset2 = CustomDataset(root='/dataset',
split="train",
transform=aug_transform2)
augset3 = CustomDataset(root='/dataset',
split="train",
transform=aug_transform3)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=1024,
shuffle=True,
num_workers=2)
augloader1 = torch.utils.data.DataLoader(augset1,
def main_worker(gpu, args):
global best_acc1
args.gpu = gpu
if args.gpu is not None:
print("Use GPU: {} for training".format(args.gpu))
# create model
print("=> creating model '{}'".format(args.arch))
model = models.__dict__[args.arch](num_classes=800,
norm_layer=SubBatchNorm2d)
# load from pre-trained, before DistributedDataParallel constructor
if args.pretrained:
if os.path.isfile(args.pretrained):
print("=> loading freezed checkpoint '{}'".format(args.pretrained))
states = torch.load(args.pretrained)
model.load_state_dict(states)
args.start_epoch = 0
print("=> checkpoint loaded '{}'".format(args.pretrained))
else:
raise ValueError("=> no freezed checkpoint found at '{}'".format(
args.pretrained))
if args.gpu is not None:
torch.cuda.set_device(args.gpu)
model = model.cuda(args.gpu)
else:
print("=> ERROR: gpu must be assigned")
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda(args.gpu)
# optimize only the linear classifier
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
cudnn.benchmark = True
# Data loading code
traindir = args.data
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transform_train = transforms.Compose([
transforms.RandomResizedCrop(96), # add crop resize
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize
])
transform_eval = transforms.Compose([
transforms.Resize(128), # add resize
transforms.CenterCrop(96), # add crop
transforms.ToTensor(),
normalize
])
train_dataset = CustomDataset(traindir, 'train', transform_train)
eval_dataset = CustomDataset(traindir, 'val', transform_eval)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
eval_loader = torch.utils.data.DataLoader(eval_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
# training code
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch, args)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args)
if epoch == 0 or (epoch + 1) % args.eval_per_n_epoch == 0:
accuracy = evaluate(eval_loader, model, args)
print(f"=> Epoch: {epoch+1}, accuracy: {accuracy:.4f}")
# remember best acc and save checkpoint
is_best = accuracy > best_acc1
best_acc1 = max(accuracy, best_acc1)
print(f"=> Epoch: {epoch+1}, isBest? : {is_best}")
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'accuracy': accuracy,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
},
is_best,
save_dir=args.checkpoint_dir,
epoch=(epoch + 1),
filename=os.path.join(
args.checkpoint_dir,
'checkpoint_{:03d}.pth.tar'.format(epoch + 1)))
def main_worker(gpu, ngpus_per_node, args):
args.gpu = gpu
if args.gpu is not None:
print("Use GPU: {} for training".format(args.gpu))
if args.multiprocessing_distributed and args.gpu != 0:
def print_pass(*args):
pass
builtins.print = print_pass
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
print("=> creating model '{}'".format(args.arch))
if args.arch == 'resnet50':
model = Model(resnet50,args,width=1)
elif args.arch == 'resnet50x2':
model = Model(resnet50,args,width=2)
elif args.arch == 'resnet50x4':
model = Model(resnet50,args,width=4)
else:
raise NotImplementedError('model not supported {}'.format(args.arch))
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.batch_size_u = int(args.batch_size_u / 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]) #find_unused_parameters=True
else:
model.cuda()
# DistributedDataParallel will divide and allocate batch_size to all
# available GPUs if device_ids are not set
model = torch.nn.parallel.DistributedDataParallel(model)
# define loss function (criterion) and optimizer
criteria_x = nn.CrossEntropyLoss().cuda(args.gpu)
optimizer = torch.optim.SGD(model.parameters(), args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay,
nesterov=True
)
# 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']
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
print("=> preparing dataset")
# Data loading code
transform_strong = transforms.Compose([
transforms.RandomResizedCrop(96, scale=(0.2, 1.)),
transforms.RandomApply([
transforms.ColorJitter(0.4, 0.4, 0.4, 0.1)
], p=0.8),
transforms.RandomGrayscale(p=0.2),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])
])
transform_weak = transforms.Compose([
transforms.RandomResizedCrop(96, scale=(0.2, 1.)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])
])
transform_eval = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])
])
three_crops_transform = loader.ThreeCropsTransform(transform_weak, transform_strong, transform_strong)
unlabeled_dataset = CustomDataset(args.data, 'unlabeled', transform=three_crops_transform)
labeled_dataset = CustomDataset(args.data, 'train', transform=transform_weak)
#labeled_sampler = torch.utils.data.distributed.DistributedSampler(labeled_dataset)
labeled_loader = torch.utils.data.DataLoader(
labeled_dataset, batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True)
#unlabeled_sampler = torch.utils.data.distributed.DistributedSampler(unlabeled_dataset)
unlabeled_loader = torch.utils.data.DataLoader(
unlabeled_dataset, batch_size=int(args.batch_size_u), shuffle=True,
num_workers=args.workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(CustomDataset('/content/dataset', 'val', transform_eval),
batch_size=64, shuffle=False,
num_workers=args.workers, pin_memory=True)
# create loggers
if args.gpu==0:
tb_logger = tensorboard_logger.Logger(logdir=os.path.join(args.exp_dir,'tensorboard'), flush_secs=2)
logger = setup_default_logging(args)
logger.info(dict(args._get_kwargs()))
else:
tb_logger = None
logger = None
for epoch in range(args.start_epoch, args.epochs):
if epoch==0:
args.m = 0.99 # larger update in first epoch
else:
args.m = args.moco_m
# args.lr=0.01
adjust_learning_rate(optimizer, epoch, args)
train(labeled_loader, unlabeled_loader, model, criteria_x, optimizer, epoch, args, logger, tb_logger)
# evaluate on validation set
acc1 = validate(val_loader, model, args, logger, tb_logger, epoch)
if not args.multiprocessing_distributed or (args.multiprocessing_distributed
and args.rank % ngpus_per_node == 0):
save_checkpoint({
'args': args,
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'optimizer' : optimizer.state_dict()
},filename='{}/checkpoint_{:04d}.pt'.format(args.exp_dir,epoch))
# evaluate ema model
acc1 = validate(val_loader, model, args, logger, tb_logger, -1)
def main():
#TODO: Get args
# python3 train_fixmatch.py --checkpoint-path ./checkpoint_path/model.pth --batch-size 1 --num-epochs 1 --num-steps 1 --train-from-start 1 --dataset-folder ./dataset
parser = argparse.ArgumentParser()
parser.add_argument('--checkpoint-path', type=str, default= "./checkpoints/model_barlow_20h.pth.tar")
parser.add_argument('--batch-size', type=int, default= 512)
parser.add_argument('--num-epochs', type=int, default= 10)
parser.add_argument('--num-steps', type=int, default= 1)
parser.add_argument('--train-from-start', type= int, default= 0)
parser.add_argument('--dataset-folder', type= str, default= "./dataset")
parser.add_argument('--learning-rate', type = float, default= 0.01)
parser.add_argument('--threshold', type = float, default= 0.5)
parser.add_argument('--mu', type= int, default= 7)
parser.add_argument('--lambd', type= float, default= 0.005)
parser.add_argument('--momentum', type= float, default= 0.9)
parser.add_argument('--weight-decay', type= float, default= 1.5*1e-6)
parser.add_argument('--warmup-epochs', type= int, default= 2)
parser.add_argument('--scale-loss', type = float, default= 1.0/32.0)
parser.add_argument('--wide', type= int, default= 0)
args = parser.parse_args()
dataset_folder = args.dataset_folder
batch_size = args.batch_size
n_epochs = args.num_epochs
num_classes = 800
lambd = args.lambd
weight_decay = args.weight_decay
checkpoint_path = args.checkpoint_path
train_from_start = args.train_from_start
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
unlabeled_train_dataset = CustomDataset(root= dataset_folder, split = "unlabeled", transform = TransformBarlowTwins())
unlabeled_train_loader = DataLoader(unlabeled_train_dataset, batch_size= batch_size, shuffle= True, num_workers= 4)
if args.wide == 1:
model = lightly.models.BarlowTwins(wide_resnet50_2(pretrained= False), num_ftrs= 2048)
else:
model = lightly.models.BarlowTwins(resnet18(pretrained= False), num_ftrs= 512)
optimizer = LARS(model.parameters(), lr=0, weight_decay=weight_decay,
weight_decay_filter=exclude_bias_and_norm,
lars_adaptation_filter=exclude_bias_and_norm)
criterion = lightly.loss.BarlowTwinsLoss()
start_epoch = 0
model.train()
losses = Average()
model = model.to(device)
criterion = criterion.to(device)
if train_from_start == 0:
assert os.path.isfile(checkpoint_path), "Error: no checkpoint directory found!"
print("Restoring model from checkpoint")
# args.out = os.path.dirname(args.resume)
checkpoint = torch.load(checkpoint_path, map_location= device)
if args.wide == 0:
model = torch.nn.DataParallel(model)
# best_acc = checkpoint['best_acc']
start_epoch = checkpoint['epoch'] - 1
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
model = torch.nn.DataParallel(model)
criterion = torch.nn.DataParallel(criterion)
model = model.to(device)
criterion = criterion.to(device)
#TODO
# scaler = torch.cuda.amp.GradScaler()
# model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
for epoch in tqdm(range(start_epoch, n_epochs)):
# for batch_idx in tqdm(range(n_steps)): ## CHECK
for batch_idx, batch in enumerate(tqdm(unlabeled_train_loader)):
y_a = batch[0][0].to(device)
y_b = batch[0][1].to(device)
z_a, z_b = model(y_a, y_b)
loss = criterion(z_a, z_b).mean()
lr = adjust_learning_rate(args, optimizer, unlabeled_train_loader, epoch * len(unlabeled_train_loader) + batch_idx)
optimizer.zero_grad()
# scaler.scale(loss).backward()
# scaler.step(optimizer)
# scaler.update()
loss.backward()
optimizer.step()
losses.update(loss.item())
if batch_idx % 25 == 0:
print(f"Epoch number: {epoch}, loss_avg: {losses.avg}, loss: {loss.item()}, lr: {lr}", flush= True)
if torch.cuda.device_count() > 1:
save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.module.state_dict(),
'optimizer': optimizer.state_dict()
}, checkpoint_path)
else:
save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}, checkpoint_path)
from torchvision import datasets, transforms, models
from dataloader import CustomDataset
from submission import get_model
#parser part
parser = argparse.ArgumentParser()
parser.add_argument('--checkpoint-dir', type=str)
args = parser.parse_args()
#train composition operations
train_transform = transforms.Compose([
transforms.ToTensor(),
])
#trainset and trainloaders
trainset = CustomDataset(root='/dataset',
split="train",
transform=train_transform)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=256,
shuffle=True,
num_workers=2)
#net = get_model().cuda()
net = get_model()
criterion = nn.CrossEntropyLoss()
#optimizer = torch.optim.Adam(net.parameters(), lr=0.05)
print('Start Training')
#training part
def main_worker(gpu, args):
# args.rank += gpu
#
# torch.distributed.init_process_group(
# backend='nccl', init_method=args.dist_url,
# world_size=args.world_size, rank=args.rank)
# if args.rank == 0:
args.checkpoint_dir.mkdir(parents=True, exist_ok=True)
stats_file = open(args.checkpoint_dir /
'stats_{}.txt'.format(args.resnet_layers),
'a',
buffering=1)
print(' '.join(sys.argv))
print(' '.join(sys.argv), file=stats_file)
torch.cuda.set_device(gpu)
# torch.backends.cudnn.benchmark = True
model = BarlowTwins(args).cuda(gpu)
# model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
# model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[gpu])
optimizer = LARS(model.parameters(),
lr=0,
weight_decay=args.weight_decay,
weight_decay_filter=exclude_bias_and_norm,
lars_adaptation_filter=exclude_bias_and_norm)
# automatically resume from checkpoint if it exists
if (args.checkpoint_dir /
'checkpoint_{}.pth'.format(args.resnet_layers)).is_file():
ckpt = torch.load(args.checkpoint_dir /
'checkpoint_{}.pth'.format(args.resnet_layers),
map_location='cpu')
start_epoch = ckpt['epoch']
model.load_state_dict(ckpt['model'])
optimizer.load_state_dict(ckpt['optimizer'])
else:
start_epoch = 0
# dataset = torchvision.datasets.ImageFolder(args.data / 'train', Transform())
# sampler = torch.utils.data.distributed.DistributedSampler(dataset)
# assert args.batch_size % args.world_size == 0
# per_device_batch_size = args.batch_size // args.world_size
# loader = torch.utils.data.DataLoader(
# dataset, batch_size=per_device_batch_size, num_workers=args.workers,
# pin_memory=True, sampler=sampler)
dataset = CustomDataset(root=args.data,
split='unlabeled',
transform=Transform(args.image_size))
loader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
num_workers=args.workers)
start_time = time.time()
scaler = torch.cuda.amp.GradScaler()
for epoch in range(start_epoch, args.epochs):
# sampler.set_epoch(epoch)
for step, ((y1, y2), _) in enumerate(loader,
start=epoch * len(loader)):
y1 = y1.cuda(gpu, non_blocking=True)
y2 = y2.cuda(gpu, non_blocking=True)
lr = adjust_learning_rate(args, optimizer, loader, step)
optimizer.zero_grad()
with torch.cuda.amp.autocast():
loss = model.forward(y1, y2)
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
if step % args.print_freq == 0:
# torch.distributed.reduce(loss.div_(args.world_size), 0)
# if args.rank == 0:
stats = dict(epoch=epoch,
step=step,
learning_rate=lr,
loss=loss.item(),
time=int(time.time() - start_time))
print(json.dumps(stats))
print(json.dumps(stats), file=stats_file)
# if args.rank == 0:
# save checkpoint
state = dict(epoch=epoch + 1,
model=model.state_dict(),
optimizer=optimizer.state_dict())
torch.save(
state, args.checkpoint_dir /
'checkpoint_{}.pth'.format(args.resnet_layers))
# if args.rank == 0:
# save final model
torch.save(model.backbone.state_dict(),
args.checkpoint_dir / 'resnet{}.pth'.format(args.resnet_layers))
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data.sampler import SubsetRandomSampler
from torch.utils.data import DataLoader
from torch.optim.rmsprop import RMSprop
import torch.nn as nn
from tqdm import tqdm
import torch
import numpy as np
import sys
configs = Configs()
device = 'cuda' if torch.cuda.is_available() else 'cpu'
DATASET_PATH = r"D:\Code\CV_project\lstm_fast_fcn\dataset.pt"
dataset = CustomDataset(configs)
dataset_size = len(dataset)
indices = list(range(dataset_size))
np.random.seed(0)
np.random.shuffle(indices)
split = int(np.floor(configs.valSplit * dataset_size))
trainIndices, valIndices = indices[split:], indices[:split]
trainLoader = DataLoader(dataset,
batch_size=configs.batchSize,
num_workers=0,
sampler=SubsetRandomSampler(trainIndices))
valLoader = DataLoader(dataset,
batch_size=configs.batchSize,
num_workers=0,
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--checkpoint-path',
type=str,
default="./checkpoints/model_transfer.pth.tar")
parser.add_argument('--transfer-path',
type=str,
default="./checkpoints/model_barlow.pth.tar")
parser.add_argument('--best-path',
type=str,
default="./checkpoints/model_barlow_best.pth.tar")
parser.add_argument('--batch-size', type=int, default=10)
parser.add_argument('--num-epochs', type=int, default=100)
parser.add_argument('--dataset-folder', type=str, default="./dataset")
parser.add_argument('--new-dataset-folder', type=str, default="./dataset")
parser.add_argument('--learning-rate-classifier',
type=float,
default=0.001)
parser.add_argument('--learning-rate-model', type=float, default=0.001)
parser.add_argument('--momentum', type=float, default=0.9)
parser.add_argument('--weight-decay', type=float, default=0.001)
parser.add_argument('--fine-tune', type=int, default=0)
parser.add_argument('--wide', type=int, default=0)
parser.add_argument('--model-name', type=str, default="moco")
parser.add_argument('--dropout', type=float, default=0)
parser.add_argument('--new-data', type=int, default=0)
parser.add_argument('--seed', type=int, default=0)
args = parser.parse_args()
dataset_folder = args.dataset_folder
batch_size = args.batch_size
batch_size_val = 256 #5120
n_epochs = args.num_epochs
weight_decay = args.weight_decay
checkpoint_path = args.checkpoint_path
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
torch.backends.cudnn.deterministic = True
print(f"Training with seed {args.seed}")
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
train_transform, val_transform = get_transforms(
) #TODO Get new transforms file
if args.new_data == 0:
labeled_train_dataset = CustomDataset(root=args.dataset_folder,
split="train",
transform=train_transform)
else:
labeled_train_dataset = CustomDataset(root=args.new_dataset_folder,
split="train_new",
transform=train_transform)
val_dataset = CustomDataset(root=args.dataset_folder,
split="val",
transform=val_transform)
labeled_train_loader = DataLoader(labeled_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
val_loader = DataLoader(val_dataset,
batch_size=batch_size_val,
shuffle=False,
num_workers=4)
resnet = lightly.models.ResNetGenerator('resnet-18', 1, num_splits=0)
backbone = torch.nn.Sequential(
*list(resnet.children())[:-1],
torch.nn.AdaptiveAvgPool2d(1),
)
if args.model_name == "moco":
model = lightly.models.MoCo(backbone,
num_ftrs=512,
m=0.99,
batch_shuffle=True)
else:
if args.wide == 1:
model = lightly.models.BarlowTwins(
wide_resnet50_2(pretrained=False), num_ftrs=2048)
else:
model = lightly.models.BarlowTwins(resnet18(pretrained=False),
num_ftrs=512)
checkpoint = torch.load(args.transfer_path, map_location=device)
# print(checkpoint['state_dict'].keys())
# print("printed keys")
# print(model_barlow.state_dict().keys())
# print("printed model keys")
# if args.wide == 0:
# model = torch.nn.DataParallel(model)
model.load_state_dict(checkpoint['state_dict'])
# print(model_barlow)
if args.wide == 0:
model = model.backbone
else:
model = model.backbone
if args.wide == 1:
classifier = Classifier(ip=2048, dp=args.dropout)
else:
classifier = Classifier(ip=512, dp=args.dropout)
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
model = torch.nn.DataParallel(model)
classifier = torch.nn.DataParallel(classifier)
if not args.fine_tune:
model.requires_grad_(False)
model = model.to(device)
classifier = classifier.to(device)
param_groups = [
dict(params=classifier.parameters(), lr=args.learning_rate_classifier)
]
if args.fine_tune:
param_groups.append(
dict(params=model.parameters(), lr=args.learning_rate_model))
optimizer = optim.Adam(param_groups, weight_decay=weight_decay)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, n_epochs)
start_epoch = 0
losses = Average()
criterion = torch.nn.CrossEntropyLoss().to(device)
best_val_accuracy = 25.0 #TODO
for epoch in tqdm(range(start_epoch, n_epochs)):
if args.fine_tune:
model.train()
else:
model.eval()
classifier.train()
for batch_idx, batch in enumerate(tqdm(labeled_train_loader)):
img = batch[0].to(device)
labels = batch[1].to(device)
model_out = model(img)
if args.model_name == "moco":
model_out = model_out.squeeze()
model_out = torch.nn.functional.normalize(model_out, dim=1)
logits = classifier(model_out)
loss = criterion(logits, labels)
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
losses.update(loss.item())
if batch_idx % 25 == 0:
print(
f"Epoch number: {epoch}, loss_avg: {losses.avg}, loss: {loss.item()}, best accuracy: {best_val_accuracy:.2f}",
flush=True)
save_checkpoint(
{
'epoch': epoch + 1,
'model_state_dict': model.state_dict(),
'classifier_state_dict': classifier.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict()
}, checkpoint_path)
model.eval()
with torch.no_grad():
val_loss = 0
val_size = 0
total = 0
correct = 0
for batch in val_loader:
model_out = model(batch[0].to(device))
if args.model_name == "moco":
model_out = model_out.squeeze()
model_out = torch.nn.functional.normalize(model_out, dim=1)
logits_val = classifier(model_out)
labels = batch[1].to(device)
val_loss += F.cross_entropy(logits_val, labels)
_, predicted = torch.max(logits_val.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
val_size += 1
# break
print(
f"Val loss: {val_loss/val_size}, Accuracy: {(100 * correct / total):.2f}%",
flush=True)
if 100 * correct / total > best_val_accuracy:
best_val_accuracy = 100 * correct / total
best_val_loss = val_loss / val_size
print(
f"Saving the best model with {best_val_accuracy:.2f}% accuracy and {best_val_loss:.2f} loss",
flush=True)
save_checkpoint(
{
'epoch': epoch + 1,
'model_state_dict': model.state_dict(),
'classifier_state_dict': classifier.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'best_val_accuracy': best_val_accuracy,
'best_val_loss': best_val_loss
}, args.best_path)
def main_worker(gpu, ngpus_per_node, args):
args.gpu = gpu
if args.gpu is not None:
print("=> Use GPU: {} for training".format(args.gpu))
print("=> creating model '{}'".format(args.arch))
model = MoCo(models.__dict__[args.arch], args.moco_dim, args.moco_k,
args.moco_m, args.moco_t, args.mlp)
print(model)
if args.gpu is not None:
torch.cuda.set_device(args.gpu)
model = model.cuda(args.gpu)
else:
raise NotImplementedError("Only Single GPU 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):
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']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
del checkpoint # release GPU memory
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# Data loading code
traindir = args.data
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(
96, scale=(0.2, 1.)), # Add back RandomResizedCrop
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(),
normalize
]
else:
# MoCo v1's aug: the same as InstDisc https://arxiv.org/abs/1805.01978
augmentation = [
transforms.RandomResizedCrop(96, 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 = CustomDataset(
traindir, "unlabeled",
TwoCropsTransform(transforms.Compose(augmentation)))
if args.small_set:
print('=> Using 1/10 unlabeled set')
train_sampler = torch.utils.data.RandomSampler(train_dataset,
replacement=True,
num_samples=51200)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler)
else:
print('=> Using full unlabeled set')
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True,
drop_last=True)
print("=> Start Training.")
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch, args)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args)
if epoch == 0 or (epoch + 1) % args.save_checkpoint_per_epoch == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
},
is_best=False,
filename=os.path.join(
args.checkpoint_dir,
'checkpoint_{:03d}.pth'.format(epoch + 1)))
def main():
# Data loading
train_dataset = CustomDataset(root=config.root_train,
annFile=config.annFile_train,
transforms=config.train_transforms,
catagory=config.CATEGORY_FILTER)
val_dataset = CustomDataset(root=config.root_val,
annFile=config.annFile_val,
transforms=config.val_transforms,
catagory=config.CATEGORY_FILTER)
train_loader = DataLoader(dataset=train_dataset,
batch_size=8,
num_workers=2,
pin_memory=True,
shuffle=True,
drop_last=True)
val_loader = DataLoader(dataset=val_dataset,
batch_size=8,
num_workers=2,
pin_memory=True,
shuffle=False,
drop_last=True)
# Model
model = YoloV3(num_classes=config.C).to(device=config.DEVICE)
# from model_external import YOLOv3
# model = YOLOv3(num_classes=90).to(device=config.DEVICE)
optimizer = optim.Adam(model.parameters(),
lr=config.LEARNING_RATE,
weight_decay=config.WEIGHT_DECAY)
from loss_external import YoloLoss
loss_function = YoloLoss().to(device=config.DEVICE)
# Miscellaneous
scaled_anchors = (torch.tensor(config.anchors) * torch.tensor(
config.Scale).unsqueeze(1).unsqueeze(1).repeat(1, 3, 2)).to(
config.DEVICE)
writer = SummaryWriter()
current_time = time.time()
# Loading previously saved model weights
if config.LOAD_MODEL:
load_checkpoint("cp.pth.tar", model, optimizer, config.LEARNING_RATE)
print(torch.cuda.memory_summary(device=None, abbreviated=False))
# Training loop
for cycle in range(config.CYCLES):
print("Cycle:", cycle)
x, y = next(iter(val_loader))
x = x.to(config.DEVICE)
y0, y1, y2 = (y[0].to(config.DEVICE), y[1].to(config.DEVICE),
y[2].to(config.DEVICE))
yp = model(x)
loss_0 = loss_function(predictions=yp[0],
target=y0,
anchors=scaled_anchors[0])
loss_1 = loss_function(predictions=yp[1],
target=y1,
anchors=scaled_anchors[1])
loss_2 = loss_function(predictions=yp[2],
target=y2,
anchors=scaled_anchors[2])
loss = loss_0 + loss_1 + loss_2
loss.backward()
optimizer.step()
optimizer.zero_grad()
# Run validation
if cycle % 100 == 0 and cycle != 0:
model.eval()
losses = []
with torch.no_grad():
x, y = next(iter(val_loader))
x = x.to(config.DEVICE)
y0, y1, y2 = (y[0].to(config.DEVICE), y[1].to(config.DEVICE),
y[2].to(config.DEVICE))
yp = model(x)
loss_0 = loss_function(predictions=yp[0],
target=y0,
anchors=scaled_anchors[0])
loss_1 = loss_function(predictions=yp[1],
target=y1,
anchors=scaled_anchors[1])
loss_2 = loss_function(predictions=yp[2],
target=y2,
anchors=scaled_anchors[2])
loss = loss_0 + loss_1 + loss_2
losses.append(loss)
avg_val_loss = sum(losses) / len(losses)
writer.add_scalar("val_loss: ", avg_val_loss, cycle)
model.train()
# Run validation
"""
if cycle % 100 == 0 and cycle != 0:
model.eval()
x, y = next(iter(val_loader))
x = x.float()
x = x.to(config.DEVICE)
# y0, y1, y2 = (y[0].to(config.DEVICE), y[1].to(config.DEVICE), y[2].to(config.DEVICE))
with torch.no_grad():
yp = model(x)
# Move predictions to cpu
yp = [yp[0].to('cpu'), yp[1].to('cpu'), yp[2].to('cpu')]
# boxes_from_yp(yp) returns all yp bboxes in a batch
yp_boxes = boxes_from_yp(yp=yp, iou_threshold=config.MAP_IOU_THRESH, threshold=config.CONF_THRESHOLD)
# boxes_from_y(y) returns all y bboxes in a batch
y_boxes = boxes_from_y(y=y)
"""
# Save model
if cycle % 1000 == 0 and cycle != 0:
save_checkpoint(model,
optimizer,
cycle,
filename=config.CHECKPOINT_FILE)
# Rendering loop
if cycle % 100 == 0 and cycle != 0:
model.eval()
x, y = next(iter(val_loader))
with torch.no_grad():
x_gpu = x.to(config.DEVICE)
yp = model(x_gpu)
yp = [yp[0].to('cpu'), yp[1].to('cpu'), yp[2].to('cpu')]
x = denormalize(x) * 255
draw_y_on_x(x, y)
draw_yp_on_x(x,
yp,
probability_threshold=0.5,
anchors=config.anchors)
# Save batch grid as image
image_dir = "./batch_dir"
image_dir_exists = os.path.exists(image_dir)
if not image_dir_exists:
os.makedirs(image_dir)
img_name = str(image_dir) + "/batch_" + str(cycle) + ".png"
save_image(x / 255, img_name)
model.train()
writer.add_scalar("train_loss: ", loss.item(), cycle)
delta_time, current_time = time_function(current_time)
writer.add_scalar("Epoch Duration [s]", delta_time, cycle)
writer.flush()
parser = argparse.ArgumentParser()
parser.add_argument('--checkpoint-dir', type=str, default='checkpoints/')
parser.add_argument('--model-name', type=str, default='simsiam')
parser.add_argument('--epochs', type=int, default=1)
parser.add_argument('--net_size', type=int, default=50)
parser.add_argument('--temperature', type=int, default=1)
args = parser.parse_args()
checkpoint_path = args.checkpoint_dir + args.model_name
# sys.path.insert(1, args.checkpoint_dir)
# PATH = '/Users/colinwan/Desktop/NYU_MSDS/2572/FinalProject/DL21SP20'
PATH = ''
train_dataset = CustomDataset(root=PATH+'/dataset', split='unlabeled', transform=get_aug(train=True, image_size=96))
BATCH_SIZE = 256
print(len(train_dataset))
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=1)
if torch.cuda.is_available():
device = torch.device("cuda:0")
else:
device = torch.device("cpu")
model = SimSiam().to(device)
check = os.path.exists(
os.path.join(checkpoint_path,
args.model_name+"_encoder_{}.pth".format(args.net_size)))
print(os.path.join(checkpoint_path,
args.model_name+"_encoder_{}.pth".format(args.net_size)))
from dataloader import CustomDataset
from model.seq2seq import Seq2Seq
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if torch.cuda.is_available():
with_cuda = True
device_ids = None
print("------- GPU Working -------")
print("[Current GPU]:" + str(torch.cuda.get_device_name(0)))
else:
with_cuda = False
device_ids = None
print("------- CPU Working -------")
data_loaded = CustomDataset(path='data/eng-fra.txt')
pad_idx = data_loaded.vocab_stoi['<pad>']
hidden_size = 1000
vocab_len = len(data_loaded.vocab_stoi)
embedding_size = 620
batch_size = 80
train_loader = torchdata.DataLoader(dataset=data_loaded,
collate_fn=data_loaded.custom_collate_fn,
batch_size=batch_size)
trg_max_seq_len = next(
iter(train_loader))[1].size(1) - 1 # <s> is not included
epochs = 1
class fullmodel(nn.Module):
def __init__(self, num_classes=800):
super(fullmodel, self).__init__()
self.pretrain = resnet50()
self.pretrain.fc = nn.Linear(self.pretrain.fc.in_features, num_classes)
self.relu = nn.ReLU()
self.linear = nn.Linear(num_classes, num_classes)
def forward(self, x):
x = self.relu(self.pretrain(x))
outputs = self.linear(x)
return outputs
unlabeled_dataset = CustomDataset(PATH + '/dataset',
'unlabeled',
transform=transform_unlabled)
unlabeled_trainloader = torch.utils.data.DataLoader(unlabeled_dataset,
batch_size=BATCH_SIZE * mu,
shuffle=True,
num_workers=1)
labeled_dataset = CustomDataset(PATH + '/dataset',
'train',
transform=transform_labeled)
labeled_trainloader = torch.utils.data.DataLoader(labeled_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=1)
validation_dataset = CustomDataset(PATH + '/dataset', 'val',
transforms.ColorJitter(hue=.1, saturation=.1, contrast=.1),
transforms.RandomRotation(20, resample=Image.BILINEAR),
# transforms.GaussianBlur(7, sigma=(0.1, 1.0)),
transforms.ToTensor(), # convert PIL to Pytorch Tensor
normalize,
])
validation_transforms = transforms.Compose([
transforms.Resize((96, 96)),
transforms.ToTensor(),
normalize,
])
# path = '/Users/colinwan/Desktop/NYU_MSDS/2572/FinalProject/DL21SP20'
path = ''
train_dataset = CustomDataset(root=path + '/dataset',
split='train',
transform=train_transforms)
validation_dataset = CustomDataset(root=path + '/dataset',
split='val',
transform=validation_transforms)
BATCH_SIZE = 512
train_dataloader = torch.utils.data.DataLoader(train_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=1)
validation_dataloader = torch.utils.data.DataLoader(validation_dataset,
batch_size=BATCH_SIZE,
num_workers=1)
# from tqdm.notebook import tqdm
import os
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import datasets, transforms, models
from dataloader import CustomDataset
from submission import get_model, eval_transform, team_id, team_name, email_address
parser = argparse.ArgumentParser()
parser.add_argument('--checkpoint-path', type=str)
args = parser.parse_args()
evalset = CustomDataset(root='/dataset', split="val", transform=eval_transform)
evalloader = torch.utils.data.DataLoader(evalset, batch_size=256, shuffle=False, num_workers=2)
net = get_model()
checkpoint = torch.load(args.checkpoint_path)
net.load_state_dict(checkpoint)
net = net.cuda()
net.eval()
correct = 0
total = 0
with torch.no_grad():
for data in evalloader:
images, labels = data
images = images.cuda()
#################################################################3
def addIndexToTrainData(trainset):
returnList = []
for iC, (x, y) in enumerate(trainset):
if iC % 1000 == 0:
print(">>>>>>>>", iC)
returnList.append((x, y, iC))
return returnList
# ==========================================================================
# ==========================================================================
from submission import get_model, eval_transform, team_id, team_name, email_address
trainset = CustomDataset(root='./dataset',
split="train",
transform=train_transform)
# trainset = addIndexToTrainData(trainset)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=256,
shuffle=True,
num_workers=2)
net = get_model().cuda()
net = torch.nn.DataParallel(net)
net = net.cuda()
# trainLabeledImage(net, trainloader)
#
unLabeledSet = CustomDataset(root='./dataset',
split="unlabeled",
transform=train_transform)
def main():
# t_img = cv2.imread(f'{root_path}/imgs/AFW_1051618982_1_0.jpg')
lr = 0.001
models = [MobileNetV1]
model_name = ['mobilev1']
for init_model in models:
img_size = [(224, 224, 3), (224, 224, 1), (256, 256, 3), (256, 256, 1)]
m_name = model_name[0]
print(f'Train Model {m_name}')
for i_size in img_size:
print(f'Train Image size {i_size}')
dataset = CustomDataset(root_path, anno_name, dataset_type, i_size)
print(f"Train dataset size {len(dataset)}")
train_loader = DataLoader(dataset,
batch_size=8,
num_workers=0,
shuffle=True)
model = init_model(i_size[0], i_size[-1], num_landmark)
model.to(DEVICE)
criterion = torch.nn.MSELoss().to(DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, 'min')
pre_train = False
model_path = 'E:/models'
if pre_train:
saved_state = torch.load(
f'{model_path}/f_landmark-tcdcn-52-0.0004.pth')
model.load_state_dict(saved_state['model_state_dict'])
optimizer.load_state_dict(saved_state['optimizer_state_dict'])
init_epoch = saved_state['Epoch']
min_loss = saved_state['loss']
else:
init_epoch = 0
min_loss = 1
epochs = 10000
print(f'min loss : {min_loss}')
for epoch in range(init_epoch, epochs):
print(
f'{epoch} epoch start! : {datetime.datetime.now().strftime("%Y.%m.%d %H:%M:%S")}'
)
loss = train(train_loader, model, criterion, optimizer, DEVICE,
scheduler)
print(f" Average Loss : {loss:.6f}")
if min_loss > loss:
min_loss = loss
state = {
'Epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss
}
model_path = os.path.join(
model_save_path, f'f_landmark-{m_name}-{i_size}.pth')
torch.save(state, model_path)
print(
f'Saved model_{m_name} _ [loss : {loss:.6f}, save_path : {model_path}\n'
)
if loss < 0.000001:
break
def main():
#TODO: Get args
# python3 train_fixmatch.py --checkpoint-path ./checkpoint_path/model.pth --batch-size 1 --num-epochs 1 --num-steps 1 --train-from-start 1 --dataset-folder ./dataset
parser = argparse.ArgumentParser()
parser.add_argument('--checkpoint-path', type=str, default= "./checkpoints/model_fm_transfer.pth.tar")
parser.add_argument('--transfer-path', type=str, default= "./checkpoints/model_transfer.pth.tar")
parser.add_argument('--best-path', type= str, default= "./checkpoints/model_barlow_best.pth.tar")
parser.add_argument('--batch-size', type=int, default= 64)
parser.add_argument('--num-epochs', type=int, default= 10)
parser.add_argument('--num-steps', type=int, default= 10)
parser.add_argument('--train-from-start', type= int, default= 1)
parser.add_argument('--dataset-folder', type= str, default= "./dataset")
parser.add_argument('--new-dataset-folder', type= str, default= "./dataset")
parser.add_argument('--learning-rate', type = float, default= 0.01)
parser.add_argument('--threshold', type = float, default= 0.5)
parser.add_argument('--mu', type= int, default= 7)
parser.add_argument('--lambd', type= int, default= 1)
parser.add_argument('--momentum', type= float, default= 0.9)
parser.add_argument('--weight-decay', type= float, default= 0.001)
parser.add_argument('--layers', type= int, default= 18)
parser.add_argument('--fine-tune', type= int, default= 1)
parser.add_argument('--new-data', type= int, default= 0)
args = parser.parse_args()
dataset_folder = args.dataset_folder
batch_size_labeled = args.batch_size
mu = args.mu
batch_size_unlabeled = mu * args.batch_size
batch_size_val = 256 #5120
n_epochs = args.num_epochs
n_steps = args.num_steps
num_classes = 800
threshold = args.threshold
learning_rate = args.learning_rate
momentum = args.momentum
lamd = args.lambd
tau = 0.95
weight_decay = args.weight_decay
checkpoint_path = args.checkpoint_path
train_from_start = args.train_from_start
n_layers = args.layers
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
# print("pwd: ", os.getcwd())
train_transform, val_transform = get_transforms()
if args.new_data == 0:
labeled_train_dataset = CustomDataset(root= args.dataset_folder, split = "train", transform = train_transform)
else:
labeled_train_dataset = CustomDataset(root= args.new_dataset_folder, split = "train_new", transform = train_transform)
# labeled_train_dataset = CustomDataset(root= dataset_folder, split = "train", transform = train_transform)
unlabeled_train_dataset = CustomDataset(root= dataset_folder,
split = "unlabeled",
transform = TransformFixMatch(mean = 0, std = 0))#TODO
val_dataset = CustomDataset(root= dataset_folder, split = "val", transform = val_transform)
labeled_train_loader = DataLoader(labeled_train_dataset, batch_size= batch_size_labeled, shuffle= True, num_workers= 4)
unlabeled_train_loader = DataLoader(unlabeled_train_dataset, batch_size= batch_size_unlabeled, shuffle= True, num_workers= 4)
val_loader = DataLoader(val_dataset, batch_size= batch_size_val, shuffle= False, num_workers= 4)
labeled_iter = iter(labeled_train_loader)
unlabeled_iter = iter(unlabeled_train_loader)
model = wide_resnet50_2(pretrained=False, num_classes = 800)
classifier = Classifier(ip= 2048, dp = 0)
start_epoch = 0
checkpoint = torch.load(args.transfer_path, map_location= device)
model.load_state_dict(checkpoint['model_state_dict'])
classifier.load_state_dict(checkpoint['classifier_state_dict'])
param_groups = [dict(params=classifier.parameters(), lr=args.learning_rate)]
if args.fine_tune:
param_groups.append(dict(params=model.parameters(), lr=args.learning_rate))
optimizer = torch.optim.SGD(param_groups,
lr = learning_rate,
momentum= momentum,
nesterov= True,
weight_decay= weight_decay)
scheduler = get_cosine_schedule_with_warmup(optimizer, 0, num_training_steps= n_epochs * n_steps)
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
model = torch.nn.DataParallel(model)
classifier = torch.nn.DataParallel(classifier)
if train_from_start == 0:
assert os.path.isfile(checkpoint_path), "Error: no checkpoint directory found!"
print("Restoring model from checkpoint")
# args.out = os.path.dirname(args.resume)
checkpoint = torch.load(checkpoint_path)
# best_acc = checkpoint['best_acc']
start_epoch = checkpoint['epoch'] - 1
model.load_state_dict(checkpoint['backbone_state_dict'])
classifier.load_state_dict(checkpoint['classifier_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
model = model.to(device)
classifier = classifier.to(device)
model.train()
losses = Average()
losses_l = Average()
losses_u = Average()
mask_probs = Average()
best_val_accuracy = 25.0 #TODO
for epoch in tqdm(range(start_epoch, n_epochs)):
if args.fine_tune:
model.train()
classifier.train()
else:
model.eval()
classifier.train()
for batch_idx in tqdm(range(n_steps)):
try:
img_lab, targets_lab = labeled_iter.next()
except:
labeled_iter = iter(labeled_train_loader)
img_lab, targets_lab = labeled_iter.next()
try:
unlab, _ = unlabeled_iter.next()
img_weak = unlab[0]
img_strong = unlab[1]
except:
unlabeled_iter = iter(unlabeled_train_loader)
unlab, _ = unlabeled_iter.next()
img_weak = unlab[0]
img_strong = unlab[1]
img_lab = img_lab.to(device)
targets_lab = targets_lab.to(device)
img_weak = img_weak.to(device)
img_strong = img_strong.to(device)
img_cat = torch.cat((img_lab, img_weak, img_strong), dim = 0)
logits_cat = classifier(model(img_cat))
logits_lab = logits_cat[:batch_size_labeled]
# print(logits_lab.size())
logits_unlab = logits_cat[batch_size_labeled:]
# print(logits_unlab)
logits_weak, logits_strong = torch.chunk(logits_unlab, chunks= 2, dim = 0)
pseudo_label = torch.softmax(logits_weak.detach()/tau, dim= 1)
max_probs, targets_unlab = torch.max(pseudo_label, dim= 1)
mask = max_probs.ge(threshold).float()
loss_labeled = F.cross_entropy(logits_lab, targets_lab, reduction='mean')
# print("CE: ", F.cross_entropy(logits_strong, targets_unlab, reduction= 'none').size())
loss_unlabeled = (F.cross_entropy(logits_strong, targets_unlab, reduction= 'none') * mask).mean()
# print("Loss labelled, loss unlabelled: ", loss_labeled, loss_unlabeled)
loss_total = loss_labeled + lamd * loss_unlabeled
# print("Total loss: ", loss_total)
# loss_epoch += loss_total
# loss_lab_epoch += loss_labeled
# loss_unlab_epoch += loss_unlabeled
losses.update(loss_total.item())
losses_l.update(loss_labeled.item())
losses_u.update(loss_unlabeled.item())
mask_probs.update(mask.mean().item())
optimizer.zero_grad()
loss_total.backward()
optimizer.step()
scheduler.step()
# break
if batch_idx % 25 == 0:
print(f"Epoch number: {epoch}, loss: {losses.avg}, loss lab: {losses_l.avg}, loss unlab: {losses_u.avg}, mask: {mask_probs.avg}, loss_here: {loss_total.item()}, best accuracy: {best_val_accuracy:.2f}", flush= True)
# print(optimizer.param_groups[0]['lr'])
save_checkpoint({
'epoch': epoch + 1,
'model_state_dict': model.state_dict(),
'classifier_state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
}, checkpoint_path)
model.eval()
classifier.eval()
with torch.no_grad():
val_loss = 0
val_size = 0
total = 0
correct = 0
for batch in val_loader:
logits_val = classifier(model(batch[0].to(device)))
labels = batch[1].to(device)
val_loss += F.cross_entropy(logits_val, labels)
_, predicted = torch.max(logits_val.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
val_size += 1
# break
print(f"Val loss: {val_loss/val_size}, Accuracy: {(100 * correct / total):.2f}%", flush= True)
if 100 * correct / total > best_val_accuracy:
best_val_accuracy = 100 * correct / total
best_val_loss = val_loss/val_size
print(f"Saving the best model with {best_val_accuracy:.2f}% accuracy and {best_val_loss:.2f} loss", flush= True)
save_checkpoint({
'epoch': epoch + 1,
'model_state_dict': model.state_dict(),
'classifier_state_dict': classifier.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'best_val_accuracy': best_val_accuracy,
'best_val_loss': best_val_loss
}, args.best_path)
model.train()
classifier.train()
def infer(img_dir,classes_csv,model_fname,resnet_depth,score_thresh,out_dir, results_fname):
# Create dataset
img_list = []
if not isinstance(img_dir, list):
img_dir = [img_dir]
for dir in img_dir:
for file in os.listdir(dir):
if file.endswith(".png"):
img_list.append(dir + file)
dataset_val = CustomDataset(img_list=img_list, class_list=classes_csv, transform=transforms.Compose([Normalizer(), Resizer()]))
sampler_val = AspectRatioBasedSampler(dataset_val, batch_size=1, drop_last=False)
dataloader_val = DataLoader(dataset_val, num_workers=1, collate_fn=collater, batch_sampler=sampler_val)
print(dataset_val.num_classes())
# Create the model
if resnet_depth == 18:
retinanet = model.resnet18(num_classes=dataset_val.num_classes())
elif resnet_depth == 34:
retinanet = model.resnet34(num_classes=dataset_val.num_classes())
elif resnet_depth == 50:
retinanet = model.resnet50(num_classes=dataset_val.num_classes())
elif resnet_depth == 101:
retinanet = model.resnet101(num_classes=dataset_val.num_classes())
elif resnet_depth == 152:
retinanet = model.resnet152(num_classes=dataset_val.num_classes())
else:
raise ValueError('Unsupported model depth, must be one of 18, 34, 50, 101, 152')
state_dict = torch.load(model_fname)
from collections import OrderedDict
new_state_dict = OrderedDict()
for k, v in state_dict.items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
# load params
retinanet.load_state_dict(new_state_dict)
if use_gpu:
retinanet = retinanet.cuda()
retinanet.eval()
unnormalize = UnNormalizer()
def draw_caption(image, box, caption):
b = np.array(box).astype(int)
cv2.putText(image, caption, (b[0], b[1] - 10), cv2.FONT_HERSHEY_PLAIN, 1, (0, 0, 0), 2)
cv2.putText(image, caption, (b[0], b[1] - 10), cv2.FONT_HERSHEY_PLAIN, 1, (255, 255, 255), 1)
results = []
for idx, data in enumerate(dataloader_val):
with torch.no_grad():
st = time.time()
scores, classification, transformed_anchors = retinanet(data['img'].cuda().float())
print('Elapsed time: {}, Num objects: {}'.format(time.time() - st, len(scores)))
idxs = np.where(scores > score_thresh)
img = np.array(255 * unnormalize(data['img'][0, :, :, :])).copy()
img[img < 0] = 0
img[img > 255] = 255
img = np.transpose(img, (1, 2, 0)).astype(np.uint8).copy()
bboxes = []
for j in range(idxs[0].shape[0]):
bbox = transformed_anchors[idxs[0][j], :]
x1 = int(bbox[0] / data['scale'][0])
y1 = int(bbox[1] / data['scale'][0])
x2 = int(bbox[2] / data['scale'][0])
y2 = int(bbox[3] / data['scale'][0])
label_name = dataset_val.labels[int(classification[idxs[0][j]])]
draw_caption(img, (x1, y1, x2, y2), label_name)
cv2.rectangle(img, (x1, y1), (x2, y2), color=(0, 0, 255), thickness=2)
score = float(scores[idxs[0][j]])
bboxes.append([x1, y1, x2, y2, score])
img_fname = ntpath.basename(data['img_fname'][0])
results.append([img_fname, bboxes])
# fig, ax = plt.subplots(figsize=(12, 12))
# ax.imshow(img, interpolation='bilinear')
with open(out_dir+results_fname,"wb") as output_file:
pickle.dump(results, output_file)
aug_transform3 = transforms.Compose([
transforms.RandomResizedCrop((96, 96),
scale=(0.08, 1.0),
ratio=(0.75, 1.3333333333333333)),
rnd_color_jitter3,
rnd_gray,
transforms.ToTensor(),
])
train_transform = transforms.Compose([
transforms.ToTensor(),
])
trainset = CustomDataset(root='/dataset',
split="train",
transform=train_transform)
augset1 = CustomDataset(root='/dataset',
split="train",
transform=aug_transform1)
augset2 = CustomDataset(root='/dataset',
split="train",
transform=aug_transform2)
augset3 = CustomDataset(root='/dataset',
split="train",
transform=aug_transform3)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=256,
shuffle=True,
num_workers=2)
augloader1 = torch.utils.data.DataLoader(augset1,
def main_worker(gpu, args):
global best_acc1
args.gpu = gpu
if args.gpu is not None:
print("Use GPU: {} for training".format(args.gpu))
# create model
print("=> creating model '{}'".format(args.arch))
model = models.__dict__[args.arch](num_classes=800,
norm_layer=SubBatchNorm2d)
# model = models.__dict__[args.arch](num_classes=800)
# # freeze all layers but the last fc
for name, param in model.named_parameters():
if name not in ['fc.weight', 'fc.bias']:
param.requires_grad = False
# init the fc layer
model.fc.weight.data.normal_(mean=0.0, std=0.1)
model.fc.bias.data.zero_()
# load from pre-trained, before DistributedDataParallel constructor
if args.pretrained:
if os.path.isfile(args.pretrained):
print("=> loading checkpoint '{}'".format(args.pretrained))
checkpoint = torch.load(args.pretrained, map_location="cpu")
# rename moco pre-trained keys
state_dict = checkpoint['state_dict']
for k in list(state_dict.keys()):
# retain only encoder_q up to before the embedding layer
if k.startswith(
'encoder_q') and not k.startswith('encoder_q.fc'):
# remove prefix
state_dict[k[len("encoder_q."):]] = state_dict[k]
# delete renamed or unused k
del state_dict[k]
args.start_epoch = 0
msg = model.load_state_dict(state_dict, strict=False)
# print(f"=> loading state_dict: \n{list(state_dict.keys())}")
# print(f"=> missing state keys: \n{msg.missing_keys}")
assert set(msg.missing_keys) == {"fc.weight", "fc.bias"}
print("=> loaded pre-trained model '{}'".format(args.pretrained))
else:
raise ValueError("=> no pre-trained model found at '{}'".format(
args.pretrained))
if args.gpu is not None:
torch.cuda.set_device(args.gpu)
model = model.cuda(args.gpu)
else:
print("=> ERROR: gpu must be assigned")
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda(args.gpu)
# optimize only the linear classifier
parameters = list(filter(lambda p: p.requires_grad, model.parameters()))
assert len(parameters) == 2 # fc.weight, fc.bias
optimizer = torch.optim.SGD(parameters,
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# 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))
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'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
del checkpoint
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# Data loading code
traindir = args.data
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transform_train = transforms.Compose([
transforms.RandomResizedCrop(96), # add crop resize
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize
])
transform_eval = transforms.Compose([
transforms.Resize(128), # add resize
transforms.CenterCrop(96), # add crop
transforms.ToTensor(),
normalize
])
train_dataset = CustomDataset(traindir, 'train', transform_train)
eval_dataset = CustomDataset(traindir, 'val', transform_eval)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
eval_loader = torch.utils.data.DataLoader(eval_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
# training code
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch, args)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args)
if epoch == 0 or (epoch + 1) % args.eval_per_n_epoch == 0:
accuracy = evaluate(eval_loader, model, args)
print(f"=> Epoch: {epoch+1}, accuracy: {accuracy:.4f}")
# remember best acc and save checkpoint
is_best = accuracy > best_acc1
best_acc1 = max(accuracy, best_acc1)
print(f"=> Epoch: {epoch+1}, isBest? : {is_best}")
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'accuracy': accuracy,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
},
is_best,
save_dir=args.checkpoint_dir,
epoch=(epoch + 1),
filename=os.path.join(
args.checkpoint_dir,
'checkpoint_{:03d}.pth.tar'.format(epoch + 1)))
