def main():
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
parser = argparse.ArgumentParser(
description="PyTorch Query Localization in Videos Training")
parser.add_argument(
"--config-file",
default="experiments/charades_sta_train.yaml",
# default="experiments/anet_cap_train.yaml",
# default="experiments/tacos_train.yaml",
metavar="FILE",
help="path to config file",
type=str,
)
args = parser.parse_args()
experiment_name = args.config_file.split("/")[-1]
log_directory = args.config_file.replace(experiment_name, "logs/")
vis_directory = args.config_file.replace(experiment_name, "visualization/")
experiment_name = experiment_name.replace(".yaml", "")
cfg.merge_from_list([
'EXPERIMENT_NAME', experiment_name, 'LOG_DIRECTORY', log_directory,
"VISUALIZATION_DIRECTORY", vis_directory
])
cfg.merge_from_file(args.config_file)
output_dir = "./{}".format(cfg.LOG_DIRECTORY)
if output_dir:
mkdir(output_dir)
mkdir("./checkpoints/{}".format(cfg.EXPERIMENT_NAME))
logger = setup_logger("mlnlp", output_dir, cfg.EXPERIMENT_NAME + ".txt", 0)
logger.info("Starting moment localization with dynamic filters")
logger.info(cfg.EXPERIMENT_NAME)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# torch.backends.cudnn.enabled = False
if cfg.ENGINE_STAGE == "TRAINER":
print('#######')
print(cfg.DYNAMIC_FILTER.LSTM_VIDEO.DROPOUT)
trainer(cfg)
elif cfg.ENGINE_STAGE == "TESTER":
tester(cfg)
def __init__(self, cfg, dataset_size, is_train=True):
self.loss = []
self.IoU = []
self.mIoU = []
self.aux_mIoU = []
self.individual_loss = {}
self.vis_dir = "{}{}".format(cfg.VISUALIZATION_DIRECTORY, cfg.EXPERIMENT_NAME)
print(self.vis_dir)
mkdir(self.vis_dir)
self.cfg = cfg
self.s_samples = np.random.randint(dataset_size, size=4)
self.s_samples = np.insert(self.s_samples,0, 100)
for s in self.s_samples:
self.individual_loss[str(s)] = []
mkdir("{}/{}".format(self.vis_dir, str(s)))
if is_train == True:
self.state = "training"
else:
self.state = "testing"
def main():
parser = argparse.ArgumentParser(description="PyTorch Object Detection Training")
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="path to config file",
type=str,)
args = parser.parse_args()
experiment_name = args.config_file.split("/")[-1]
log_directory = args.config_file.replace(experiment_name,"logs/")
vis_directory = args.config_file.replace(experiment_name,"visualization/")
experiment_name = experiment_name.replace(".yaml","")
cfg.merge_from_list(['EXPERIMENT_NAME', experiment_name, 'LOG_DIRECTORY', log_directory, "VISUALIZATION_DIRECTORY", vis_directory])
cfg.merge_from_file(args.config_file)
output_dir = "./{}".format(cfg.LOG_DIRECTORY)
if output_dir:
mkdir(output_dir)
mkdir("./checkpoints/{}".format(cfg.EXPERIMENT_NAME))
logger = setup_logger("mlnlp", output_dir, cfg.EXPERIMENT_NAME + ".txt", 0)
logger.info("Starting moment localization with dynamic filters")
logger.info(cfg.EXPERIMENT_NAME)
# reproductibility
np.random.seed(0)
torch.manual_seed(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
if cfg.ENGINE_STAGE == "TRAINER":
trainer(cfg)
elif cfg.ENGINE_STAGE == "TESTER":
tester(cfg)
def tsv_writer(values, tsv_file, sep='\t'):
mkdir(op.dirname(tsv_file))
lineidx_file = op.splitext(tsv_file)[0] + '.lineidx'
idx = 0
tsv_file_tmp = tsv_file + '.tmp'
lineidx_file_tmp = lineidx_file + '.tmp'
with open(tsv_file_tmp, 'w') as fp, open(lineidx_file_tmp, 'w') as fpidx:
assert values is not None
for value in values:
assert value is not None
# this step makes sure python2 and python3 encoded img string are the same.
# for python2 encoded image string, it is a str class starts with "/".
# for python3 encoded image string, it is a bytes class starts with "b'/".
# v.decode('utf-8') converts bytes to str so the content is the same.
# v.decode('utf-8') should only be applied to bytes class type.
value = [
v if type(v) != bytes else v.decode('utf-8') for v in value
]
v = '{0}\n'.format(sep.join(map(str, value)))
fp.write(v)
fpidx.write(str(idx) + '\n')
idx = idx + len(v)
os.rename(tsv_file_tmp, tsv_file)
os.rename(lineidx_file_tmp, lineidx_file)
is_distributed=cfg.distributed)
ntrain = len(trainloader.dataset)
steps_per_epoch = len(trainloader)
cfg.SOLVER.STEPS_PER_EPOCH = steps_per_epoch
logs_per_epoch = steps_per_epoch / cfg.LOG_FREQ
warmup = cfg.OPTIM.WARMUP * steps_per_epoch
cfg.OPTIM.WARMUP = warmup
cfg.SOLVER.MAX_ITER = steps_per_epoch * cfg.OPTIM.EPOCHS
# get the starting checkpoint's iteration
iteration = config_iteration(cfg.OUTPUT_DIR, steps_per_epoch)
logging.info("Experiment settings:")
logging.info(cfg)
if cfg.OUTPUT_DIR:
mkdir(cfg.OUTPUT_DIR)
# save full config to a file in output_dir for future reference
with open(os.path.join(cfg.OUTPUT_DIR, 'config.yaml'), 'w') as f:
f.write(str(cfg))
cfg.freeze()
# mix-up
aug = cfg.AUG
mixup_fn = Mixup(
mixup_alpha=aug.MIXUP,
cutmix_alpha=aug.MIXCUT,
cutmix_minmax=aug.MIXCUT_MINMAX if aug.MIXCUT_MINMAX else None,
prob=aug.MIXUP_PROB,
switch_prob=aug.MIXUP_SWITCH_PROB,
mode=aug.MIXUP_MODE,
def main():
# Add augments
parser = argparse.ArgumentParser(description="Vision Research Toolkit by PyTorch")
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"--skip-test",
dest="skip_test",
help="Do not test the final model",
action="store_true"
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER
)
args = parser.parse_args()
num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = num_gpus > 1
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(
backend="nccl", init_method="env://"
)
synchronize()
# make config
cfg = make_config(args.config_file, args.opts)
# obtain absolute dir of project
project_dir = os.path.dirname(os.path.dirname(os.path.realpath(__file__)))
if cfg.CHECKPOINTER.DIR:
if cfg.CHECKPOINTER.DIR[0] is not os.sep:
# if the saver_dir is not absolute dir
cfg.CHECKPOINTER.DIR = os.path.join(project_dir, cfg.CHECKPOINTER.DIR)
else:
cfg.CHECKPOINTER.DIR = os.path.join(project_dir, 'log')
if not cfg.CHECKPOINTER.NAME:
cfg.CHECKPOINTER.NAME = strftime("%Y-%m-%d-%H-%M-%S", localtime())
cfg.freeze()
save_dir = os.path.join(cfg.CHECKPOINTER.DIR, cfg.CHECKPOINTER.NAME)
mkdir(save_dir)
# Init logger
logger = setup_logger(cfg.NAME, save_dir, get_rank())
logger.info("Using {} GPU".format(num_gpus))
logger.info(args)
logger.info("Collecting env info ...")
logger.info("\n" + collect_env_info())
logger.info("Loaded configuration file {}".format(args.config_file))
logger.info("Running with config:\n{}".format(cfg))
output_config_path = os.path.join(save_dir, os.path.basename(args.config_file))
logger.info("Saving config into: {}".format(output_config_path))
# save overloaded model config in the output directory
save_config(cfg, output_config_path)
train(cfg, args.local_rank, args.distributed)
return
def main(cfgs: DictConfig):
logger = logging.getLogger(cfgs.arch)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
# create model
logger.info("getting model '{}' from torch hub".format(cfgs.arch))
model, input_size = initialize_model(
model_name=cfgs.arch,
num_classes=cfgs.num_classes,
feature_extract=cfgs.feature_extract,
use_pretrained=cfgs.pretrained,
)
logger.info("model: '{}' is successfully loaded".format(
model.__class__.__name__))
logger.info("model structure: {}".format(model))
# Data augmentation and normalization for training
# Just normalization for validation
logger.info("Initializing Datasets and Dataloaders...")
logger.info("loading data {} from {}".format(cfgs.dataset, cfgs.data_path))
dataloaders_dict = load_data(input_size=input_size,
batch_size=cfgs.batch_size,
data_path=cfgs.data_path,
num_workers=cfgs.workers)
# Detect if we have a GPU available
device = torch.device(cfgs.device if torch.cuda.is_available() else "cpu")
# Gather the parameters to be optimized/updated in this run. If we are
# finetuning we will be updating all parameters. However, if we are
# doing feature extract method, we will only update the parameters
# that we have just initialized, i.e. the parameters with requires_grad
# is True.
params_to_update = model.parameters()
param_log_info = ''
if cfgs.feature_extract:
params_to_update = []
for name, param in model.named_parameters():
if param.requires_grad:
params_to_update.append(param)
param_log_info += "\t{}".format(name)
else:
for name, param in model.named_parameters():
if param.requires_grad:
param_log_info += "\t{}".format(name)
logger.info("Params to learn:\n" + param_log_info)
# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(params_to_update,
lr=cfgs.lr,
momentum=cfgs.momentum)
# Setup the loss fxn
criterion = nn.CrossEntropyLoss()
# Train and evaluate
model_ft = train_model(model,
dataloaders_dict,
device,
criterion,
optimizer_ft,
logger,
print_freq=cfgs.print_freq,
num_epochs=cfgs.epochs,
is_inception=(cfgs.arch == "inception"))
mkdir(cfgs.weight_dir)
torch.save(model_ft.state_dict(),
os.path.join(cfgs.weight_dir, cfgs.arch) + '.ckpt')
logger.info("model is saved at {}".format(
os.path.abspath(os.path.join(cfgs.weight_dir, cfgs.arch) + '.ckpt')))
def main():
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
global best_acc1
run_time = str(datetime.datetime.now())
cfgs = load_configs(CONFIG_FILE)
# create log dir and weight dir
mkdir(cfgs['weight_dir'])
mkdir(cfgs['log_dir'])
# create logger
log_dir = osp.join(cfgs['log_dir'], cfgs['arch'])
mkdir(log_dir)
cfgs['log_name'] = cfgs['arch'] + '_' + cfgs['dataset']
logger = setup_logger(cfgs['log_name'], log_dir, get_rank(),
run_time + '.txt')
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
logger.info("Loaded configuration file {}".format(CONFIG_FILE))
logger.info("Running with config:\n{}".format(cfgs))
# create model
logger.info("=> creating model '{}'".format(cfgs['arch']))
model = models.__dict__[cfgs['arch']]()
if cfgs['arch'].lower().startswith('wideresnet'):
# a customized resnet model with last feature map size as 14x14 for better class activation mapping
model = wideresnet.resnet50(num_classes=cfgs['num_classes'])
else:
model = models.__dict__[cfgs['arch']](num_classes=cfgs['num_classes'])
if cfgs['arch'].lower().startswith(
'alexnet') or cfgs['arch'].lower().startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
logger.info("=> created model '{}'".format(model.__class__.__name__))
logger.info("model structure: {}".format(model))
num_gpus = torch.cuda.device_count()
logger.info("using {} GPUs".format(num_gpus))
# optionally resume from a checkpoint
if cfgs['resume']:
if osp.isfile(cfgs['resume']):
logger.info("=> loading checkpoint '{}'".format(cfgs['resume']))
checkpoint = torch.load(cfgs['resume'])
cfgs['start_epoch'] = checkpoint['epoch']
best_acc1 = checkpoint['best_acc1']
model.load_state_dict(checkpoint['state_dict'])
logger.info("=> loaded checkpoint '{}' (epoch {})".format(
cfgs['resume'], checkpoint['epoch']))
else:
logger.info("=> no checkpoint found at '{}'".format(
cfgs['resume']))
torch.backends.cudnn.benchmark = True
# Data loading code
traindir = osp.join(cfgs['data_path'], 'train')
valdir = osp.join(cfgs['data_path'], 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_loader = torch.utils.data.DataLoader(datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])),
batch_size=cfgs['batch_size'],
shuffle=True,
num_workers=cfgs['workers'],
pin_memory=True)
val_loader = torch.utils.data.DataLoader(datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=cfgs['batch_size'],
shuffle=False,
num_workers=cfgs['workers'],
pin_memory=True)
# define loss function (criterion) and pptimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
cfgs['lr'],
momentum=cfgs['momentum'],
weight_decay=float(cfgs['weight_decay']))
# if cfgs['evaluate']:
# validate(val_loader, model, criterion, cfgs)
# return
# for epoch in range(cfgs['start_epoch'], cfgs['epochs']):
# adjust_learning_rate(optimizer, epoch, cfgs)
# # train for one epoch
# train(train_loader, model, criterion, optimizer, epoch, cfgs)
# # evaluate on validation set
# acc1 = validate(val_loader, model, criterion, cfgs)
# # remember best acc@1 and save checkpoint
# is_best = acc1 > best_acc1
# best_acc1 = max(acc1, best_acc1)
# save_checkpoint({
# 'epoch': epoch + 1,
# 'arch': cfgs['arch'],
# 'state_dict': model.state_dict(),
# 'best_acc1': best_acc1,
# }, is_best, cfgs['weight_dir'] + '/' + cfgs['arch'].lower())
logger.info("start to test the best model")
best_weight = cfgs['weight_dir'] + '/' + cfgs['arch'].lower(
) + '_best.pth.tar'
if osp.isfile(best_weight):
logger.info("=> loading best model '{}'".format(best_weight))
checkpoint = torch.load(best_weight)
best_acc1 = checkpoint['best_acc1']
epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
logger.info("=> loaded checkpoint '{}' (val Acc@1 {})".format(
best_weight, best_acc1))
else:
logger.info("=> no best model found at '{}'".format(best_weight))
acc1 = validate(val_loader, model, criterion, cfgs)
