def main():
global args, batchSize, kSaveModel, bb_params
args = parser.parse_args()
print(args)
batchSize = args.batch_size
kSaveModel = args.save_freq
np.random.seed(args.manual_seed)
torch.manual_seed(args.manual_seed)
if cuda:
torch.cuda.manual_seed_all(args.manual_seed)
# load bounding box motion model params
bb_params['lambda_shift_frac'] = args.lambda_shift_frac
bb_params['lambda_scale_frac'] = args.lambda_scale_frac
bb_params['min_scale'] = args.min_scale
bb_params['max_scale'] = args.max_scale
# load datasets
alov = ALOVDataset(
os.path.join("../../pygoturn/data", 'imagedata++/'),
os.path.join("../../pygoturn/data",
'alov300++_rectangleAnnotation_full/'),
NormalizeToTensor(), input_size)
# imagenet = ILSVRC2014_DET_Dataset(os.path.join("../../pygoturn/data",
# 'ILSVRC2014_DET_train/'),
# os.path.join("../../pygoturn/data",
# 'ILSVRC2014_DET_bbox_train/'),
# bb_params,
# transform,
# input_size)
#list of datasets to train on
datasets = [
alov,
]
# load model
net = model.SPPGoNet().to(device)
# summary(net, [(3, 224, 224), (3, 224, 224)])
loss_fn = torch.nn.L1Loss(size_average=False).to(device)
# initialize optimizer
optimizer = optim.SGD(net.classifier.parameters(),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
if os.path.exists(args.save_directory):
print('Directory %s already exists' % (args.save_directory))
else:
os.makedirs(args.save_directory)
# start training
net = train_model(net, datasets, loss_fn, optimizer)
# save trained model
checkpoint = {'state_dict': net.state_dict()}
path = os.path.join(args.save_directory, 'pytorch_goturn.pth.tar')
torch.save(checkpoint, path)
def main():
global args
args = parser.parse_args()
print(args)
np.random.seed(args.manual_seed)
torch.manual_seed(args.manual_seed)
if use_gpu:
torch.cuda.manual_seed(args.manual_seed)
# load datasets
alov = ALOVDataset('../data/alov300/imagedata++/',
'../data/alov300/alov300++_rectangleAnnotation_full/',
transform)
imagenet = ILSVRC2014_DET_Dataset('../data/imagenet_img/',
'../data/imagenet_bbox/', transform,
args.lambda_shift_frac,
args.lambda_scale_frac, args.min_scale,
args.max_scale)
# list of datasets to train on
datasets = [alov, imagenet]
# load model
net = model.GoNet()
loss_fn = torch.nn.L1Loss(size_average=False)
if use_gpu:
net = net.cuda()
loss_fn = loss_fn.cuda()
# initialize optimizer
optimizer = optim.SGD(net.classifier.parameters(),
lr=args.learning_rate,
weight_decay=0.0005)
if os.path.exists(args.save_directory):
print('Directory %s already exists' % (args.save_directory))
else:
os.makedirs(args.save_directory)
# start training
net = train_model(net, datasets, loss_fn, optimizer)
model_dir = 0
if args.model_dir:
params = util.Params()
params.update(args.model_dir)
model_dir = args.model_dir
else:
model_dir_path = os.path.join(".", "model")
if not os.path.isdir(model_dir_path):
os.mkdir(model_dir_path)
model_dir = model_dir_path
params.cuda = torch.cuda.is_available()
alov = ALOVDataset('/large_storage/imagedata++',
'/large_storage/alov300++_rectangleAnnotation_full',
transform)
dataloader = DataLoader(alov, batch_size=params.batch_size)
use_gpu = torch.cuda.is_available()
model = model.Re3Net().cuda() if use_gpu else model.Re3Net()
optimizer = optim.Adam(model.parameters(), lr=params.learning_rate)
net = 0
loss_fn = model.loss_fn(params.cuda)
# Train the model
logging.info("Starting training for {} epoch(s)".format(params.num_epochs))
batchSize = args.batch_size
kSaveModel = args.save_freq
np.random.seed(args.manual_seed)
torch.manual_seed(args.manual_seed)
bb_params = {}
bb_params['lambda_shift_frac'] = 5
bb_params['lambda_scale_frac'] = 15
bb_params['min_scale'] = -0.4
bb_params['max_scale'] = 0.4
transform = NormalizeToTensor()
input_size = 224
alov = ALOVDataset(
os.path.join("../pygoturn/data", 'imagedata++/'),
os.path.join("../pygoturn/data", 'alov300++_rectangleAnnotation_full/'),
NormalizeToTensor(), input_size)
# imagenet = ILSVRC2014_DET_Dataset(os.path.join("../pygoturn/data",
# 'ILSVRC2014_DET_train/'),
# os.path.join("../pygoturn/data",
# 'ILSVRC2014_DET_bbox_train/'),
# bb_params,
# transform,
# input_size)
cuda = torch.cuda.is_available()
device = torch.device('cuda:0' if cuda else 'cpu')
net = GOTURN_AlexNET()
def main():
global args
args = parser.parse_args()
print(args)
np.random.seed(args.manual_seed)
torch.manual_seed(args.manual_seed)
if use_gpu:
torch.cuda.manual_seed(args.manual_seed)
# load datasets
alov = ALOVDataset('../data/alov300/imagedata++/',
'../data/alov300/alov300++_rectangleAnnotation_full/',
transform)
imagenet = ILSVRC2014_DET_Dataset('../data/imagenet_img/',
'../data/imagenet_bbox/', transform,
args.lambda_shift_frac,
args.lambda_scale_frac, args.min_scale,
args.max_scale)
# list of datasets to train on
datasets = [alov, imagenet]
# load model
net = model.GoNet()
loss_fn = torch.nn.L1Loss(size_average=False)
if use_gpu:
net = net.cuda()
loss_fn = loss_fn.cuda()
# initialize optimizer
trainable_weights = []
trainable_bias = []
for name, param in net.classifier.named_parameters():
if 'weight' in name:
trainable_weights.append(param)
elif 'bias' in name:
trainable_bias.append(param)
optimizer = optim.SGD([{
'params': trainable_weights,
'lr': args.learning_rate * 10
}, {
'params': trainable_bias,
'lr': args.learning_rate * 20
}],
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
if os.path.exists(args.save_directory):
print('Directory %s already exists' % (args.save_directory))
else:
os.makedirs(args.save_directory)
# start training
net = train_model(net, datasets, loss_fn, optimizer)
# save trained model
path = os.path.join(args.save_directory, 'final_model.pth')
torch.save(net.state_dict(), path)