def main():
args = parse_inputs()
# Load Model
net = GazeNet()
net.cuda()
resume_path = args.resume_path
net, optimizer, start_epoch = resume_checkpoint(net, None, resume_path)
#Prepare dataloaders
test_images_dir = args.test_dir
test_pickle_path = args.test_annotation
#For GOO
val_set = GooDataset(test_images_dir,
test_pickle_path,
'test',
use_bboxes=True)
test_only = True
if not test_only:
test_and_save(net, test_data_loader)
# Get a random sample image from the dataset
idx = np.random.randint(len(val_set))
image_path = val_set[idx]['image_path']
eyes = val_set[idx]['eye_position']
bboxes = val_set[idx]['gt_bboxes']
heatmap, x, y = test_on_image(net, image_path, eyes)
draw_results(image_path, eyes, (x, y), bboxes)
def main():
args = parse_inputs()
# Load Model
net = ModelSpatial()
net.cuda()
resume_path = args.resume_path
net, optimizer, start_epoch = resume_checkpoint(net, None, resume_path)
#Prepare dataloaders
test_images_dir = args.test_dir
test_pickle_path = args.test_annotation
#For GOO
val_set = GazeDataset(test_images_dir, test_pickle_path, 'test')
test_set = GazeDataset(test_images_dir, test_pickle_path, 'test')
test_data_loader = torch.utils.data.DataLoader(test_set,
batch_size=16,
num_workers=8,
shuffle=False)
predictions_npz = args.predictions_npz
if predictions_npz is None:
print(
'==> No npzfile provided. Inference will be done on the test dataset and will be saved to predictions.npz'
)
test_and_save(net, test_data_loader)
predictions_npz = './predictions.npz'
print('==> Calculating eval metrics...')
metrics = calculate_metrics(predictions_npz, val_set)
PA_count = metrics['pa']
# CPA_count = metrics['cpa']
print('AUC:', metrics['auc'])
print('L2 Distance: ', metrics['l2'])
print('Angular Error:', metrics['angular'])
print("Percentage Distances: ",
[0.01, 0.03, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30])
print(
"Proximate Accuracy: \t%.3f\t%.3f\t%.3f\t%.3f\t%.3f\t%.3f\t%.3f\t%.3f\t"
% (
PA_count[0],
PA_count[1],
PA_count[2],
PA_count[3],
PA_count[4],
PA_count[5],
PA_count[6],
PA_count[7],
))
def main():
# Dataloaders for GOO
batch_size = args.batch_size
print('==> Loading Train Dataset')
train_set = GooDataset(args.train_dir, args.train_annotation, 'train', use_gazemask=args.gazemask)
train_data_loader = DataLoader(train_set, batch_size=batch_size,
shuffle=True, num_workers=16)
if args.test_dir is not None:
print('==> Loading Test Dataset')
test_set = GooDataset(args.test_dir, args.test_annotation, 'test')
test_data_loader = DataLoader(test_set, batch_size=batch_size//2,
shuffle=False, num_workers=8)
# Loads model
net = GazeNet()
net.cuda()
# Hyperparameters
start_epoch = 0
max_epoch = 25
learning_rate = args.init_lr
# Initializes Optimizer
gaze_opt = GazeOptimizer(net, learning_rate)
optimizer = gaze_opt.getOptimizer(start_epoch)
# Resuming Training
resume_training = args.resume_training
if resume_training:
net, optimizer, start_epoch = resume_checkpoint(net, optimizer, args.resume_path)
if args.test_dir is not None:
test(net, test_data_loader,logger)
for epoch in range(start_epoch, max_epoch):
# Update optimizer
optimizer = gaze_opt.getOptimizer(epoch)
# Train model
train(net, train_data_loader, optimizer, epoch, logger)
# Save model and optimizer at the last 5 epochs
if epoch > max_epoch-5:
save_path = args.save_model_dir
save_checkpoint(net, optimizer, epoch+1, save_path)
# Evaluate model
if args.test_dir is not None:
test(net, test_data_loader, logger)
def main():
# transform = _get_transform(args.input_resolution)
# Prepare data
print("Loading Data")
batch_size = args.batch_size
train_set = GazeDataset(args.train_dir, args.train_annotation, 'train')
train_data_loader = DataLoader(dataset=train_set,
batch_size=batch_size,
shuffle=False,
num_workers=8)
if args.test_dir is not None:
print('==> Loading Test Dataset')
test_set = GazeDataset(args.test_dir, args.test_annotation, 'test')
test_data_loader = DataLoader(test_set,
batch_size=1,
shuffle=False,
num_workers=0) # half train batch
# Loads model
print("Constructing model")
net = ModelSpatial()
net.cuda()
# net.cuda().to(device)
# Hyperparameters
start_epoch = 0
# max_epoch = 25
max_epoch = 45
learning_rate = args.init_lr
# Initial weights chong
if args.init_weights:
model_dict = net.state_dict()
pretrained_dict = torch.load(args.init_weights)
pretrained_dict = pretrained_dict['model']
model_dict.update(pretrained_dict)
net.load_state_dict(model_dict)
# Initializes Optimizer
gaze_opt = GazeOptimizer(net, learning_rate)
optimizer = gaze_opt.getOptimizer(start_epoch)
# Resuming Training
resume_training = args.resume_training
print(resume_training)
if resume_training:
net, optimizer, start_epoch = resume_checkpoint(
net, optimizer, args.resume_path)
if args.test_dir is not None:
test(net, test_data_loader, logger)
for epoch in range(start_epoch, max_epoch):
# Update optimizer
optimizer = gaze_opt.getOptimizer(epoch)
# Train model
train(net, train_data_loader, optimizer, epoch, logger)
# # Save model and optimizer at the last 5 epochs
if epoch % 4 == 0:
save_path = './saved_models_gazefollow/temp_chong/'
save_checkpoint(net, optimizer, epoch + 1, save_path)