batch += 1
_, _, _, _, c = correct_preds(probs, labels.squeeze())
if disp:
print(i, c)
correct.append(c)
PCE = np.mean(correct)
return PCE
if __name__ == "__main__":
split = 1
seq_length = 64
n_cpu = 6
model = EventDetector(
pretrain=True,
width_mult=1.0,
lstm_layers=1,
lstm_hidden=256,
bidirectional=True,
dropout=False,
)
save_dict = torch.load("models/swingnet_1800.pth.tar")
model.load_state_dict(save_dict["model_state_dict"])
model.cuda()
model.eval()
PCE = eval(model, split, seq_length, n_cpu, True)
print("Average PCE: {}".format(PCE))
n_cpu = 6
seq_length = args.seq_length
bs = args.batch_size # batch size
k = 10 # frozen layers
use_no_element = args.use_no_element
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Load Model')
model = EventDetector(pretrain=True,
width_mult=1.,
lstm_layers=1,
lstm_hidden=256,
device=device,
bidirectional=True,
dropout=False,
use_no_element=use_no_element
)
#print('model.py, class EventDetector()')
freeze_layers(k, model)
#print('utils.py, func freeze_laters()')
model.train()
model.to(device)
print('Loading Data')
# TODO: vid_dirのpathをかえる。stsqの動画を切り出したimage全部が含まれているdirにする
if use_no_element == False:
version_name = 'original_' + str(split) + '_' + str(noise_level)
print(version_name)
if __name__ == '__main__':
# training configuration
iterations = 10000
it_save = 10000 # save model every 100 iterations
n_cpu = 8
seq_length = 64
bs = 22 # batch size
k = 10 # frozen layers
model = EventDetector(pretrain=True,
width_mult=1.,
lstm_layers=1,
lstm_hidden=256,
bidirectional=True,
dropout=False)
freeze_layers(k, model)
model.train()
model.cuda()
dataset = GolfDB(data_file='data/train_split_{}.pkl'.format(split),
vid_dir='data/videos_160/',
seq_length=seq_length,
transform=transforms.Compose([
ToTensor(),
Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
]),
train=True,
_, _, _, _, c = correct_preds(probs, labels.squeeze())
if disp:
print(i, c)
correct.append(c)
PCE = np.mean(correct)
return PCE
if __name__ == '__main__':
seq_length = 64
n_cpu = 6
model = EventDetector(pretrain=True,
width_mult=1.,
lstm_layers=1,
lstm_hidden=256,
bidirectional=True,
dropout=False)
save_dict = torch.load('models/' + version_name + '_10000.pth.tar',
map_location=lambda storage, loc: storage)
model.load_state_dict(save_dict['model_state_dict'])
model.to(device)
model.eval()
PCE = eval(model, split, seq_length, n_cpu, True)
print('Average PCE: {}'.format(PCE))
if not os.path.exists('results'):
os.mkdir('results')
if bool_classical_loss:
rgb_count += 1
else:
preds[i] = optical_preds
deltas = np.abs(events - preds)
correct = (deltas <= tol).astype(np.uint8)
return preds, deltas, correct
if __name__ == '__main__':
split = 1
seq_length = 64
n_cpu = 6
model = EventDetector(pretrain=True,
width_mult=1,
lstm_layers=1,
lstm_hidden=256,
bidirectional=True,
dropout=False)
rgb_save_dict = torch.load('swingnet_1600.pth.tar')
model.load_state_dict(rgb_save_dict['model_state_dict'])
model.cuda()
model.eval()
_, _, rgb_probs, rgb_tols, rgb_events = myeval(model, split, seq_length,
n_cpu, False, 1)
optical_save_dict = torch.load('swingnet_1200.pth.tar')
model.load_state_dict(optical_save_dict['model_state_dict'])
model.cuda()
model.eval()
if __name__ == '__main__':
# training configuration
split = cfg.SPLIT
iterations = cfg.ITERATIONS
it_save = cfg.IT_SAVE # save model every 100 iterations
n_cpu = cfg.CPU_NUM
seq_length = cfg.SEQUENCE_LENGTH
bs = cfg.BATCH_SIZE # batch size
k = 10 # frozen layers
model = EventDetector(pretrain=True,
width_mult=1.,
lstm_layers=1,
lstm_hidden=256,
bidirectional=True,
dropout=False)
freeze_layers(k, model)
model.train()
model.cuda()
# model = nn.DataParallel(model)
# 用来训练非光流部分
# dataset = GolfDB(data_file='./data/train_split_{}.pkl'.format(split),
# vid_dir='/home/zqr/data/videos_160/',
# seq_length=seq_length,
# transform=transforms.Compose([transforms.ToPILImage(),
# transforms.RandomHorizontalFlip(0.5),
# transforms.RandomAffine(5,shear=5),
# transforms.ToTensor()]),
# training configuration
split = 1
iterations = 2000
it_save = 100 # save model every 100 iterations
n_cpu = 6
seq_length = 64
bs = 4 # batch size
k = 10 # frozen layers
device = 'cuda:0'
model = EventDetector(
# pretrain=False,
pretrain=True,
width_mult=1.0,
lstm_layers=1,
lstm_hidden=256,
bidirectional=True,
dropout=False,
device=device,
)
base_list = [-1]
event_th = 30
dataset = PointImgGolfDB(
data_file="data/train_split_{}.pkl".format(split),
vid_dir="data/videos_160/",
seq_length=seq_length,
# transform=transforms.Compose(
# [ToTensor(), Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]
# ),
train=True,
list_loss = {'loss':[], 'soft':[], 'count':[]}
if __name__ == '__main__':
# training configuration (From McNally et al.)
iterations = config['iterations']
it_save = 10000 # save model every 10000 iterations
n_cpu = 6
seq_length = 64
bs = 22 # batch size
k = 10 # frozen layers
model = EventDetector(pretrain=True,
width_mult=1.,
lstm_layers=1,
lstm_hidden=256,
bidirectional=True,
dropout=False)
freeze_layers(k, model)
model.train()
model.to(device)
dataset = GolfDB(data_file='data/train_split_{}.pkl'.format(split),
vid_dir='data/videos_160/',
seq_length=seq_length,
transform=transforms.Compose([ToTensor(),
Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]),
train=True,
noise_level=noise_level)
data_loader = DataLoader(dataset,
it_save = args.it_save # save model every 100 iterations
n_cpu = 6
seq_length = args.seq_length
bs = args.batch_size # batch size
k = 10 # frozen layers
use_no_element = args.use_no_element
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Load Model')
model = EventDetector(pretrain=True,
width_mult=1.,
lstm_layers=1,
lstm_hidden=256,
device=device,
bidirectional=True,
dropout=False,
use_no_element=use_no_element)
#print('model.py, class EventDetector()')
freeze_layers(k, model)
#print('utils.py, func freeze_laters()')
model.train()
model.to(device)
print('Loading Data')
# TODO: vid_dirのpathをかえる。stsqの動画を切り出したimage全部が含まれているdirにする
if use_no_element == False:
dataset = StsqDB(
data_file='data/no_ele/seq_length_{}/train_split_{}.pkl'.format(
