action='store_true',
default=True,
help='save log')
# Working directory - model_path. Currently disabled
# parser.add_argument('--work-dir', type=str, default=model_path, metavar='WD',
# help='path to save')
# TO ADD: save_result
args = parser.parse_args()
device = 'cuda:0'
# pr = processor1.Processor(args, None, device)
# data, labels, data_train, labels_train, data_test, labels_test = \
# loader.load_data(data_path)
a = loader.TrainTestLoader(False)
d, l = a.__getitem__(0)
print(type(d))
print(type(l))
print(l.shape)
print(d.shape)
# print(np.amax(d))
img__ = d.numpy()
# print(np.amax(l))
img_ = np.zeros((d.shape[1], d.shape[2], 3))
img_[:, :, 0] = d[0, :, :]
img_[:, :, 1] = d[1, :, :]
img_[:, :, 2] = d[2, :, :]
img__ = img_.astype('float32')
io.imshow(img__)
io.show()
parser.add_argument('--work-dir',
type=str,
default=model_path,
metavar='WD',
help='path to save')
# TO ADD: save_result
args = parser.parse_args()
device = 'cuda:0'
data, labels, data_train, labels_train, data_test, labels_test =\
loader.load_data(data_path, ftype_real, ftype_synth, coords, joints, cycles=cycles)
num_classes = np.unique(labels_train).shape[0]
data_loader = {
'train':
torch.utils.data.DataLoader(dataset=loader.TrainTestLoader(
data_train, labels_train, joints, coords, num_classes),
batch_size=args.batch_size,
shuffle=True,
drop_last=True),
'test':
torch.utils.data.DataLoader(dataset=loader.TrainTestLoader(
data_test, labels_test, joints, coords, num_classes),
batch_size=args.batch_size,
shuffle=True,
drop_last=True)
}
graph_dict = {'strategy': 'spatial'}
print('Train set size: {:d}'.format(len(data_train)))
print('Test set size: {:d}'.format(len(data_test)))
print('Number of classes: {:d}'.format(num_classes))
pr = processor.Processor(args,
metrics_file_full_path = 'metrics.txt'
if not os.path.exists(metrics_file_full_path):
for init_idx in range(num_inits):
for fold_idx, (data_train, labels_train, data_test, labels_test) in enumerate(zip(data_train_all_folds, labels_train_all_folds,
data_test_all_folds, labels_test_all_folds)):
print('Running init {:02d}, fold {:02d}'.format(init_idx, fold_idx))
# saving trained models for each init and split in separate folders
model_path = os.path.join(base_path, 'model_classifier_combined_lstm_init_{:02d}_fold_{:02d}/features'.format(init_idx, fold_idx) + ftype)
args.work_dir = model_path
os.makedirs(model_path, exist_ok=True)
aff_features = len(data_train[0][0])
num_classes = np.unique(labels_train).shape[0]
data_loader = list()
data_loader.append(torch.utils.data.DataLoader(
dataset=loader.TrainTestLoader(data_train, labels_train, joints, coords),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_worker * ngpu(device),
drop_last=True))
data_loader.append(torch.utils.data.DataLoader(
dataset=loader.TrainTestLoader(data_test, labels_test, joints, coords),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_worker * ngpu(device),
drop_last=True))
data_loader = dict(train=data_loader[0], test=data_loader[1])
graph_dict = {'strategy': 'spatial'}
pr = processor.Processor(args, data_loader, coords*joints, aff_features, num_classes, graph_dict, device=device)
if args.train:
pr.train()
# Working directory - model_path. Currently disabled
# parser.add_argument('--work-dir', type=str, default=model_path, metavar='WD',
# help='path to save')
# TO ADD: save_result
args = parser.parse_args()
device = 'cuda:0'
#%% TBD: Load the dataset
if DEBUG == False:
# data, labels, data_train, labels_train, data_test, labels_test = \
# loader.load_data(data_path, ftype, coords, joints, cycles=cycles)
# num_classes = np.unique(labels_train).shape[0]
data_loader_train_test = list()
data_loader_train_test.append(
torch.utils.data.DataLoader(dataset=loader.TrainTestLoader(train=True),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_worker *
torchlight.ngpu(device),
drop_last=True))
data_loader_train_test.append(
torch.utils.data.DataLoader(
dataset=loader.TrainTestLoader(train=False),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_worker * torchlight.ngpu(device),
drop_last=True))
data_loader_train_test = dict(train=data_loader_train_test[0],
test=data_loader_train_test[1])
else:
affs_dim = affective_features_train.shape[-1] + deep_dim
affective_features = np.concatenate(
(affective_features_train, affective_features_test), axis=0)
affective_features, affs_max, affs_min = loader.scale_data(affective_features)
affective_features_train, _, _ = loader.scale_data(affective_features_train,
affs_max, affs_min)
affective_features_test, _, _ = loader.scale_data(affective_features_test,
affs_max, affs_min)
num_frames_max = rotations_train.shape[1]
num_frames_out = num_frames_max - 1
num_frames_train_norm = num_frames_train / num_frames_max
num_frames_test_norm = num_frames_test / num_frames_max
data_loader = list()
data_loader.append(
torch.utils.data.DataLoader(dataset=loader.TrainTestLoader(
data_train, poses_train, rotations_train, translations_train,
affective_features_train, num_frames_train_norm, labels_train, coords,
num_labels),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_worker * ngpu(device),
drop_last=True))
data_loader.append(
torch.utils.data.DataLoader(dataset=loader.TrainTestLoader(
data_test, poses_test, rotations_test, translations_test,
affective_features_test, num_frames_test_norm, labels_test, coords,
num_labels),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_worker * ngpu(device),
drop_last=True))
data_loader = dict(train=data_loader[0], test=data_loader[1])
