# test_input = None
# test_label = None
feature_dimension = train_data[0].shape[1]
if args.type == "conv":
train_data = np.expand_dims(train_data, axis=1)
label_data = np.expand_dims(label_data, axis=1)
net = conv_transform_net(f_dimension=feature_dimension, \
mid_dimension=args.mid_dimen, mid_num=args.mid_num, \
seq_length=args.seq_length)
if (args.test_input_mat is not None):
test_input = np.expand_dims(test_input, axis=1)
test_label = np.expand_dims(test_label, axis=1)
else:
net = transform_net(f_dimension=feature_dimension, \
mid_dimension=args.mid_dimen, mid_num=args.mid_num)
save_name = args.saving_title + args.type + 'transform_{}-midNum_{}-midDimen'.format(
args.mid_num, args.mid_dimen)
net = train(train_data, label_data, net, args.batch_size, args.epoch_num,
args.learning_rate, test_input_mat_list, test_label_mat,
args.whether_save, save_name)
# net = train(train_data, label_data, net, args.batch_size, args.epoch_num, args.learning_rate)
if (args.whether_save is True):
save_name = args.saving_title + args.type + 'transform_{}-midNum_{}-midDimen-epoch_{}.pth'.format(
args.mid_num, args.mid_dimen, args.epoch_num)
save_model(net, save_name)
type=str,
default="transform_lfw_feature.mat")
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.use_gpu
predicts = []
net = getattr(net_sphere, args.net)()
net.load_state_dict(torch.load(args.model))
net.cuda()
net.eval()
net.feature = True
net.requires_grad = True
transform_net = transform_net(f_dimension=args.feature_dimen, \
mid_dimension=args.mid_dimen, mid_num=args.mid_num)
zfile = zipfile.ZipFile(args.dataset)
landmark = {}
with open(args.BLUFR) as f:
landmark_lines = f.readlines()
num_person = len(landmark_lines) # number of the total images
img = []
feature_list = []
batch_num = int(num_person / args.batch_size)
for i in range(batch_num):
img_batch = []
for j in range(args.batch_size):
line = landmark_lines[i * args.batch_size + j].strip('\n')
plt.plot(tensor_dict_resize[i][j, 0],
tensor_dict_resize[i][j, 1],
color=color_param_mark[i],
marker='.')
plt.subplot(221)
for i in range(class_num):
a, b = np.shape(tensor_dict[i])
for j in range(a):
plt.plot(tensor_dict[i][j, 0],
tensor_dict[i][j, 1],
color=color_param[i],
marker='.')
transform_net = transform_net(f_dimension=2, \
mid_dimension=args.mid_dimen, mid_num=args.mid_num)
plt.subplot(223)
for i in range(class_num):
a, b = np.shape(tensor_dict[i])
for j in range(a):
plt.plot(tensor_dict[i][j, 0],
tensor_dict[i][j, 1],
color=color_param[i],
marker='.')
transform_dict = {}
for i in range(class_num):
tensor_i = torch.from_numpy(tensor_dict[i]).type(
torch.FloatTensor) # .cuda()
feature = transform_net(tensor_i)