def main():
parser = ArgumentParser()
parser.add_argument('-i',
dest='input_data',
default='freiburg1_xyz.sqlite3',
help='input data set')
parser.add_argument('-t',
dest='test_size',
default=10000,
type=int,
help='test data size')
parser.add_argument('-b',
dest='batch_size',
default=100,
type=int,
help='batch size for testing')
parser.add_argument('-c',
dest='checkpoint',
default='KinectPose2/KinectPose-100000')
args = parser.parse_args()
input_width, input_height, input_channel = 640, 480, 1
output_size = 7
if os.path.isfile(args.checkpoint + '.index') and \
os.path.isfile(args.checkpoint + '.meta'):
print('preparing data loader...')
data_loader = FreiburgData(args.input_data)
print('loading pre-train model...')
model = KinectPoseModel(input_width,
input_height,
input_channel,
output_size,
model_name='KinectPose',
saving=False)
with tf.Session() as sess:
model.load(sess, args.checkpoint)
batches = args.test_size / args.batch_size
outputs = np.zeros(shape=[output_size], dtype=np.float32)
for i in range(batches):
p, n, l = data_loader.diff_depth_batch(args.batch_size)
prediction = model.predict(sess, p, n)
outputs += np.sum(np.abs(prediction - l), axis=0)
print('processing batch %d...' % (i))
print(prediction[0, :])
print(l[0, :])
p_image = p[0, :, :, 0]
n_image = n[0, :, :, 0]
plt.figure()
plt.imshow(p_image, cmap='gray')
plt.figure()
plt.imshow(n_image, cmap='gray')
plt.show()
outputs /= args.test_size
print(outputs)
else:
print('checkpoint not exists.')
def main():
parser = ArgumentParser()
parser.add_argument('--input-db', dest='input_db',
default='freiburg1_xyz.sqlite3',
help='input dataset')
parser.add_argument('--checkpoint', dest='checkpoint',
default='KinectPose2/KinectPose-100000',
help='checkpoint path')
args = parser.parse_args()
data = FreiburgData(args.input_db)
logger.info(data.depth_images.shape)
predicted_poses = predict_seq(data.depth_images, args.checkpoint,
data.depth_labels[0, :])
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
ax.plot(data.depth_labels[:, 0], data.depth_labels[:, 1],
data.depth_labels[:, 2], 'g')
ax.plot(predicted_poses[:, 0], predicted_poses[:, 1],
predicted_poses[:, 2], 'r')
ax.set_xlabel('$X$')
ax.set_ylabel('$Y$')
plt.show()
def main():
parser = ArgumentParser()
parser.add_argument('--db', dest='db', default='freiburg1_xyz.sqlite3',
help='input database')
parser.add_argument('--batch-size', type=int, dest='batch_size',
default=64, help='batch size for training')
parser.add_argument('--max-epoch', type=int, dest='max_epoch',
default=100000, help='max epoch for training')
parser.add_argument('--output-epoch', type=int, dest='output_epoch',
default=10, help='epoch for display and saving')
parser.add_argument('--keep-prob', type=float, dest='keep_prob',
default=0.8, help='keep probability for drop out')
parser.add_argument('--decay-epoch', type=int, dest='decay_epoch',
default=5000, help='decay learning rate epoch')
args = parser.parse_args()
data_loader = FreiburgData(args.db)
input_width = 640
input_height = 480
input_channel = 1
output_size = 7
model = KinectPoseModel(input_width, input_height, input_channel,
output_size, model_name='KinectPose', saving=True)
with tf.Session() as sess:
model.train_with_loader(sess, data_loader,
batch_size=args.batch_size,
output_period=args.output_epoch,
decay_epoch=args.decay_epoch,
keep_prob=args.keep_prob,
max_epoch=args.max_epoch)
def main():
parser = ArgumentParser()
parser.add_argument('--input-db',
dest='input_db',
default='freiburg1_360.sqlite3',
help='input dataset')
parser.add_argument(
'--checkpoint',
dest='checkpoint',
default='saved_KP_320_240_3_diff10_c11/KinectPose-3400',
help='checkpoint path')
parser.add_argument('--diff',
dest='diff',
type=int,
default='10',
help='prediction frame diff')
args = parser.parse_args()
data = FreiburgData(args.input_db)
logger.info(data.depth_images.shape)
predicted_poses, acc = predict_seq(data.rgb_images,
args.checkpoint,
data.rgb_labels,
diff=args.diff)
# print(data.rgb_labels[0:10])
# print(predicted_poses[0:10])
print(acc)
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
ax.plot(data.depth_labels[:, 0], data.depth_labels[:, 1],
data.depth_labels[:, 2], 'g') #depth?????
ax.plot(predicted_poses[:, 0], predicted_poses[:, 1],
predicted_poses[:, 2], 'r')
ax.set_xlabel('$X$')
ax.set_ylabel('$Y$')
plt.show()
def main():
logging.basicConfig()
logger = logging.getLogger('train')
logger.setLevel(logging.INFO)
parser = ArgumentParser()
parser.add_argument('--db',
dest='db',
default='freiburg1_xyz.sqlite3',
help='input database')
parser.add_argument('--batch-size',
type=int,
dest='batch_size',
default=256,
help='batch size for training')
parser.add_argument('--max-epoch',
type=int,
dest='max_epoch',
default=10000,
help='max epoch for training')
parser.add_argument('--output-epoch',
type=int,
dest='output_epoch',
default=10,
help='epoch for display and saving')
parser.add_argument('--keep-prob',
type=float,
dest='keep_prob',
default=0.8,
help='keep probability for drop out')
parser.add_argument('--decay-epoch',
type=int,
dest='decay_epoch',
default=1000,
help='decay learning rate epoch')
parser.add_argument('--input-type',
dest='input_type',
default='rgbd',
help='type of input(channels)')
args = parser.parse_args()
data_loader = FreiburgData(args.db)
input_width = 320
input_height = 240
if args.input_type == 'depth':
input_channel = 1
logger.info('input type: depth')
elif args.input_type == 'rgb':
input_channel = 3
logger.info('input type: rgb')
elif args.input_type == 'rgbd':
input_channel = 4
logger.info('input type: rgbd')
output_size = 3
model = KinectPoseModel(input_width,
input_height,
input_channel,
output_size,
kernel_size=3,
model_name='KinectPose',
saving=True)
with tf.Session() as sess:
model.train_with_loader(sess,
data_loader,
batch_size=args.batch_size,
output_period=args.output_epoch,
decay_epoch=args.decay_epoch,
keep_prob=args.keep_prob,
max_epoch=args.max_epoch,
input_type=args.input_type)