def train(self,
canvas: Canvas,
ground_truth,
iteration=-1,
test=False,
initialize=False):
sess = canvas.sess
if iteration < 0:
iteration = self.iter
if initialize:
sess.run(tf.global_variables_initializer())
for i in range(iteration):
self.global_step += 1
self.train_step(sess, ground_truth.get_batch(self.batch_size),
i + 1)
if (i + 1) % self.show_steps == 0:
self.eval_step(sess, ground_truth.get_batch(self.batch_size),
i + 1)
if self.save_step is not None and self.save_step > 0 and (
i + 1) % self.save_step == 0:
self.save_graph()
canvas.save_step(self.global_step)
if test and ground_truth.batch_manager.epoched:
print("an epoch finished")
self.test_step(sess, ground_truth.get_test())
ground_truth.batch_manager.epoched = False
if test:
self.test_step(sess, ground_truth.get_test())
def get_canvas():
global canvas
if canvas is None:
if load_step >= 0:
canvas = Canvas()
else:
canvas = Canvas()
return canvas
def test_mlp():
mlp_canvas = Canvas()
sr = ShapeRegressor(mlp_canvas, conf_path('model/mlp/1'))
res = sr.gen(np.zeros(10))
print(np.mean(np.square(res)))
mlp_canvas.close()
def test_gru():
gru_canvas = Canvas()
pr = PoseRegressor(gru_canvas, conf_path('model/sgru/1'), 5)
res = pr.gen(np.zeros(10), np.zeros((24, 3)))
print(np.mean(np.square(res)))
gru_canvas.close()
def serve():
smpl = SMPLModel(conf_path('smpl'))
from com.learning.canvas import Canvas
from com.learning.mlp import Graph
canvas = Canvas()
g = Graph()
canvas.load_graph(conf_path('temp/mlp/sklt/3'), g)
smpl = SMPLModel(conf_path('smpl'))
vf = VirtualFitting()
def idle(m):
m.seek(0)
kinect_pos = m.read(size(['k']))
kinect_pos = np.fromstring(kinect_pos, 'f').reshape((21, 3))
axangle = kinect_pos[0]
if not math.isnan(axangle[0]) and not np.linalg.norm(
np.array(axangle)) == 0:
ai, aj, ak = tr.euler.mat2euler(
tr.axangles.axangle2mat(axangle,
math.radians(np.linalg.norm(axangle))))
root_rot = [ai, aj, ak]
else:
root_rot = [0, 0, 0]
kinect_joints = kinect_pos[1:]
pose = kinect_joints_to_smpl(kinect_joints, root_rot, g, smpl)
print(pose)
vf.pose = pose
vf.recalculate_pose_displacement()
verts = np.array(vf.cloth.vertices, 'f').tobytes()
faces = np.array(vf.cloth.faces, 'i').tobytes()
m.seek(size(['k']))
m.write(verts)
m.write(faces)
m.flush()
cn = create_shared_connect(size(['k', 'v', 'f', 'b']), 'lbac', 0.01)
cn.idle.insert(0, idle)
print('start')
try:
cn.start()
except KeyboardInterrupt:
print('stopped!')
canvas.close()
vf.close()
return
canvas.close()
vf.close()
def serve():
smpl = SMPLModel(conf_path('smpl'))
from com.learning.canvas import Canvas
from com.learning.mlp import Graph
canvas = Canvas()
g = Graph()
canvas.load_graph(conf_path('temp/mlp/sklt/3'), g)
smpl = SMPLModel(conf_path('smpl'))
def idle(m):
timer = Timer(False)
m.seek(0)
kinect_pos = m.read(size(['k']))
kinect_pos = np.fromstring(kinect_pos, 'f').reshape((21, 3))
# 坐标系转换
# kinect_pos[1:, 2] *= -1
axangle = kinect_pos[0]
print(np.linalg.norm(axangle))
if not math.isnan(axangle[0]):
ai, aj, ak = tr.euler.mat2euler(
tr.axangles.axangle2mat(axangle,
math.radians(np.linalg.norm(axangle))))
root_rot = [ai, aj, ak]
else:
root_rot = [0, 0, 0]
print(root_rot)
timer.tick('rot')
# for i in range(3):
# root_rot[i] = math.radians(root_rot[i])
kinect_joints = kinect_pos[1:]
# std = np.array([0, -0.295, 0])
# tmp = kinect_joints[0] - std
# for i in range(20):
# kinect_joints[i] -= tmp
pose = kinect_joints_to_smpl(kinect_joints, root_rot, g, smpl)
timer.tick('pose')
print(pose)
smpl.set_params(pose)
verts = np.array(smpl.verts, 'f').tobytes()
faces = np.array(smpl.faces, 'i').tobytes()
m.seek(size(['k']))
m.write(verts)
m.write(faces)
timer.tick('mesh')
m.flush()
cn = create_shared_connect(size(['k', 'v', 'f']), 'pstm', 0.01)
cn.idle.insert(0, idle)
print('start')
cn.start()
canvas.close()
def train(module, gt):
set_attributes(module)
canvas = Canvas()
canvas.open()
if load_step is not None:
canvas.load_step(load_step, output_dir)
g = module.Graph()
if load_step is not None:
g.restore()
g.global_step = load_step
else:
g.generate()
canvas.path = output_dir
g.train(canvas,
gt,
test=module.test_flag,
initialize=load_step is None)
if output_dir is not None:
canvas.save_all()
smpl = SMPLModel(conf_path('smpl'))
pose = np.zeros((24, 3))
pose[0] = (0, 0, 2)
pose[1:] += 0
smpl.set_params(pose)
joints = smpl.J
joints = apply(smpl, np.eye(24), joints)
joints = process_joints(joints, pose[0])
show_joints(joints)
if __name__ == '__main__':
# show_20_joints()
# save_20_joints()
sklt = load_json('sklt.json')
from com.learning.canvas import Canvas
from com.learning.mlp import Graph
canvas = Canvas()
g = Graph()
canvas.load_graph(conf_path('temp/mlp/sklt/3'), g)
kinect_joints_to_smpl(sklt[1:], [0, 0, 0], g)
canvas.close()
# test_process_joints()