def do_test(self, sess, summary_writer, step, names=None):
'''execute computation of the inference
a fast version of inference
'''
# during training
"""
self.uvd_pts = uvd_pts
self.xyz_pts = xyz_pts
self.val_dms = dms
self.est_hms = est_hms
self.gt_pose = poses
self.gt_uvd_pts = gt_uvd_pts
"""
if names is None:
f = open(self._log_path, 'a')
summary_str, gt_vals, xyz_vals = sess.run(
[self.val_summary_op, self.gt_pose, self.xyz_pts])
summary_writer.add_summary(summary_str, step)
maxJntError = []
f.write('[%s] step %d\n' % (datetime.now(), step))
for xyz_val, gt_val in zip(xyz_vals, gt_vals):
maxJntError.append(Evaluation.maxJntError(xyz_val, gt_val))
diff = (xyz_val - gt_val).reshape(-1, 3)
dist = alg.norm(diff, axis=1).reshape(-1, 1)
error_mat = np.concatenate((diff, dist), axis=1)
print(error_mat)
f.write(np.array_str(error_mat) + '\n')
print('validate error:', maxJntError)
f.write('validation error: {}\n'.format(maxJntError))
f.flush()
f.close()
return
if step % 100 == 0:
summary_str, xyz_vals, gt_vals, names, val_dms_v, gt_uvd_pts_v, uvd_pts_v = sess.run(
[
self.val_summary_op, self.xyz_pts, self.gt_pose, names,
self.val_dms, self.gt_uvd_pts, self.uvd_pts
])
summary_writer.add_summary(summary_str, step)
maxJntError = []
for xyz_val, gt_val in zip(xyz_vals, gt_vals):
maxJntError.append(Evaluation.maxJntError(xyz_val, gt_val))
diff = (xyz_val - gt_val).reshape(-1, 3)
dist = alg.norm(diff, axis=1).reshape(-1, 1)
print(np.concatenate((diff, dist), axis=1))
print('[step: %d]test error:' % step, maxJntError)
print('---\n')
return gt_vals, xyz_vals, names, val_dms_v, gt_uvd_pts_v, uvd_pts_v
gt_vals, xyz_vals, names, val_dms_v, gt_uvd_pts_v, uvd_pts_v = \
sess.run([self.gt_pose, self.xyz_pts, names, self.val_dms, self.gt_uvd_pts, self.uvd_pts])
return gt_vals, xyz_vals, names, val_dms_v, gt_uvd_pts_v, uvd_pts_v
import sys
import numpy as np
from data.evaluation import Evaluation
xyz_gt_list = np.loadtxt("xyz_gt_list_all.txt")
xyz_val_list1 = np.loadtxt(
"/home/chen/Documents/GNN_demo/train_values_all_1.txt")
xyz_val_list1 = xyz_val_list1.reshape([-1, 42])
xyz_val_list2 = np.loadtxt(
"/home/chen/Documents/GNN_demo/train_values_all_2.txt")
xyz_val_list2 = xyz_val_list2.reshape([-1, 42])
xyz_val_list3 = np.loadtxt(
"/home/chen/Documents/GNN_demo/train_values_all_3.txt")
xyz_val_list3 = xyz_val_list3.reshape([-1, 42])
xyz_val_list4 = np.loadtxt(
"/home/chen/Documents/GNN_demo/train_values_all_4.txt")
xyz_val_list4 = xyz_val_list4.reshape([-1, 42])
xyz_val_list = np.concatenate(
(xyz_val_list1, xyz_val_list2, xyz_val_list3, xyz_val_list4), axis=0)
xyz_val_list_ori = np.loadtxt(
"/home/chen/Documents/denseReg-master/model/xyz_val_test_all.txt")
meanJntError = []
err_path = "/home/chen/Documents/denseReg-master/exp/train_cache/saved/GNN/_error.txt"
for i in range(8252):
#xyz_val = xyz_val_list[i,:]#图网络优化后的输出，误差test10.060924231540694mm
xyz_val = xyz_val_list_ori[i, :] #原始输出，误差10.233949127173911mm
gt_val = xyz_gt_list[i, :]
meanJntError.append(Evaluation.meanJntError(xyz_val, gt_val))
Evaluation.plotError(meanJntError, err_path)
'--initial-state',
'-s',
type=int,
default=1,
help=
'Defines the initial state (i.e. initial city, where teams are spawned)')
parser.add_argument(
'--local-search',
'-l',
choices=LocalSearchFactory.choices(),
default=None,
help='Choose an available local search method to improve output paths')
args = parser.parse_args()
evaluation = Evaluation.minmax()
loader = TSPLIBLoader(args.problem)
initializer = Initializer.fixed_state(args.initial_state)
local_search = LocalSearchFactory.build(args.local_search)
stop_criterion = StopCriterion.iteration_limit(args.max_iterations)
taco = TeamAntColonyOptimization(args.num_teams, args.team_size, initializer,
evaluation)
track_base = 'm{}_{}minmax_track.png'
solution_base = '{}2opt_m{}_{}minmax.png'
solutions, score, track = taco.optimize(loader, stop_criterion, local_search)
plot_evolution_track(track_base.format(args.team_size, score), track)
loader.plot(solution_base.format(loader.name, args.team_size, score),
solutions=solutions)
def test(model, selected_step):
with tf.Graph().as_default():
total_test_num = model.val_dataset.exact_num
dms, poses, cfgs, coms, names = model.batch_input_test(
model.val_dataset)
model.test(dms, poses, cfgs, coms, reuse_variables=None)
# dms, poses, names = model.batch_input_test(model.val_dataset)
# model.test(dms, poses, reuse_variables=None)
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False))
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
if selected_step is not None:
checkpoint_path = os.path.join(model.train_dir,
'model.ckpt-%d' % selected_step)
saver = tf.train.Saver(tf.global_variables())
saver.restore(sess, checkpoint_path)
print('[test_model]model has been resotored from %s' %
checkpoint_path)
tf.train.start_queue_runners(sess=sess)
summary_writer = tf.summary.FileWriter(model.summary_dir + '_' +
model.val_dataset.subset,
graph=sess.graph)
res_path = os.path.join(
model.train_dir,
'%s-%s-result' % (model.val_dataset.subset, datetime.now()))
res_path = res_path.replace(' ', '_')
res_path = res_path.replace(':', '_')
res_txt_path = res_path + '.txt'
if os.path.exists(res_txt_path):
os.remove(res_txt_path)
err_path = res_path + '_error.txt'
f = open(res_txt_path, 'w')
# res_vid_path = res_path+'.avi'
# codec = cv2.cv.CV_FOURCC('X','V','I','D')
# the output size is defined by the visualization tool of matplotlib
# vid = cv2.VideoWriter(res_vid_path, codec, 25, (640, 480))
print('[test_model]begin test')
test_num = 0
step = 0
maxJntError = []
while True:
start_time = time.time()
try:
gt_vals, xyz_vals, name_vals = model.do_test(
sess, summary_writer, step, names)
except tf.errors.OutOfRangeError:
print('run out of range')
break
duration = time.time() - start_time
for xyz_val, gt_val, name_val in zip(xyz_vals, gt_vals, name_vals):
maxJntError.append(Evaluation.maxJntError(xyz_val, gt_val))
xyz_val = xyz_val.tolist()
res_str = '%s\t%s\n' % (name_val, '\t'.join(
format(pt, '.4f') for pt in xyz_val))
res_str = res_str.replace('/', '\\')
f.write(res_str)
# vid.write(vis_val)
test_num += 1
if test_num >= total_test_num:
print('finish test')
f.close()
Evaluation.plotError(maxJntError, err_path)
return
f.flush()
if step % 101 == 0:
print('[%s]: %d/%d computed, with %.2fs' %
(datetime.now(), step, model.max_steps, duration))
step += 1
print('finish test')
f.close()
Evaluation.plotError(maxJntError, 'result.txt')
def test(self, test_stream, desc='dummy', modelIdx='-1'):
cache_dir = os.path.join(
globalConfig.model_dir,
'gan_render/%s_%s' % (globalConfig.dataset, desc))
if not os.path.exists(cache_dir):
raise IOError('%s does not exists' % cache_dir)
img_dir = os.path.join(cache_dir, 'test_img')
if os.path.exists(img_dir):
shutil.rmtree(img_dir)
os.mkdir(img_dir)
model_path = os.path.join(cache_dir, 'params', modelIdx)
self.loadParam(model_path)
idx, total_time, recons_err = 0, 0, 0
maxJntError = []
codec = cv2.cv.CV_FOURCC('X', 'V', 'I', 'D')
vid = cv2.VideoWriter(os.path.join(img_dir, 'res.avi'), codec, 25,
(128 * 4, 128))
for skel, orig, trans, com, depth in\
test_stream.iterate(batchsize=1, shuffle=False):
noise = np.zeros((1, self.pose_z_dim), np.float32)
reco_depth = self.render_fn(skel, orig, noise)
reco_pose = self.vae_reco_fn(skel, noise)
pose = self.resumePose(reco_pose[0], orig[0])
fake_img = self.visPair(reco_depth[0], pose, trans[0], com[0],
50.0)
gt_pose = self.resumePose(skel[0], orig[0])
real_img = self.visPair(depth[0], gt_pose, trans[0], com[0], 50.0)
# real calculation part
start_time = time.time()
est_z = self.z_est_fn(depth)
est_z.shape = (23, )
est_z, est_orig = est_z[:20], est_z[20:]
est_z.shape = (1, 20)
est_orig.shape = (1, 3)
est_pose = self.pose_decode_fn(est_z)
end_time = time.time()
est_depth = self.render_fn(est_pose, est_orig, noise)
est_pose = self.resumePose(est_pose[0], est_orig[0])
est_img = self.visPair(est_depth[0], est_pose, trans[0], com[0],
50.0)
com_img = self.visPair(depth[0], est_pose, trans[0], com[0], 50.0)
recons_err += (abs(reco_depth - depth)).mean()
recons_depth = np.hstack((real_img, fake_img, est_img, com_img))
# cv2.imwrite(os.path.join(img_dir,'%d_0.jpg'%(idx)),\
# recons_depth.astype('uint8'))
idx += 1
maxJntError.append(Evaluation.maxJntError(gt_pose, est_pose))
total_time += end_time - start_time
vid.write(recons_depth.astype('uint8'))
print 'average running time = %fs' % (total_time / idx)
print 'average reconstruction error = %f' % (recons_err / idx)
fig_path = os.path.join(img_dir, 'maxError.txt')
Evaluation.plotError(maxJntError, fig_path)
vid.release()
def test(model, selected_step):
with tf.Graph().as_default():
total_test_num = model.val_dataset.exact_num
#获得原始数据，唯一的处理就是把深度图crop了一下
dms, poses, cfgs, coms, names = model.batch_input_test(
model.val_dataset)
model.test(dms, poses, cfgs, coms, reuse_variables=None)
# dms, poses, names = model.batch_input_test(model.val_dataset)
# model.test(dms, poses, reuse_variables=None)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True,
log_device_placement=False))
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
if selected_step is not None:
checkpoint_path = os.path.join(model.train_dir,
'model.ckpt-%d' % selected_step)
saver = tf.train.Saver(tf.global_variables())
saver.restore(sess, checkpoint_path)
print('[test_model]model has been resotored from %s' %
checkpoint_path)
tf.train.start_queue_runners(sess=sess)
summary_writer = tf.summary.FileWriter(model.summary_dir + '_' +
model.val_dataset.subset,
graph=sess.graph)
res_path = os.path.join(
model.train_dir,
'%s-%s-result' % (model.val_dataset.subset, datetime.now()))
res_path = res_path.replace(' ', '_')
res_txt_path = res_path + '.txt'
if os.path.exists(res_txt_path):
os.remove(res_txt_path)
err_path = res_path + '_error.txt'
f = open(res_txt_path, 'w')
# res_vid_path = res_path+'.avi'
# codec = cv2.cv.CV_FOURCC('X','V','I','D')
# the output size is defined by the visualization tool of matplotlib
# vid = cv2.VideoWriter(res_vid_path, codec, 25, (640, 480))
print('[test_model]begin test')
test_num = 0
step = 0
meanJntError = []
xyz_val_list = []
xyz_gt_list = []
while True:
start_time = time.time()
try:
gt_vals, xyz_vals, name_vals, val_dms_v, gt_uvd_pts_v, uvd_pts_v = model.do_test(
sess, summary_writer, step, names)
except tf.errors.OutOfRangeError:
print('run out of range')
break
duration = time.time() - start_time
for xyz_val, gt_val, name_val, val_dms_v1, gt_uvd_pts_v1, uvd_pts_v1 in zip(
xyz_vals, gt_vals, name_vals, val_dms_v, gt_uvd_pts_v,
uvd_pts_v):
#maxJntError.append(Evaluation.maxJntError(xyz_val, gt_val))
#figure_joint_skeleton2(np.squeeze(val_dms_v1),gt_uvd_pts_v1,test_num,"gt")#.savefig("/home/chen/Documents/denseReg-master/exp/train_cache/saved/pretrain/image/"+str(test_num).zfill(7)+"_gt.png")
#figure_joint_skeleton2(np.squeeze(val_dms_v1), uvd_pts_v1,test_num,"pt")#.savefig("/home/chen/Documents/denseReg-master/exp/train_cache/saved/pretrain/image/"+str(test_num).zfill(7)+"_pt.png")
meanJntError.append(Evaluation.meanJntError(xyz_val, gt_val))
xyz_val_list.append(xyz_val[np.newaxis, :])
xyz_gt_list.append(gt_val[np.newaxis, :])
xyz_val = xyz_val.tolist()
res_str = '%s\t%s\n' % (name_val, '\t'.join(
format(pt, '.4f') for pt in xyz_val))
res_str = res_str.replace('/', '\\')
f.write(res_str)
# vid.write(vis_val)
test_num += 1
if test_num >= total_test_num:
xyz_gt_list = np.concatenate(xyz_gt_list, axis=0)
np.savetxt("xyz_gt_test_5.txt", xyz_gt_list, fmt="%.5f")
np.savetxt("xyz_gt_test_all.txt",
xyz_gt_list) # 缺省按照'%.18e'格式保存数据，以空格分隔
xyz_val_list = np.concatenate(xyz_val_list, axis=0)
np.savetxt("xyz_val_test_5.txt", xyz_val_list, fmt="%.5f")
np.savetxt("xyz_val_test_all.txt",
xyz_val_list) # 缺省按照'%.18e'格式保存数据，以空格分隔
#xyz_val_list2 = np.loadtxt("a.txt")
print('finish test')
f.close()
Evaluation.plotError(meanJntError, err_path)
return
f.flush()
if step % 101 == 0:
print('[%s]: %d/%d computed, with %.2fs' %
(datetime.now(), step, model.max_steps, duration))
step += 1