type=float,
default=0.0001,
help='could be changed')
parser.add_argument('--ADAM_BETA_D',
type=float,
default=0.5,
help='could be changed')
parser.add_argument('--ADAM_BETA_G',
type=float,
default=0.5,
help='could be changed')
FLAGS = parser.parse_args()
# 这个函数把两个参数对应的数据集放进来
Train_examples = ioUtil.load_examples(FLAGS.train_hdf5, 'names')
Test_examples = ioUtil.load_examples(FLAGS.test_hdf5, 'names')
'''重新排列数据'''
Train_data = ioUtil.arrange_datas(Train_examples)
Test_data = ioUtil.arrange_datas(Test_examples)
############# FALG things #################################
FLAGS.point_num_in = Train_examples.skeleton_in.shape[
1] # shape gives nums in every dims
FLAGS.point_num_out = Train_examples.pointSet_out.shape[1]
FLAGS.example_num = Train_examples.skeleton_in.shape[0]
EXAMPLE_NUM = FLAGS.example_num
TRAINING_EPOCHES = FLAGS.epoch
yy = tf.expand_dims(y, -1) # [batch_size, npts, 3]
yy = tf.tile(yy, tf.stack([1, 1, 1, nump_x]))
yy = tf.transpose(yy, perm=[0, 3, 2, 1]) # 交换张量的不同维度相当于1和3维的转置
diff = tf.subtract(xx, yy) # 做差，xx中每个点和yy中每个点的差
square_diff = tf.square(diff) # 平方
square_dist = tf.reduce_sum(square_diff, 2) # 平方和，特征维求平方和，降维
return square_dist
# 这个函数把两个参数对应的数据集放进来
Test_examples = ioUtil.load_examples('../data_hdf5/horse_seg_test.hdf5', 'names')
#segmentation tags
Tags = txtReader.read_txt("../data_hdf5/seg.txt")
tags = np.squeeze(Tags)
shape_in = Test_examples.pointSet_out
inputpoints = shape_in[0:1, ...]
outputpoints = shape_in[7:8, ...]
#inputpoints = np.random.rand(1, 10, 3)
#transform = np.random.rand(1, 10, 3)
#outputpoints = inputpoints + transform
# FLAGS
FLAGS = collections.namedtuple("FLAGS", "gpu, batch_size, point_num_out, point_num_in, range_max, radiusScal")
FLAGS.gpu = 0
#print('names : ', nametosave[0])
ioUtil.output_point_cloud_ply( Predicted_xyz1, nametosave, output_dir,
'Ep' + '_predicted_' + 'X1')
#ioUtil.output_point_cloud_ply( Predicted_xyz4, nametosave, output_dir,
# 'Ep' + '_predicted_' + 'X4')
#ioUtil.output_point_cloud_ply( Predicted_xyz8, nametosave, output_dir,
# 'Ep' + '_predicted_' + 'X8')
tf.Session.close
Predicted_xyz = np.squeeze(Predicted_xyz1)
return Predicted_xyz
# main function
if __name__ == '__main__':
modelPath = 'myNet/trained_models/'
data = ioUtil.load_examples('Skeletons/horse.hdf5', 'names')
Y = data.skeleton_in
N = Y.shape[0]
D = Y.shape[1]
pointSet_in = Y[0,:] # [19 * 3]
nametosave = data.names[0]
print(type(pointSet_in))
mustSavePly = True
with tf.Graph().as_default():
model = load_model()
# check point: 二进制文件，它包含的权重变量，biases变量和其他变量
config = tf.ConfigProto(allow_soft_placement = True)
sess = tf.Session(config = config)
#metaPath = modelPath + 'epoch_200.ckpt.meta'
print('\t locate results:\n', results)
#main_ax.scatter(xx_ske[:,0], xx_ske[:,1], s=50, alpha=0.3)
main_ax.scatter(final_X[0], final_X[1], s=30, alpha=0.3)
plt.draw()
nav_init = np.array(final_X).reshape([-1,2])
np.random.seed(42) #用于指定随机数生成时所用算法开始的整数值
gpflow.settings.numerics.quadrature = 'error' # throw error if quadrature is used for kernel expectations
if __name__ == '__main__':
data = ioUtil.load_examples(FLAGS.hdf5, 'names')
skeletons = data.skeleton_in
pointSet = data.pointSet_out
print('Number of points * number of dimensions', skeletons.shape)
# create model
Q = 2 #pca降维的维度
M = 10 #支撑点的数目
N = skeletons.shape[0]
D = skeletons.shape[1]
Y = np.reshape(skeletons, (N, D*3))
pointSet = np.reshape(pointSet, (N, 2048, 3))
# PCA降维，提取前面5维向量作为基 100*5
X_mean = gpflow.gplvm.PCA_reduce(Y, Q)
# GPLVM