Remaining tasks:
1. the percentage of missing points, 非0个数6677/7852 = 0.8503，所以0的比率是15%.
3. modify paper according to pdf.
4. check collective anomalies.
5. 代码部分：
- 多个lstm层
- 结果sim_score的分析，
- 画图： TSNE, attention等。
2. tuning hyperparameters: 目前调了z_dim, h_dim; 还差参数layer_depth； lambda_kl + eta_kl (目前是0.01)
pre-processing: when removing noise: 11 and 2
x[each] = savgol_filter(x[each], 11, 2) # (seq_length,)
hyperparameters for model training:
# Model hyperparameters
T_w = 240 # Window length
batch_size = 1024
n_epochs = 550 # 到550时会收敛，可以固定。
z_dim = 3 # 3,4,5,6
h_dim = 128,256,712 # Number of LSTM units in each direction
layer_depth = 1,2,3 # >1 会报错。
x_dim = 23 #固定
lr = 0.001 #learning rate, 固定
criterion = ELBO_loss or M_ELBO_loss.