how can i plot training figures by Plot.py

[stu-yzZ]

I modified the state vector x = [p_x, p_y, v_x, v_y, a_x, a_y] and also adjusted F, H, and Q accordingly based on the original state vector.

## modify code
m_2d = 6 # dim of state for 2D CA model
F_gen_2d = torch.tensor([
    [1, 0, delta_t_gen, 0, 0.5 * delta_t_gen ** 2, 0],
    [0, 1, 0, delta_t_gen, 0, 0.5 * delta_t_gen ** 2],
    [0, 0, 1, 0, delta_t_gen, 0],
    [0, 0, 0, 1, 0, delta_t_gen],
    [0, 0, 0, 0, 1, 0],
    [0, 0, 0, 0, 0, 1],
]).float()
H_identity_2d = torch.eye(6)
Q_gen_2d = q2 * torch.tensor([
    [1/20*delta_t_gen**5, 0,                 1/8*delta_t_gen**4, 0,                 1/6*delta_t_gen**3, 0],
    [0,                 1/20*delta_t_gen**5, 0,                 1/8*delta_t_gen**4, 0,                 1/6*delta_t_gen**3],
    [1/8*delta_t_gen**4, 0,                 1/3*delta_t_gen**3, 0,                 1/2*delta_t_gen**2, 0],
    [0,                 1/8*delta_t_gen**4, 0,                 1/3*delta_t_gen**3, 0,                 1/2*delta_t_gen**2],
    [1/6*delta_t_gen**3, 0,                 1/2*delta_t_gen**2, 0,                 delta_t_gen,        0],
    [0,                 1/6*delta_t_gen**3, 0,                 1/2*delta_t_gen**2, 0,                 delta_t_gen],
]).float()
R_3_2d = r2 * torch.eye(6)
args.n_steps = 2000
args.n_batch = 40
args.lr = 1e-4
args.wd = 1e-4

## Generation model (CA)
sys_model_gen = SystemModel(F_gen_2d, Q_gen_2d, #H_onlyPos, R_onlyPos, 
                            H_identity_2d, R_3_2d,
                            args.T, args.T_test)
sys_model_gen.InitSequence(m1x_0_2d, m2x_0_gen_2d)# x0 and P0

## Feed model (to KF, KalmanNet) 
 sys_model = SystemModel(F_gen_2d, Q_gen_2d, #H_onlyPos, R_onlyPos, 
                         H_identity_2d, R_3_2d,
                         args.T, args.T_test)
 sys_model.InitSequence(m1x_0_2d, m2x_0_2d)# x0 and P0

However,

1. Q1：I noticed that the MSE (Mean Squared Error) doesn’t perform well during training. I need to inspect the entire training process, so I might need to use the plot.py function. Could you provide a concrete example?

2. Q2：Should normalization be considered across these six dimensions [p_x, p_y, v_x, v_y, a_x, a_y]? The key code I modified is shown above (primarily aligning some state dimensions).

[XiaoyongNI]

Hi there, thanks for your interest in our work.

Q1: Inspecting Training with plot.py
plot.py is primarily for visualizing final results, not the live training process. However, it's straightforward to write your own code to plot the loss against training steps or epochs to inspect performance.

Q2: Normalization Across Dimensions
Yes, normalization is recommended. Without it, the Mean Squared Error (MSE) loss will inherently prioritize optimizing elements with larger magnitudes, potentially leading to unfair optimization across your state variables. Normalization helps ensure each element contributes more equally to the overall error.

Best,
Xiaoyong