edouardelasalles / stnn Goto Github PK

Could you explain the dataset heat.csv and heat_relations.csv explain what are the rows and columns that you have used or update the readme file as to how it actually works? @edouardelasalles

Hello,

Your code looks great and is very helpful! Sorry if I miss any details, but could you please explain the heat_m dataset?

Thank you in advance for replying!

may I use the code also for more than one dimension like 2D problems as in your paper?

I map the 2D into 1D. I guess from the technical point it is the same, right?!

best

Post Scriptum: does it makes sense when you have also property dimension attached. I imagine that you could also calculate some distances in this multi dimensional property space for getting a reasonable adjacent matrix.

What do you think?

Dear Professor:
I have reading your work , and I think that is a very interesting research. I'm trying to replicate some experiments with your STNN, but in addition to MSE, I need to calculate accuracy, precision and recall. Could you help me to know if with the output data of the model and the prediction I can calculate these metrics. Is the first time that I work with neural networks and I have few experience in this area. Thanks in advance by your time,
Best regard, Yamel.

For example, in this formula, the parameters which need to be optimized are d, Z, Theta0, Theta1.
In the code, I can find the params list, in which model.factors_parameters() is Z, model.decoder.parameters() is d. That means model.dynamic.parameters() is Theta0 and Theta1, but I can't understand how the Thetas participate in the training.
The function

show the method to calculate the content of h(), in which I can find Z_t and W, I guess z_context is correspond to Theta, but I don't know why should it calculate like this.
If you see the issue, can you give me some instructions? I will be very grateful!

Hi! Recently I try to use the dataset pst (2520 * 399, same as the paper) to run the code. But I can't get similar result as those in the paper. Maybe I made some mistakes.So I have some questions to ask you.
1.In the paper you said there are 8 relations in the dataset pst (north, east and so on), does this mean that I should generate 8 dir_relasiton.csv files correspond to the directions in the original dataset? What I'm using is a single file and W_{ij} is 1 if j is in one of these 8 directions of i, Is it wrong?
2.In the paper you said using size T to train and the evaluation of the quality of the model will be made over T + 1 to T + 5 time steps.Does this means, for example, if I use 0-200 to train, I should use 201-205 to test? Last time I used 201~399 to test, the result was a little bit awful.
I'm very sorry to disturb you again. I would appreciate it if you could reply to me. Thanks!

Hello, I am running the code for a dataset (LOBO dataset) containing 7 time series of length 24888 in which each data has depth 8.
My goal is to do regression, I would like to predict the evolution of each time series taking into consideration their spatial correlation. Even though during training the loss gets pretty low, the predicted values are pretty much constant and stay in the middle of the evolution of the ground truth, like a mean value. Do you have any idea why the prediction work so badly?

I am using these parameters:
nz=80 mode=discover batch_size=200 dropout_f=0.05 dropout_d=0.125 lambd=1 lr=3e-3 nhid=8 nlayers=8 l1_rel=3e-6

Dear,
Your code looks great.
Could you please provide the real processed dataset, so that I can reproduce the result, instead of the synthetic dataset?
Many thanks.