rose-stl-lab / deepstpp Goto Github PK

If you change the value of the lookback variable from the default of 20 in the SlidingWindowWrapper, it results in a dimensional mismatch in the loss function of the model at following line

  File "C:\Users\Adam\Documents\PyCharmProjects\DeepSTPP-master\src\util.py", line 145, in train
    loss, sll, tll = model.loss(st_x, st_y)
  File "C:\Users\Adam\Documents\PyCharmProjects\DeepSTPP-master\src\model.py", line 247, in loss
    sll = torch.log(s_intensity(w_i, b_i, t_ti, s_diff, inv_var))
  File "C:\Users\Adam\Documents\PyCharmProjects\DeepSTPP-master\src\model.py", line 185, in s_intensity
    v_i = w_i * torch.exp(-b_i * t_ti)
RuntimeError: The size of tensor a (40) must match the size of tensor b (30) at non-singleton dimension 1

I'm using my customized data and find the total training loss and validation loss may become negative when the iteration number increases, and I noticed the negative loss also appeared in the provided notebook file. Does it really matter? Can you explain a little bit? Thank you so much

I believe there is a bug in the code that plots the lamb results somewhere.

Using the code provided in the text notebooks, I have used a simple hand-crafted dataset, where I have added a very small amount of gaussian noise to the sequence shown in the first image (green images are the ones to be predicted).

With normalisation turned off, the resultant lamb diagram looks like the x-y axis is flipped.

With normalisation turned on, again it looks flipped but also the scaling is off,

Hi, would it be possible to make the earthquake data folder available? Just want to try out the tutorial. Thanks