loss computing in sample_validation_data

[kingwmk]

Hi, ret_x_seq, loss = sample_validation_data(x_seq, PedsList_seq, grid_seq, args, net, lookup_seq, numPedsList_seq, dataloader)

loss = Gaussian2DLikelihood(out_[0].view(1, out_.size()[1], out_.size()[2]), x_seq[tstep].view(1, numx_seq, 2), [Pedlist[tstep]], look_up)

why use x_seq[tstep] to compute loss, why not use x_seq[tstep+1].?

[quancore]

As far as I remember, it is because I am vectorizing it (subtract the first frame to all frames so every time first position is 0) however, I am not remembering clearly. I need to print variables and length of each sequence but I am unable to debug right now. Can you do it?

[kingwmk]

Ok, I will spend some time learning your code. By the way, May i ask that which result/rank was computed by this code in the World Plane Human-Human Dataset challenge(http://trajnet.stanford.edu/result.php?cid=1). social lstm_v2 or v3 ? Thanks!

[quancore]

both of them.

[william-yan]

I wonder the code is run under windows or ubuntu?

[quancore]

I have experimented on Ubuntu 16.04, never tried on Windows.

[william-yan]

I wonder why you clone the x_seq to ret_x_seq in the sample function under test.py? That means the error between predict point the to original point is zero. How did you get the simulation result?

[quancore]

@william-yan please check the issue https://github.com/quancore/social-lstm/issues/6

[JackFram]

During the training, I think the model is actually trying to predict the current location instead of the next time step given the current location.