Open cuihenggang opened 4 years ago
Hi, I have the same problem. Did you end up fixing this issue? Did training until t = 4200 helps?
No luck :(
activation using relu or leakyrelu
did you try to use a larger learning rate like e.g. 1e-3? try to reuse hyperparameters from run_traj.sh maybe it will fix your problem
I have finally figured out the issue. You need to train with the run_traj.sh
script. The default arguments in train.py
don't work. The most important argument is --l2_loss_weight 1
which adds the L2 loss to the generator. Social-GAN needs the L2 loss to train and doesn't work with GAN loss only.
I have finally figured out the issue. You need to train with the
run_traj.sh
script. The default arguments intrain.py
don't work. The most important argument is--l2_loss_weight 1
which adds the L2 loss to the generator. Social-GAN needs the L2 loss to train and doesn't work with GAN loss only.
Hi, I have the same problem. I set --l2_loss_weight
to 1. But the G loss and D loss keep unchanged still(1.386 and 0.693 respectively) and l2 loss keeps changing. Do you know how to fix it?
I am trying to train Social-GAN with the code in the repository, but it looks like the G loss and D loss never change after 1.386 and 0.693.
Also, the validation FDE error is 11.058
Am I doing the training correctly?