Problem in loss function

[mephisto28]

minRF/advanced/main.py

Line 60 in 72feb0c

 batchwise_mse = ((z1 - x - vtheta) ** 2).mean(dim=list(range(1, len(x.shape)))) 

I guess the loss should be batchwise_mse = ((z1 - x - vtheta * texp) ** 2).mean(dim=list(range(1, len(x.shape))))... is it?

The current code says the predict target in training should be the distance other than velocity. If not changing the training target the inference code could be modified by dividing the predicted value by timestep to gain the velocity.

[Simo Ryu]

Sorry what?

[mephisto28]

Sorry what?

If mse = (model_pred - x) ** 2, it should be model_pred = z1 - vtheta * t instead of model_pred = z1 - vtheta ... is it?

[Simo Ryu]

nah