seominseok0429
commented
1 year ago Thank you for your interest in our work.
I haven't posted our inference code yet!
It will be uploaded by this week.
If you need to work in advance, you can write inference code in a stacked autoregressive manner.
Thank you!
The following code is located in your valid function:
pred_y = self.ema_model(batch_x, batch_y, t)I have a question: can predictive models take Ground-Truth as input? If the GT is the input of the model in the validation, why not output the GT directly? This means the predictive results in your paper are generated by _batchx and _batchy, so the performance is very good, even with a performance improvement of nearly 250% than SimVP on human datasets. If you directly output the GT, the MSE will become 0! This is ridiculous. Although the test function without GT as input, it simply won't run after the training and validation for the lack of ground-truthpred_y = self.model(batch_x.to(self.device))Your model does not work without GT, this is a fatal problem that your model is completely unable to predict future frames. In other words, if I already know the actual value to be predicted in advance, why do I still need a model to predict？