Junru Gu

When we were developing the code, only Argoverse 1 was available. And the evaluation in Argoverse 1 did not require outputting scores (minFDE, Miss Rate), so the implementation of the...

Maybe multiplying these scores to represent a trajectory's score is not reasonable. It would be better to train a score for each trajectory, treating it as a 6-class classification instead...

1. The first initialization of `self.layers` can be removed as it is covered by the second initialization 2. `linear(hidden_states, self.key.weight)` indicates not use the bias of linear layer `self.key`.

The code can not run without GPU. You can use [Colab](https://[colab.google](https://colab.google/)/) or [Kaggle](https://www.kaggle.com/) to run with GPU for free.