DLAV2024 PROJECT : TEAM SPONGIFLEX

Student Name	Email	Sciper
Martin Rollet	[email protected]	300780
Julien Ars	[email protected]	314545

Github repo : https://github.com/merlebleue/DLAV2024-Spongiflex/tree/main

Milestone 1 : 28th april 2024

Introduction

This project is about prediting precise vehicle trajectory prediction, using the UniTraj framework from VITA lab @ EPFL. For this part, the objective is mainly familiarising ourselves with the framework. We use the provided ptr model, with the default configuration, evaluated using minADE6 error.

Code

The code of the model can be consulted here : ptr.py The configuration is here : ptr.yaml

It consists of the provided code, with some parts we had to fill in. Here is our code :

• Function temporal_attn_fn() :

  ######################## Your code here ########################
  for n in range(agents_emb.shape[2]): #per agent, assuming N is the number of agents
    agents_emb[:,:,n,:] = layer(agents_emb[:,:,n,:], src_key_padding_mask=agent_masks[:,:,n])
  ################################################################

• Functionsocial_attn_fn() :

  ######################## Your code here ########################
  for t in range(agents_emb.shape[0]): #per time step, assuming T is the mnumber of time steps
    agents_emb[t,:,:,:] = layer(agents_emb[t,:,:,:], src_key_padding_mask=agent_masks[:,t,:].permute(1,0))
  ################################################################

• In the function _forward():

  ######################## Your code here ########################
  # Apply temporal attention layers and then the social attention layers on agents_emb, each for L_enc times.
  for i in range(self.L_enc):
    agents_emb = self.temporal_attn_fn(agents_emb, opps_masks, self.temporal_attn_layers[i])
    agents_emb = self.social_attn_fn(agents_emb, opps_masks, self.social_attn_layers[i])
  ################################################################

Results

Sadly, we encountered issues due to the presence of nan values in the dataset. This seemed to arrise from agents that do not exist in the sequence, but we did not manage to get rid of this error.