Gutianpei / MID

[CVPR2022] Code for CVPR 2022 paper "Stochastic Trajectory Prediction via Motion Indeterminacy Diffusion"
182 stars 30 forks source link

Issue in training with custom dataset #28

Open plchen86157 opened 11 months ago

plchen86157 commented 11 months ago

Hello, @Gutianpei

Thank you for your excellent work and for making your code available!

I am interested in training your model for trajectory prediction involving 'vehicles,' utilizing datasets such as the Argoverse Motion Forecasting Dataset or the NuScenes Dataset. I have made modifications to the 'process_data.py' file, and now the model can be trained on the vehicle dataset.

However, I am encountering an issue where the training loss does not converge, and the ADE remains at 10 meters. I am wondering if there are any modifications that need to be addressed that I may have overlooked. Additionally, is there any specific initialization hypothesis for 'PEDESTRIAN'? For example, should the 'standardization' in 'process_data.py' be adjusted? train_loss