search

opendilab / SmartRefine

[CVPR 2024] SmartRefine: A Scenario-Adaptive Refinement Framework for Efficient Motion Prediction
Apache License 2.0
97 stars 7 forks source link

Training details #3

Open Harry333777 opened 3 months ago

Harry333777 commented 3 months ago

We greatly appreciate the author's contribution. We are attempting to replicate this work, but we have encountered obstacles during the training process. The dataset is 'reprocessed' at the start of each training session.Below is the list of files from our preprocessed dataset and the training command. L`B65X$3OW0GCFZXDG)P61X TC}8 H6(6U 7$R MX9)~9CT

youngzhou1999 commented 3 months ago

Hi, thanks for running our code.

Sorry for the confusion. The downloaded pkl files (as the p1_root argument) are the prediction results by hivt (we refer to 'trajectory generation backbone' in our original paper). For training, it's necessary to process raw data to pkl files for our refinement.

If you do have your own preprocessed dataset, you can edit this function (like return directly) https://github.com/opendilab/SmartRefine/blob/main/datamodules/argoverse_v1_datamodule.py#L37, so the pipeline won't process data again. Also, this may need some extra editing in the dataset.py file case by case.

Hope it helps. Feel free to ask if you have any further concerns.

Family-Liao commented 3 months ago

have you met this error？ @Harry333777 QQ图片20240604102950 QQ图片20240604102956