Nachiket Deo

It looks like K-means returned an empty cluster. This is very strange and has not happened during any of my training runs. Can you consistently reproduce the error? Were any...

Turns out training from scratch without pre-training with ground truth traversals leads to the same results and makes training simpler. That's why I decided to just drop the pre-training step....

Thanks for the clarifications, this is helpful :) There's just one small issue that ties in to (1): the TrajNet _challenge_ considers only a small subset of all trajectories in...

Yeah, it's a little slow. It takes about a day to run with 6 workers.

The data files are too large to upload to Github. The 'preprocess_data.m' file generates these files from the raw NGSIM US-101 and I-80 text files. These are publicly available datasets.

These are the maneuver classes as described in the paper

Hi, the files may have different names. There should be three files in the US 101 dataset and three in the I-80 dataset, unless they changed the data format.

Thank you for your interest! Yes, a model where the encoding is used to initialize the decoder state could have been used as well.

The decoder generates outputs over the 5 second horizon. The output will have size [batch_size, args['out_length'], 5]

Sorry for the late response. I'm not sure if I'm allowed to host the data. I'll try to include a line to download it in the pre-processing script.