Questions about Navtrain Dataset and warmup_test about navsim HOT 3 CLOSED

Hi @sakjdas,

Thanks for your feedback. Please find the answers to your questions below.

Missing files when downloading navtrain

Thanks for pointing out that the files current_split_5.tgz and history_split_5.tgz are missing on the aws download server. We will look into this and come back to you.

Why is there no pickle file for Navtrain data?

As described in the docs, Navtrain is just a fraction of trainval that was obtained by filtering it for challenging scenarios. Moreover, it also includes overlapping scenes, while trainval consists only of non-overlapping scenes.
Then Navtrain download script, simply loads the sensor data for these filtered scenes.
As Navtrain is a fraction of trainval, no additional .pkl file is needed for the logs, since all the scenes are already included in the .pkl file for trainval (which is not very large). So you can just download the .pkl file for trainval.
With that being said, given that the file is small, we might add it to the download script to make the setup even more convenient. Thus, thanks for your feedback.

Current vs. History Frames

The script to load Navtrain downloads sensor-data for current and history frames. While current means the sensor data for t=0 of each scenes, history refers to the sensor data of past timesteps. Consequently, if you want to run an agent, that does not need access to past sensor data, you don't need the history sensor data.

How is mini split into train, val, test ?

mini is a dataset-split that contains log files for training, validation and testing. Our training pipeline automatically ensures, that the respective logs are loaded into the training and validation data loaders (see splitter). So if so far you used split=mini and scene_filter=all_scenes(which is the default) for training, then you trained on these logs. The test-logs of mini are left untouched in training.

test-logs in mini vs. warmup_test_e2e

The warmup_test_e2e scene-filter filters the test-logs from the mini dataset for challenging scenarios by employing the same strategy that is used to generate navtrain from trainval. Thus it subsamples these logs.

Let me know if this answers your questions.

from navsim.

Thank you for your patient replies; they have addressed all of my questions!

from navsim.

Hey @sakjdas, we don’t have a current_split_5 or history_split_5. It was a bug in the for-loop of the download script. Thanks for pointing this out! 

We updated the download_navsim.sh script, which additionally downloads the logs now. 

Best, Daniel

from navsim.