Adaickalavan
commented
2 years ago Hi @TianQi-777,
-
You may consider using the Waymo dataset
waymo_open_dataset_motion_v_1_1_0/uncompressed/scenario. Convert the Waymo dataset to an equivalent SMARTS dataset, and then use the equivalent SMARTS dataset to train your model.- In order to browse and replay a Waymo dataset, use the
scl waymo overviewandscl waymo previewcommands. - In order to convert a Waymo dataset into an equivalent SMARTS dataset, do the following:
- First, convert the Waymo dataset into an equivalent SMARTS scenario by using the
scl waymo exportcommand. - Then, feed the equivalent SMARTS scenario into the traffic_histories_to_observations.py script to generate equivalent SMARTS dataset.
- First, convert the Waymo dataset into an equivalent SMARTS scenario by using the
- In order to browse and replay a Waymo dataset, use the
-
Please note that we are unable to provide or distribute a Waymo dataset which has already been converted into an equivalent SMARTS dataset, due to Waymo dataset's terms and conditions.
-
A subset of Waymo dataset which have useful and interesting trajectories are provided here. This subset may be used to focus the training. The provided data conversion tool can be used to convert these datasets into an equivalent SMARTS dataset.
-
Please note that Track-2 is an offline learning track. Here, participants are required to submit their training code for the competition organizers to re-train a new model from scratch using hidden SMARTS dataset. The newly trained model will then be evaluated.
-
Once you have developed a solution for Track2, you are required to upload a Docker image of your training code to DockerHub. Do not include any pre-trained models in your uploaded Docker image. During evaluation in the competition organizer's server, your Docker image will be pulled and executed as follows.
$ docker run --rm -it \ --volume=<path>/offline_dataset:/SMARTS/competition/offline_dataset --volume=<path>/output:/SMARTS/competition/output <username/imagename:version>The competition organizer will supply hidden SMARTS dataset to the container via a mapped volume at
/SMARTS/competition/offline_datasetdirectory. Upon executing the above command, the training code inside the container should start training a new model from scratch by using the newly supplied SMARTS dataset from the container path/SMARTS/competition/offline_dataset. The resulting re-trained model from the container will then be copied out and evaluated in a manner identical to that of Track1.
TianQi-777
soheilalizadeh
tunglm2203
saulfield
High Level Description 2022 NeurIPS Driving SMARTS Competition - track 2 - Which datasets should we use to train the model?
Desired SMARTS version 0.4.12
Operating System ubuntu 20.04
Problems 2022 NeurIPS Driving SMARTS Competition - track 2 - Which datasets should we use to train the model? In the Data and Model section of README: Track-2: Offline Learning, the organizer said that the training data should use Waymo Open Motion or Next Generation Simulation (NGSIM). However, we found that the data of Waymo Open Motion is very large (>1T) and contains many scenarios. If Waymo Open Motion is used for training, which scenario dataset should be used? Also, we noticed that in the Dockerfile, DockerHub, Training, and Evaluation section (8), the organizer said New offline dataset is made available to the container via a mapped volume at /SMARTS/competition/offline_dataset directory. Does this mean that we should use the data in the container as training data, and if so, the description of Data and Model section is confusing and misleading.