Skip to content

Latest commit

 

History

History
62 lines (49 loc) · 2.83 KB

README.md

File metadata and controls

62 lines (49 loc) · 2.83 KB

Argoverse Dataset

Download Argoverse v1.1 dataset

Download Argoverse Motion Forecasting v1.1 dataset and Argoverse HD Maps from https://www.argoverse.org/av1.html

Install Argoverse API

Follow instructions at https://github.com/argoai/argoverse-api#installation

Convert to protobufs

Command usage:

$ python data_sources/argoverse/create_adovehicle_snippets_argoverse.py --input-dir ${ARGOVERSE_DATA_DIR} --output-dir ${OUTPUT_PROTOBUFs_DIR} --protobuf -s -r --smooth

Options include:

  • --protobuf: Save data as protobufs.
  • -s: Select candidate map centerlines close to the agent trajectory, using Argoverse API, as opposed to geofencing.
  • -r: Resample equidistant points along map centerlines.
  • --smooth: Smooth agent trajectory.
  • -v: Visualize data.
  • -l: Maximum number of examples to save.
  • --mode: Whether the data comes from train/val/test set.

Data conversion commands

Convert training data:

$ python data_sources/argoverse/create_adovehicle_snippets_argoverse.py --input-dir ~/argoverse_data/train/data/ --output-dir ~/argoverse_data/train_pb --protobuf -s -r --smooth

Convert validation data:

$ python data_sources/argoverse/create_adovehicle_snippets_argoverse.py --input-dir ~/argoverse_data/val/data/ --output-dir ~/argoverse_data/val_pb --protobuf -s -r --smooth

Convert test data:

$ python data_sources/argoverse/create_adovehicle_snippets_argoverse.py --input-dir ~/argoverse_data/test/data/ --output-dir ~/argoverse_data/test_pb --protobuf -s -r --smooth --mode test

Converted Version

For TRI users, see here to download converted protobuf.

Example training command

The following assumes that you have created a training environment as described here.

Run basic training without images

$ python intent/multiagents/train_pedestrian_trajectory_prediction.py --input-dir /home/ec2-user/converted_protobufs/argoverse/train_pb/ --scene-image-mode none --agent-image-mode none --datasets-name-lists none --use-semantic false --num-workers 32 --batch-size 32 --val-batch-size 32 --val-num-workers 32 --vis-batch-size 32 --vis-num-workers 32 --max-agents 4 --past-timesteps 20 --future-timesteps 30 --past-timestep-size 0.1 --future-timestep-size 0.1 --val-interval 10 --vis-interval 10 --disable-map-decoder false

To train other models, refer to README in its own folder.

  • For instance, to train a hybrid model, refer to here

Augment protobufs

This is usually used for non-standard training, such as language-based prediction.

Language

Command usage:

$ python data_sources/augment_protobuf_with_language.py --source-protobufs-dir [SOURCE_PROTOBUFS_DIR] --augmented-protobufs-dir [AUGMENTED_PROTOBUFS_DIR]