QCNet training code on ETRI Trajectory Dataset (ETD)

We slightly modified the original source code (https://github.com/ZikangZhou/QCNet) to make it compatible with our ETD. The deep learning environment can be easily created following the instruction (https://github.com/d1024choi/QCNetDocker). In this guide, we would like to walk you through how to train and test QCNet on ETD. If you need more details about ETD, visit the corresponding repository (https://github.com/d1024choi/ETRITrajPredChallenage).

1.Convert Argoverse2-type dataset into QCNet-compatible one

Initially, you are given two sets, one for training ('train' folder) and the other for test ('test_maked' folder). Convert scene sample files in the folders to QCNet-compatible ones with 'argoverse2_to_qcnet_format.py' in https://github.com/d1024choi/ETRITrajPredChallenage.

2.Split Dataset

To train and validate QCNet, we split samples in 'train' folder into two groups, one for training and the other for validation, according to 'train_list.txt' and 'val_list.txt'. To calculate the average FLOPs for QCNet, we copy test samples in 'test_masked' folder according to 'test_flops_list.txt'. Eventually, after going through all the processes, we have the following four folders:

• 'train_qcnet' : converted samples from 'train' folder to train QCNet
• 'val_qcnet' : converted samples from 'train' folder to validate QCNet
• 'test_qcnet' : converted samples from 'test_masked' to test QCNet
• 'test_flops_qcnet' : converted samples from 'test_masked' to calculate FLOPs of QCNet

3.Train

Run the following command:

python train.py

Before running the training, you must properly set dataset location parameters like

parser.add_argument('--root', type=str, default="/workspace/av2format")
parser.add_argument('--train_processed_dir', type=str, default="train_qcnet")
parser.add_argument('--val_processed_dir', type=str, default="val_qcnet")
parser.add_argument('--test_processed_dir', type=str, default="test_qcnet")

We stored our data samples with the following data folder structure

av2format --- train_qcnet
            |
            - val_qcnet
            |
            - test_qcnet
            |
            - test_flops_qcnet

You can control parameters for the training by editing train.py. For example, by setting --device to 8, you can use 8 GPUs for multi-gpu training.

Once you start training, a folder by the name of 'version_$number$' is automatically created and the intermediate trained parameters are stored in the folder.

4.Test

Run the following command:

python test.py

When testing your trained model, please set --batch_size to 1 to see the average time spent per scene. Once you start testing, a folder by the name of 'version_$number$/test_results' is automatically created and the prediction results are stored in the folder. You can submit the prediction results for the competition leaderboard.

5.Visualization

You can visualize your predictions by running

python visualization.py

6. Calculate FLOPs

You can calculate FLOPs of QCNet by running

python calculate_FLOPs.py

The following parameters must be properly set like

    parser.add_argument('--root', type=str, default="/home/dooseop/DATASET/ETRI/av2format")
    parser.add_argument('--train_batch_size', type=int, default=1)
    parser.add_argument('--val_batch_size', type=int, default=1)
    parser.add_argument('--test_batch_size', type=int, default=1)
    parser.add_argument('--train_processed_dir', type=str, default="train_qcnet")
    parser.add_argument('--val_processed_dir', type=str, default="val_qcnet")
    parser.add_argument('--test_processed_dir', type=str, default="test_flops_qcnet")