search

PJLab-ADG / 3DTrans

An open-source codebase for exploring autonomous driving pre-training
https://bobrown.github.io/Team_3DTrans.github.io/
Apache License 2.0
585 stars 72 forks source link

how to get the "subset of unlabeled target domain"? #3

Closed Kailthen closed 1 year ago

Kailthen commented 1 year ago

Hi, GETTING_STARTED_ADA says "Active Domain Adaptation (ADA) task is to pick up a subset..": 1 To perform the manual annotation process, how to export the subset ? 2 After manual annotation process, how to put the labeled data back to the pipeline ?

谢谢

BOBrown commented 1 year ago

Hello: Thanks for your attention to our work.

 As stated in [ADA readme](https://github.com/PJLab-ADG/3DTrans/blob/master/docs/GETTING_STARTED_ADA.md), the Bi3D consists of three stages: 1) pre-training stage, 2) active source domain data, and 3) active target domain data.
 1) you should perform the pre-training stage to obtain the pre-trained ckpt model to support the following stages, according to the corresponding script file: scripts/dist_train.sh.
 2) you should load the above pre-trainined ckpt model to perform the source-domain sampling: scripts/ADA/dist_train_active_source.sh
 3) you should perform the target-domain sampling and re-training bby this scrips: scripts/ADA/dist_train_active.sh

If you still have a problem, please contact us

JiakangYuan commented 1 year ago

We save the labeled subset and the unlabeled subset as pickle files, and reload the dataset. For more details, you can refer to tools/train_utils/train_active_target_utils.py and pcdet/utils/active_learning_utils.py