Skip to content

Latest commit

 

History

History
75 lines (55 loc) · 2.36 KB

GET_STARTED.md

File metadata and controls

75 lines (55 loc) · 2.36 KB
Raw

Training

We provide two shell files, start_multigpu.sh and train.sh, to start the training process. One can start the training with

sh train.sh

In train.sh,

[-t] or [--target] determines the experiment name.
[--gpus] controls the number of gpus.

In start_multigpu.sh,

[dataset] determines using which dataset: SemanticKITTI or nuScenes.
[model] means to use which backbone: linkunet, linkencoder, minkunet, spvcnn.

linkencoder is a decoder-free architecture, more efficient.

We backup the full runtime code into runs/. One can recover the runtime code in runs/${dataset}/${model}/${experiment name}/.

Evaluation

Reproduction

For a quick rerpduction, we provide the full run-time code along with the checkpoint.

After finishing the installation and dataset preparation, download the log file in model zoo, and uncompress them into the segmentation/runs folder. Take the cos_x[2x3] as an example. Make a hyperlink to link the data folder to the backup path.

cd runs/cos_x[2x3_g1]/2023-05-23_15:30:26/backup/
ln -s ../../../../data .

The file structure should be like:

LinK/segmentation/runs/
        cos_x[2x3_g1]/
            2023-05-23_15:30:26/
                ious.txt
                backup/
                    data/
                        SemanticKITTI/
                        nuScenes/
                    evaluate.py
                    evaluate.sh # validating on sequence 08
                    test.py
                    test.sh # generating results on test split
                    ...
                checkpoints/
                    max-iou-val.pt
                logging/
                tensorboard/
                summary/
                metainfo/

Then, run the evaluation shell:

chmod +x evaluate.sh
./evaluate.sh

Submission

For submission on SemanticKITTI, add sequence 08 in training set and modify the training_size in configs/semantic_kitti/default.yaml. After training, run

# modify test.sh accrodingly
./test.sh

The results will be stored in submission folder, validate its format with https://github.com/PRBonn/semantic-kitti-api/blob/master/validate_submission.py. If corretly, compress the zip file to the official website.