About merge_nusc_fullset_pkl.py

[qiuqc1]

merge_nusc_fullset_pkl.py Is the test_pkl = 'data/nuscenes/nuscenes_infos_temporal_test.pkl' mentioned in the file using the pkl of val or the pkl of test?

[qiuqc1]

How long does it take to train 8 A100s with only the Vidar pre-training model for the entire dataset? It takes 7 days for me to put it in.I'm shocked. Let me ask you something. 请问全数据集，仅仅训vidar预训练模型，8张A100要多久呢，我这边放进去要7天。挺震惊的，来问一下

[tomztyang]

train + test. It is safe because we test our model on val split set.

Yes. For 8 A100 GPUs, it takes around 7 days :). For our CVPR version, we use 32 GPUs and around 2 days. Speeding up ViDAR pre-training is now the highest priority. Some novel designs like sparse design or speeding up latent rendering are required, maybe future works?

Best, Zetong