Using ckpt for pretraining

[martiiv]

Hello!

I am currently working on reproducing the results from the AD-PT paper. To save some time I am trying to include a ckpt file when running the pretraining with Pointcontrast file.

sh scripts/PRETRAIN/dist_train_pointcontrast.sh 2 \
--cfg_file ./cfgs/once_models/unsupervised_model/pointcontrast_pvrcnn_res_plus_backbone.yaml \
--batch_size 4 \
--epochs 4 \
--ckpt once_1M_ckpt.pth

However the model fails when trying to load the model state giving me the following error:

Traceback (most recent call last):
  File "train_pointcontrast.py", line 212, in <module>
    main()
  File "train_pointcontrast.py", line 140, in main
    it, start_epoch = model.load_params_with_optimizer(args.ckpt, to_cpu=dist_train, optimizer=optimizer, logger=logger)
  File "../pcdet/models/detectors/detector3d_template.py", line 403, in load_params_with_optimizer
    self._load_state_dict(checkpoint['model_state'], strict=True)
KeyError: 'model_state'

Do I need to save my model state before applying the checkpoint? Do you have a solution for this problem @BOBrown

[martiiv]

Update: The code crashes on line 403 here:

Am I using the wrong ckpt file? Since it doesn't recognize the model_state key

[BOBrown]

@martiiv

Which ckpt file are you using? Please note that we provide two different pre-training methods in our repo.

• Pre-training using PointContrast: ./cfgs/once_models/unsupervised_model/pointcontrast_pvrcnn_res_plus_backbone.yaml
• Pre-training using AD-PT, the code of pre-training using AD-PT has not been released yet

[martiiv]

@BOBrown I am using pointContrast, is the ckpt file for AD-PT pre-training?

[BOBrown]

@martiiv No. The ckpt file for AD-PT is shown in here for three ckpts trained using different sizes of pre-training data (100K, 500K, 1M)