We provde the training codes of three experiments.
- [supervised_ours_mp3d]. The experiments "BiFuse++" in Table 1 (Matterport3D) of our main paper, which is trained under supervised scenario.
- [selfsupervised_ours_panosuncg_spl]. The experiments "BiFuse++ w/ SPL" (PanoSUNCG) in Table 8 of our main paper, which is trained under self-supervised scenario.
- [selfsupervised_ours_panosuncg_capl]. The experiments "BiFuse++ w/ CAPL" (PanoSUNCG) in Table 8 of our main paper, which is trained under self-supervised scenario.
Before starting training, please modify "dataset_path" in config.yaml to specify the path of your dataset. After that, you can use the following commands for training and validation.
python main.py --mode train
python main.py --mode valOur code support multi-gpu and multi-node training.
- You can change the arguments ("devices" and "num_nodes") in config.yaml to specify the number of GPUs and nodes, respectively.
- To switch between Tensorboard and W&B for logging, you can change "logger_type" argument in config.yaml.
- For supervised training, we found multi-gpu or larger batch size do not improve the final performance. So the total batch size is 8 in supervised_ours_mp3d.
- For self-supervised training, a larger batch size can significantly improve the training stability. Hence, the total batch size of selfsupervised_ours_panosuncg_spl and selfsupervised_ours_panosuncg_capl is 8 (bs) x 4 (GPUs) = 32.
For the provided self-supervised experiments, we use 4x Tesla V100 and train the models for 40 hours. The comparison between SPL and CAPL is summarized in our W&B report (link).