PRETRAINING.md

Pre-training

We provide MultiMAE pre-training scripts on (multi-modal) ImageNet-1K.
Please check SETUP.md for set-up instructions first.

All our models are pre-trained on a single node with 8 A100 GPUs.

To pre-train MultiMAE on 8 GPUs using default settings, run:

OMP_NUM_THREADS=1 torchrun --nproc_per_node=8 run_pretraining_multimae.py \
--config cfgs/pretrain/multimae-b_98_rgb+-depth-semseg_1600e.yaml \
--data_path /path/to/imagenet/train

Modifying configs

The training scripts support both YAML config files and command-line arguments. See here for pre-training config files.

To modify pre-training settings, either edit / add config files or provide additional command-line arguments.

For a list of possible arguments, see run_pretraining_multimae.py.

ℹ️ Config files arguments override default arguments, and command-line arguments override both default arguments and config arguments.

⚠️ When changing settings, make sure to modify the output_dir and wandb_run_name (if logging is activated) to reflect the changes.

Experiment logging

To activate logging to Weights & Biases, either edit the config files or use the --log_wandb flag along with any other extra logging arguments.