asymov

Running MT model

• Clone the main branch

• Install dependencies

  !pip install pytorch_lightning --upgrade
  !pip install torchmetrics==0.7
  !pip install hydra-core --upgrade
  !pip install hydra_colorlog --upgrade
  !pip install shortuuid
  !pip install tqdm
  !pip install pandas
  !pip install transformers
  !pip install psutil
  !pip install einops
  !pip install wandb

• Change the directory to TEMOS

  %cd '<dir_path>/asymov/packages/TEMOS'

• Test training code on tiny dataset

  !HYDRA_FULL_ERROR=1 python train_asymov_mt.py experiment=<exp name> user=<shared, darsh etc.> data.splitpath='${path.datasets}/kit-splits-tiny' num_mw_clusters=1000 trainer=<cpu, gpu> trainer.max_epochs=10 model.max_frames=100 model.metrics.recons_types=['naive'] viz_metrics_start_epoch=0 viz_metrics_every_n_epoch=3
  

• Run training code

  !HYDRA_FULL_ERROR=1 python train_asymov_mt.py experiment=<exp name> run_id=<optional> user=<shared, darsh etc.> num_mw_clusters=1000 trainer=gpu model.max_frames=1000 model.metrics.recons_types=['naive', 'naive_no_rep']
  

  other options that might be helpful:

  path.code_dir=<> hydra.run.dir=<> trainer.max_epochs=<> viz_metrics_start_epoch=<> viz_metrics_every_n_epoch=<> callback.last_ckpt.every_n_epochs=<> callback.viz_ckpt.start_epoch=<> callback.viz_ckpt.every_n_epochs=<>
  

‌ ```