Skip to content

Latest commit

 

History

History
20 lines (14 loc) · 1.36 KB

TRAIN_UPSTREAM_CL.md

File metadata and controls

20 lines (14 loc) · 1.36 KB
Raw

Training Upstream Continual Learning

Continual Learning experiments can be run using the file src/train/train_upstream_continual_learning.py. The following arguments need to be specified:

  • --encoder_name : Name of base vision-language encoder to use. Available VL encoders can be viewed here.
  • --pretrained_model_name : Name of pre-trained model checkpoint name to load for encoder.
  • --ordered_cl_tasks : Order of CL tasks. Tasks available for upstream CL can be seen here.
  • --cl_algorithm : Name of CL algorithm to use. Some algorithms may include additional algorithm-specific arguments
  • --climb_data_dir : Directory where the training data for all CLiMB tasks is located
  • --do_train and --do_eval (latter if only doing Knowledge Transfer/Forgetting evaluation of an already CL-trained model)
  • --output_dir : Directory where experiment outputs will be stored.
  • --batch_size

Flowchart of training upstream CL

The above flowchart shows the steps in training a model on multimodal tasks in a Continual Learning setting.

Sample CL training scripts for training a ViLT encoder on the task order VQAv2 -> NLVR2 -> SNLI-VE -> VCR, using a variety of CL algorithms, can be viewed here.