Step 1: Prepare the dataset. Our dataset download process follows the HumanMotionQA benchmark.
Step 2: Train.
jac-run scripts/trainval-humanmotion.py --desc experiments/desc_neuro_codex_humanmotion.py \
--datadir $datadir --data-split-file $data_split_file --vocab $vocab --no_gt_segments \
--parsed-train-path $train --parsed-test-path $test \
--validation-interval 1 --save-interval 1 --lr 0.0005 --epochs 5000 --batch-size 4
Here, $datadir
is the path to BABEL-QA
, $data_split_file
is the path to split_question_ids.json
. You can download the files for $vocab
and parsed $train
, $test
from this download link.
Step 3: Evaluate.
jac-run scripts/trainval-humanmotion.py --desc experiments/desc_neuro_codex_humanmotion.py \
--datadir $datadir --data-split-file $data_split_file --vocab $vocab --no_gt_segments \
--parsed-train-path $train --parsed-test-path $test \
--batch-size 4 --load $load_path --evaluate
You can find our trained checkpoint for $load_path
from this download link.