Skip to content

Latest commit

 

History

History
26 lines (19 loc) · 1.26 KB

README-HumanMotionQA.md

File metadata and controls

26 lines (19 loc) · 1.26 KB

Left

HumanMotionQA Dataset

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.