You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We would like to include your work as a key baseline in our research and tried to reproduce the results you stated in the paper. However, after running the code you provided on openreview, we only got a success rate of 37%. Can you help us get a comparable result? We used the following launch script:
#!/bin/bash# Scene IDs
scenes=(
"4ok3usBNeis""5cdEh9F2hJL""6s7QHgap2fW""7MXmsvcQjpJ""a8BtkwhxdRV""BAbdmeyTvMZ""bCPU9suPUw9""bxsVRursffK""CrMo8WxCyVb""cvZr5TUy5C5""Dd4bFSTQ8gi""DYehNKdT76V""eF36g7L6Z9M""GLAQ4DNUx5U""h1zeeAwLh9Z""HY1NcmCgn3n""k1cupFYWXJ6""LT9Jq6dN3Ea""MHPLjHsuG27""mL8ThkuaVTM""mv2HUxq3B53""Nfvxx8J5NCo""p53SfW6mjZe""q3zU7Yy5E5s""q5QZSEeHe5g""QaLdnwvtxbs""qyAac8rV8Zk""svBbv1Pavdk""TEEsavR23oF""VBzV5z6i1WS""wcojb4TFT35""XB4GS9ShBRE""y9hTuugGdiq""yr17PDCnDDW""ziup5kvtCCR""zt1RVoi7PcG"
)
# API keys and org IDs# Please replace with your API keys and org ids, use multiple to avoid hitting rate limits
openai_keys=( "sk-***""sk-***""sk-***" )
openai_orgs=( "org-***""org-***""org-***" )
# Divide scenes among the runs
scenes_per_run=$((${#scenes[@]}/5))# Start runsforiin {0..4};do
start_index=$(( i * scenes_per_run ))
end_index=$(( (i+1) * scenes_per_run -1))# Slice the array to get the scenes for this run
run_scenes=(${scenes[@]:$start_index:$scenes_per_run})
# Convert the array to a comma-separated string
run_scenes=$(printf ",%s""${run_scenes[@]}")
run_scenes=${run_scenes:1}# remove the leading comma
experiment_name="Eval_LFG_${i}"echo"Starting run $i with scenes: $run_scenes"# Set the GPU for this runexport CUDA_VISIBLE_DEVICES=$(($i))# Select OpenAI key and org id for this run
key_index=$(($i%3))# Alternates between 0, 1, 2
openai_key=${openai_keys[$key_index]}
openai_org=${openai_orgs[$key_index]}# Run the script with the appropriate scenes and experiment name
nohup python3 projects/habitat_objectnav/eval_dataset_distributed.py --scene_ids="$run_scenes" --experiment_name="$experiment_name" --openai_key="$openai_key" --openai_org="$openai_org" --max_episodes_per_scene=10 --use_gt_semantics --use_language >"${experiment_name}.out"2>&1&done
The text was updated successfully, but these errors were encountered:
Dear authors of the LFG paper,
We would like to include your work as a key baseline in our research and tried to reproduce the results you stated in the paper. However, after running the code you provided on openreview, we only got a success rate of 37%. Can you help us get a comparable result? We used the following launch script:
The text was updated successfully, but these errors were encountered: