docs/Evaluation.md

Evaluation

In LLaVA-1.5, we evaluate models on a diverse set of 12 benchmarks. To ensure the reproducibility, we evaluate the models with greedy decoding. We do not evaluate using beam search to make the inference process consistent with the chat demo of real-time outputs.

Currently, we mostly utilize the official toolkit or server for the evaluation.

Evaluate on Custom Datasets

You can evaluate LLaVA on your custom datasets by converting your dataset to LLaVA's jsonl format, and evaluate using model_vqa.py.

Below we provide a general guideline for evaluating datasets with some common formats.

1. Short-answer (e.g. VQAv2, MME).

<question>
Answer the question using a single word or phrase.

2. Option-only for multiple-choice (e.g. MMBench, SEED-Bench).

<question>
A. <option_1>
B. <option_2>
C. <option_3>
D. <option_4>
Answer with the option's letter from the given choices directly.

3. Natural QA (e.g. LLaVA-Bench, MM-Vet).

No postprocessing is needed.

Scripts

Before preparing task-specific data, you MUST first download eval.zip. It contains custom annotations, scripts, and the prediction files with LLaVA v1.5. Extract to ./playground/data/eval. This also provides a general structure for all datasets.

VQAv2

1. Download test2015 and put it under ./playground/data/eval/vqav2.
2. Multi-GPU inference.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/v1_5/eval/vqav2.sh

3. Submit the results to the evaluation server: ./playground/data/eval/vqav2/answers_upload.

GQA

1. Download the data and evaluation scripts following the official instructions and put under ./playground/data/eval/gqa/data. You may need to modify eval.py as this due to the missing assets in the GQA v1.2 release.
2. Multi-GPU inference.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/v1_5/eval/gqa.sh

VisWiz

1. Download test.json and extract test.zip to test. Put them under ./playground/data/eval/vizwiz.
2. Single-GPU inference.

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/vizwiz.sh

3. Submit the results to the evaluation server: ./playground/data/eval/vizwiz/answers_upload.

ScienceQA

1. Under ./playground/data/eval/scienceqa, download images, pid_splits.json, problems.json from the data/scienceqa folder of the ScienceQA repo.
2. Single-GPU inference and evaluate.

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/sqa.sh

TextVQA

1. Download TextVQA_0.5.1_val.json and images and extract to ./playground/data/eval/textvqa.
2. Single-GPU inference and evaluate.

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/textvqa.sh

POPE

1. Download coco from POPE and put under ./playground/data/eval/pope.
2. Single-GPU inference and evaluate.

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/pope.sh

MME

1. Download the data following the official instructions here.
2. Downloaded images to MME_Benchmark_release_version.
3. put the official eval_tool and MME_Benchmark_release_version under ./playground/data/eval/MME.
4. Single-GPU inference and evaluate.

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/mme.sh

MMBench

1. Download mmbench_dev_20230712.tsv and put under ./playground/data/eval/mmbench.
2. Single-GPU inference.

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/mmbench.sh

3. Submit the results to the evaluation server: ./playground/data/eval/mmbench/answers_upload/mmbench_dev_20230712.

MMBench-CN

1. Download mmbench_dev_cn_20231003.tsv and put under ./playground/data/eval/mmbench.
2. Single-GPU inference.

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/mmbench_cn.sh

3. Submit the results to the evaluation server: ./playground/data/eval/mmbench/answers_upload/mmbench_dev_cn_20231003.

SEED-Bench

1. Following the official instructions to download the images and the videos. Put images under ./playground/data/eval/seed_bench/SEED-Bench-image.
2. Extract the video frame in the middle from the downloaded videos, and put them under ./playground/data/eval/seed_bench/SEED-Bench-video-image. We provide our script extract_video_frames.py modified from the official one.
3. Multiple-GPU inference and evaluate.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/v1_5/eval/seed.sh

4. Optionally, submit the results to the leaderboard: ./playground/data/eval/seed_bench/answers_upload using the official jupyter notebook.

LLaVA-Bench-in-the-Wild

1. Extract contents of llava-bench-in-the-wild to ./playground/data/eval/llava-bench-in-the-wild.
2. Single-GPU inference and evaluate.

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/llavabench.sh

MM-Vet

1. Extract mm-vet.zip to ./playground/data/eval/mmvet.
2. Single-GPU inference.

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/mmvet.sh

3. Evaluate the predictions in ./playground/data/eval/mmvet/results using the official jupyter notebook.

More Benchmarks

Below are awesome benchmarks for multimodal understanding from the research community, that are not initially included in the LLaVA-1.5 release.

Q-Bench

1. Download llvisionqa_dev.json (for dev-subset) and llvisionqa_test.json (for test-subset). Put them under ./playground/data/eval/qbench.
2. Download and extract images and put all the images directly under ./playground/data/eval/qbench/images_llviqionqa.
3. Single-GPU inference (change dev to test for evaluation on test set).

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/qbench.sh dev

4. Submit the results by instruction here: ./playground/data/eval/qbench/llvisionqa_dev_answers.jsonl.

Chinese-Q-Bench

1. Download 质衡-问答-验证集.json (for dev-subset) and 质衡-问答-测试集.json (for test-subset). Put them under ./playground/data/eval/qbench.
2. Download and extract images and put all the images directly under ./playground/data/eval/qbench/images_llviqionqa.
3. Single-GPU inference (change dev to test for evaluation on test set).

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/qbench_zh.sh dev

4. Submit the results by instruction here: ./playground/data/eval/qbench/llvisionqa_zh_dev_answers.jsonl.