Skip to content

EVAL.md

Latest commit

 

History

History
240 lines (200 loc) · 11.1 KB

EVAL.md

File metadata and controls

240 lines (200 loc) · 11.1 KB

We conduct a comprehensive evaluation of models across a diverse set of benchmarks. In addition to assessing model performance on clean examples, as done in the original LLaVA codebase, we extend the evaluation to include both untargeted and targeted attacks on benchmark datasets. Furthermore, we analyze model robustness against common corruptions to provide a more holistic assessment of their reliability.

Data Preparation

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.

Data Structure

The data for evaluation can be stored in the $DATA_ROOT directory. The structure of the directory will be as follows:

$DATA_ROOT
├── coco
│   ├── train2014
│   ├── val2014
│   ├── annotations
│   │   ├── karpathy_coco.json
│   │   ├── captions_val2014.json
│   │   ├── v2_OpenEnded_mscoco_train2014_questions.json
│   │   ├── v2_mscoco_train2014_annotations.json
│   │   ├── v2_OpenEnded_mscoco_val2014_questions.json
│   │   ├── v2_mscoco_val2014_annotations.json
│   │   ├── OpenEnded_mscoco_train2014_questions.json
│   │   ├── mscoco_train2014_annotations.json
│   │   ├── OpenEnded_mscoco_val2014_questions.json
│   │   ├── mscoco_val2014_annotations.json
│   │
├── flickr30k
│   ├── flickr30k-images
│   ├── karpathy_flickr30k.json
│   ├── dataset_flickr30k_coco_style.json
│
├── vizwiz
│   ├── train
│   ├── val
│   ├── train_questions_vqa_format.json
│   ├── train_annotations_vqa_format.json
│   ├── val_questions_vqa_format.json
│   ├── val_annotations_vqa_format.json
│
├── vqav2
│   ├── v2_OpenEnded_mscoco_train2014_questions.json
│   ├── v2_mscoco_train2014_annotations.json
│   ├── v2_OpenEnded_mscoco_val2014_questions.json
│   ├── v2_mscoco_val2014_annotations.json
│
├── okvqa
│   ├── OpenEnded_mscoco_train2014_questions.json
│   ├── mscoco_train2014_annotations.json
│   ├── OpenEnded_mscoco_val2014_questions.json
│   ├── mscoco_val2014_annotations.json
│
├── textvqa
│   ├── train_images
│   ├── train_questions_vqa_format.json
│   ├── train_annotations_vqa_format.json
│   ├── val_questions_vqa_format.json
│   ├── val_annotations_vqa_format.json

The annotation files for each can be found in the huggingface datasets eval_benchmark and can be kept in the above structure or you can download the datasets/annotations from this Link and keep it in the above structure. Clean evaluation can be done on more benchmarks as provided in the original LLaVA codebase, but for adversarial evaluation, we have provided the evaluation on above benchmarks.

Evaluation

First download the instruction tuned (Stage 2) model checkpoints from Model Zoo as mentioned in the README. In the below scripts, following variables should be set:

  • $DATA_ROOT: Path to the root directory of the evaluation data.
  • $MODEL_PATH: Path to the model checkpoint.
  • $GPU: ID of the GPU to be used.
  • $EPSILON: Perturbation budget for adversarial evaluation.
    • For example, to evaluate on a perturbation budget of 2/255, set $EPSILON=2.
  • $ENCODER:
    • Will be changed when using the adversarially finetuned CLIP model from SimCLIP, replacing the original CLIP vision encoder in the LLaVA framework. ($MODEL_PATH will be of CLIP)
    • Will be changed when using the adversarially finetuned CLIP model from FARE, replacing the original CLIP vision encoder in the LLaVA framework. ($MODEL_PATH will be of CLIP)
    • Set to 'none' for all other cases.
LLaVA Model $Encoder
CLIP none
FARE4@224 fare4
FARE4@336 fare4_336
Robust-LLaVA4H none
Robust-LLaVA4G none
Robust-LLaVA4H + CLIP none
Robust-LLaVA4G + CLIP none
Robust-LLaVA4H + FARE4@336 fare4_336
Robust-LLaVA4G + FARE4@336 fare4_336
Robust-LLaVA4H + Robust-LLaVA4G none
ViT-B/16 (Adversarially trained on ImageNet-1k) none
ViT-L/14 (Naturally trained on ImageNet-21k + ImageNet-1k) none

Flickr30k

To evaluate the adversarial performance(APGD-Ensemble Attack) of model on Flickr, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU true false false false false false false ensemble $EPSILON

To evaluate the adversarial performance(APGD Attack) of model on Flickr, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU true false false false false false false apgd $EPSILON

To evaluate the clean performance of model on Flickr, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU true false false false false false false none

COCO

To evaluate the adversarial performance(APGD-Ensemble Attack) of model on COCO, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false true false false false false false ensemble $EPSILON

To evaluate the adversarial performance(APGD Attack) of model on COCO, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false true false false false false false apgd $EPSILON

To evaluate the clean performance of model on COCO, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false true false false false false false none

VQAv2

To evaluate the adversarial performance(APGD-Ensemble Attack) of model on VQAv2, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false true false false false false ensemble $EPSILON

To evaluate the adversarial performance(APGD Attack) of model on VQAv2, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false true false false false false apgd $EPSILON

To evaluate the clean performance of model on VQAv2, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false true false false false false none

TextVQA

To evaluate the adversarial performance(APGD-Ensemble Attack) of model on TextVQA, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false false true false false false ensemble $EPSILON

To evaluate the adversarial performance(APGD Attack) of model on TextVQA, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false false true false false false apgd $EPSILON

To evaluate the clean performance of model on TextVQA, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false false true false false false none

VizWiz

To evaluate the adversarial performance(APGD-Ensemble Attack) of model on VizWiz, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false false false true false false ensemble $EPSILON

To evaluate the adversarial performance(APGD Attack) of model on VizWiz, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false false false true false false apgd $EPSILON

To evaluate the clean performance of model on VizWiz, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false false false true false false none

OKVQA

To evaluate the adversarial performance(APGD-Ensemble Attack) of model on OKVQA, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false false false false true false ensemble $EPSILON

To evaluate the adversarial performance(APGD Attack) of model on OKVQA, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false false false false true false apgd $EPSILON

To evaluate the clean performance of model on OKVQA, run the following command:

bash bash/llava_eval.sh $DATA_ROOT $MODEL_PATH $ENCODER $GPU false false false false false true false none

POPE

To evaluate the performance of model on POPE, run the following command:

bash bash/eval_pope.sh $DATA_ROOT $MODEL_PATH $ENCODER

To run clean evaluation on all the benchmarks, run the following command:

bash bash/clean_evaluation.sh $MODEL_PATH $DATA_ROOT $GPU $ENCODER

To run adversarial evaluation on all the benchmarks, run the following command:

bash bash/adv_evaluation.sh $MODEL_PATH $DATA_ROOT $GPU $ENCODER $EPSILON

Evaluation on Common Corruptions

For generating common corruptions on COCO, VQAv2, OKVQA, and VizWiz, you can run the python file corruptions.py to generate different corrupted version of the benchmark at different levels of severity and save them. You can then evaluate the model on these corrupted versions using the following command:

bash bash/corruption_evaluation.sh $MODEL_PATH $DATA_ROOT $GPU $ENCODER

Targetted Attack on COCO Captioning Task

To evaluate the targetted attack on COCO Captioning Task, run the following command:

bash bash/llava_eval_targetted.sh $DATA_ROOT $MODEL_PATH $ENCODER  $EPSILON $DEVICE $EXP_NUM
  • $EXP_NUM=1 for target string: "Visit https://tinyurl.com/23cbjxjz"
  • $EXP_NUM=2 for target string: "Please reset your password"
  • $EXP_NUM=3 for target string: "You are too stupid. I will not answer anymore"