# RobustVLM (Foundation Models) via Object-centric Learning


## Table of Contents
- [Installation](#installation)
- [Stage1: Get Object-centric Models](#models)
    - [Dataset](#loading-pretrained-models)
    - [Training](#summary-of-results)

## Installation
Create and activate anaconda environment:
```shell
conda create -n robustclip python==3.11
```
```shell
conda activate robustclip 
```

The code is tested with Python 3.11. To install the required packages, run:
```shell
pip install -r requirements.txt
```

To install the open_clip_torch locally run:
```shell
cd ./open_clip_torch
```
```shell
python setup.py develop
```

## Stage1: Get Object-centric Models

### Dataset
Prepare the ImageNet dataset in a torch.ImageFolder style format:
```
dataset_path
└─imagenet
  └─train
    └─n01440764
        xxxxxx.JPEG
        .....
    └─......
  └─val
    └─n04254680
        xxxxxx.JPEG
        .....
    └─......
```
### Training
- Slot-Attention on 4GPUs
```shell
CUDA_VISIBLE_DEVICES=0,1,2,3 python -m train.training_clip_slots --clip_model_name ViT-L-14 --pretrained openai --dataset imagenet --imagenet_root /.../.../dataset_path/imagenet --template std --output_normalize False --steps 1000000 --warmup 10000 --batch_size 128 --loss l2 --opt adamw --lr 5e-5 --wd 1e-4 --attack pgd --inner_loss l2 --norm linf --eps 4 --iterations_adv 10 --stepsize_adv 1 --wandb False --output_dir ./output_slots --experiment_name SLOTS --log_freq 1000 --eval_freq 1000```
```

The results of reconstruction after slot-attention and ckps are stored in './output_slots/ViT-L-14_openai_imagenet_l2_imagenet_SLOTS_xxxxx'



## Stage2: Training and Evaluation with Object-centric Representations

- SlotVLM<sup>4</sup>
```shell
CUDA_VISIBLE_DEVICES=0,1,2,3 python -m train.adversarial_training_clip_with_object_token --clip_model_name ViT-L-14 --slots_ckp ./ckps/model_slots_step_300000.pt --pretrained openai --dataset imagenet --imagenet_root /path/to/imagenet --template std --output_normalize False --steps 20000 --warmup 1400 --batch_size 128 --loss l2 --opt adamw --lr 1e-5 --wd 1e-4 --attack pgd --inner_loss l2 --norm linf --eps 4 --iterations_adv 10 --stepsize_adv 1 --wandb False --output_dir ./output --experiment_name with_OT --log_freq 10 --eval_freq 10
```

Set `--eps 2` to obtain SlotVLM<sup>2</sup> models.

\
If you want to resume your training, just add some params like: 
\
--optimizer_state /xxx/checkpoints/fallback_80000_opt.pt --start_step 80000 --pretrained none

```shell
CUDA_VISIBLE_DEVICES=0,1,2,3 python -m train.adversarial_training_clip_with_object_token --clip_model_name ViT-L-14 --slots_ckp ./ckps/model_slots_step_300000.pt --dataset imagenet --imagenet_root /path/to/imagenet --template std --output_normalize False --steps 20000 --warmup 1400 --batch_size 128 --loss l2 --opt adamw --lr 1e-5 --wd 1e-4 --attack pgd --inner_loss l2 --norm linf --eps 4 --iterations_adv 10 --stepsize_adv 1 --wandb False --output_dir ./output --experiment_name with_OT --log_freq 10 --eval_freq 10 --optimizer_state /home/xxx/RobustVLM/output/ViT-L-14_openai_imagenet_l2_imagenet_with_Object_Token_xxxxx/checkpoints/fallback_80000_opt.pt --start_step 80000 --pretrained none
```

## Evaluation
Make sure files in `bash` directory are executable: `chmod +x bash/*`
### CLIP ImageNet
```shell
python -m CLIP_eval.clip_robustbench --clip_model_name ViT-L-14 --pretrained /path/to/ckpt.pt --dataset imagenet --imagenet_root /path/to/imagenet --wandb False --norm linf --eps 2
```
⬆ You should notice the `--pretrained` and the `--eps 2/4` for  SlotVLM<sup>2/4</sup> models.


### CLIP Zero-Shot
Set models to be evaluated in `CLIP_benchmark/benchmark/models.txt` and datasets in `CLIP_benchmark/benchmark/datasets.txt`
(the datasets are downloaded from HuggingFace). Then run
```shell
cd CLIP_benchmark
./bash/run_benchmark_adv.sh
```

### VLM Captioning and VQA
#### LLaVA
In `/bash/llava_eval.sh` supply paths for the datasets. The required annotation files for the datasets can be obtained from this [HuggingFace repository](https://huggingface.co/datasets/openflamingo/eval_benchmark/tree/main).
Set `--vision_encoder_pretrained` to `openai` or supply path to fine-tuned CLIP model checkpoint.
Then run
```shell
./bash/llava_eval.sh
```
The LLaVA model will be automatically downloaded from HuggingFace.

#### OpenFlamingo
Download the OpenFlamingo 9B [model](https://huggingface.co/openflamingo/OpenFlamingo-9B-vitl-mpt7b/tree/main), supply paths in `/bash/of_eval_9B.sh` and run
```shell
./bash/of_eval_9B.sh
```

Some non-standard annotation files are supplied [here](https://nc.mlcloud.uni-tuebingen.de/index.php/s/mtRnQFaZJkR9zaX) and [here](https://github.com/mlfoundations/open_flamingo/tree/main/open_flamingo/eval/data).

### VLM Stealthy Targeted Attacks
For targeted attacks on COCO, run
```shell
./bash/llava_eval_targeted.sh
```
For targeted attacks on self-selected images, set images and target captions in `vlm_eval/run_evaluation_qualitative.py` and run
```shell
python -m vlm_eval.run_evaluation_qualitative --precision float32 --attack apgd --eps 2 --steps 10000 --vlm_model_name llava --vision_encoder_pretrained openai --verbose
```
With 10,000 iterations it takes about 2 hours per image on an A100 GPU.

### POPE
```shell
./bash/eval_pope.sh openai  # for clean  model evaluation
./bash/eval_pope.sh   # for robust  model evaluation - add path_to_ckpt in bash file
```
### SQA
```shell
./bash/eval_scienceqa.sh openai  # for clean  model evaluation
./bash/eval_scienceqa.sh   # for robust  model evaluation - add path_to_ckpt in bash file
```