/shikra

Primary LanguagePythonOtherNOASSERTION

Logo

Shikra: Unleashing Multimodal LLM’s Referential Dialogue Magic


Shikra, an MLLM designed to kick off referential dialogue by excelling in spatial coordinate inputs/outputs in natural language, without additional vocabularies, position encoders, pre-/post-detection, or external plug-in models.

teaser

News

[07/16] We released the shikra demo. You can access our online demo or deploy it locally.

[07/03] We released the code, data and shikra-7b checkpoint.

[06/28] We released Shikra: Unleashing Multimodal LLM’s Referential Dialogue Magic, which is designed to kick off referential dialogue. Checkout the paper.

Contents

Install

conda create -n shikra python=3.10
conda activate shikra
pip install -r requirements.txt

configure accelerate

accelerate config

Checkpoint

we maintain shikra-7b-delta-v1 and a frequently updated ckpt on huggingface. shikra-7b-delta-v1 is the checkpoint we evaluate and report in the paper, and the frequently updated ckpt is constantly updated to add new features.

We release Shikra weights as delta weights to comply with the LLaMA model license. You can add our delta to the original LLaMA weights to obtain the Shikra weights.

Use shikra-7b-delta-v1 Instructions:

  1. Get the original LLaMA weights in the huggingface format by following the instructions here.
  2. Use the following scripts to get weights by applying our delta (shikra-7b-delta). It will automatically download delta weights from our Hugging Face account.
python mllm/models/shikra/apply_delta.py \
    --base /path/to/llama-7b \
    --target /output/path/to/shikra-7b \
    --delta shikras/shikra-7b-delta-v1

Use the frequently updated ckpt Instructions:

  1. Get the original LLaMA weights in the huggingface format by following the instructions here.
  2. Use the following scripts to get weights by applying our delta (frequently updated ckpt). It will automatically download delta weights from our Hugging Face account.
python mllm/models/shikra/apply_delta.py \
    --base /path/to/llama-7b \
    --target /output/path/to/shikra-7b \
    --delta shikras/shikra-7b-delta-v1-0708

Demo

Gradio Web Demo

To launch a Gradio web demo, use the following command. Please note that the model evaluates in the torch.float16 format, which requires a GPU with at least 16GB of memory.

python mllm/demo/webdemo.py --model_path /path/to/shikra/ckpt

It is also possible to use it in 8-bit quantization, albeit at the expense of sacrificing some performance.

python mllm/demo/webdemo.py --model_path /path/to/shikra/ckpt --load_in_8bit

Server-Client Demo

launch a shikra server:

python mllm/demo/server.py --model_path /path/to/shikra/ckpt

a client example is in mllm/demo/client.py, check the example results by

python mllm/demo/client.py

Train

After preparing data, you can train the model using the command:

accelerate launch --num_processes 4 \
        --main_process_port 23786 \
        mllm/pipeline/finetune.py \
        config/shikra_pretrain_final19_stage2.py \
        --cfg-options model_args.model_name_or_path=/path/to/init/checkpoint

mmengine style args and huggingface:Trainer args are supported. For example, you can change epoch and output_dir like this:

accelerate launch --num_processes 4 \
        --main_process_port 23786 \
        mllm/pipeline/finetune.py \
        config/shikra_pretrain_final19_stage2.py \
        --cfg-options model_args.model_name_or_path=/path/to/init/checkpoint \
        --num_train_epochs 3 \
        --output_dir /path/to/new/exp/dir

where --cfg-options a=balabala b=balabala is mmengine style argument. They will overwrite the argument predefined in config file. And --num_train_epochs , --output_dir are huggingface:Trainer argument.

Inference

After preparing data, you can inference the model using the command:

accelerate launch --num_processes 4 \
        --main_process_port 23786 \
        mllm/pipeline/finetune.py \
        config/shikra_eval_multi_pope.py \
        --cfg-options model_args.model_name_or_path=/path/to/checkpoint

mmengine style args and huggingface:Trainer args are supported. for example, you can change eval batchsize like this:

accelerate launch --num_processes 4 \
        --main_process_port 23786 \
        mllm/pipeline/finetune.py \
        config/shikra_eval_multi_pope.py \
        --cfg-options model_args.model_name_or_path=/path/to/checkpoint \
        --per_device_eval_batch_size 1

where --cfg-options a=balabala b=balabala is mmengine style argument. They will overwrite the argument predefined in config file. And --per_device_eval_batch_size is huggingface:Trainer argument.

the prediction result will be saved in output_dir/multitest_xxxx_extra_prediction.jsonl, which hold the same order as the input dataset.

Examples

shikra_case_1

Cite

@article{chen2023shikra,
  title={Shikra: Unleashing Multimodal LLM's Referential Dialogue Magic},
  author={Chen, Keqin and Zhang, Zhao and Zeng, Weili and Zhang, Richong and Zhu, Feng and Zhao, Rui},
  journal={arXiv preprint arXiv:2306.15195},
  year={2023}
}

Acknowledgement

This repo benefits from LLaVA, Vicuna, ChatGLM-Efficient-Tuning and GLIGEN. Thanks for their wonderful works.