dagelf / visual-llama

Open LLaMA Eyes to See the World

Open LLaMA Eyes to See the World

This project aims to optimize LLaMA model for visual information understanding like GPT-4 and further explore the potentional of large language model.

Generally, we use CLIP vision encoder to extract image features, then image features are projected with MLP-based or Transformer-based connection network into text embedding dimensionality. Then, visual representation (including additional special tokens [boi] and [eoi]) is concatenated with text representation to learn in a autoregressive manner. The framework is similar to kosmos-1 and PaLM-E.

• Code adjustation to support for multi-modal generation. Download clip and LLaMA models from huggingface. Meantime, we test the scripts are also compatible with other LLaMA model size. Please use script preprocess.py to deal with the data.

• Supervised training stage: freeze llama and clip-encoder models and only optimize the connection network. In this stage, we use COCO, CC-3M and COYO-700M datasets with training scripts train.py. We provide the training hyper-parameter used in our experiemnts on 80G A100 GPU. We also evaluate the image captioning performance in COCO testing set.

  Argument	Values
  batch size	1 * 8 * 8
  epochs	3
  cut length	256
  learning rate	4e-3
  image sequence length	10

• Instructing tuning stage: fine-tuning full model with mixed VQA and language-only instructing dataset. We use lora strategy to optimize the entire model with fine-tuning scripts finetune.py.

  Argument	Values
  batch size	1024
  epochs	3
  cut length	256
  learning rate	2e-5
  image sequence length	10

• Open source trained ckpt on huggingface and gradio interface for multi-model generation.

Reference

[1] https://github.com/facebookresearch/llama

[2] https://github.com/tloen/alpaca-lora