Skip to content

Latest commit

 

History

History
44 lines (28 loc) · 1.62 KB

training.md

File metadata and controls

44 lines (28 loc) · 1.62 KB
Raw

Overview

We decide not to integrate the training code for each part due to high maintenance cost. Please refer to the original repositories. However, we provide a set of scripts to convert the OpenCOLE dataset into some dataset formats that are ready to use for the following models.

Data Generation

Text-to-image

We use SimpleTuner for training text-to-image part. The following script saves the input prompt and the target image given each sample for SimpleTuner.

poetry run python -m opencole.preprocess.create_dataset_for_text_to_image --output_dir <OUTPUT_DIR> --max_size <MAX_SIZE>

TypographyLMM

We use LLaVA1.5 for training TypographyLMM part.

poetry run python -m opencole.preprocess.create_dataset_for_typographylmm --output_dir <OUTPUT_DIR>

Model Conversion

Convert LLaVA1.5 model to a model compatible with transformers

  1. Saving all the weights in a single file

Note: this should be done using LLaVA's codebase.

python dump_full_llava_weight_dicts.py --model_base <MODEL_BASE> --model_base liuhaotian/llava-v1.5-7b
  1. Conversion

For example, to convert a llava1.5-based model, please run the following.

poetry run python -m opencole.hf_support.convert_llava_to_hf --model_path <MODEL_PATH> --output_dir <OUTPUT_DIR> --text_model_id lmsys/vicuna-7b-v1.5 --vision_model_id openai/clip-vit-large-patch14-336