Skip to content

Latest commit

 

History

History
39 lines (25 loc) · 2.06 KB

README_cont.md

File metadata and controls

39 lines (25 loc) · 2.06 KB

Prismatic Repository

Directory Structure

  • Makefile: Self-explanatory, work in progress.
  • README.md: Project documentation.
  • pyproject.toml: Centralizes configuration for building the project, specifying dependencies and tools.
  • requirements-min.txt: Minimum requirements for the project.

prismatic/

Contains miscellaneous setup for models, configs, processing, and training. Should not require editing.

scripts/

  • generate.py: Simple CLI script for interactive testing of generating from a pretrained VLM. Provides a minimal REPL for specifying image URLs, prompts, and language generation parameters.
  • preprocess.py: Core script for automatically downloading raw VLM pretraining datasets, including LLaVA v1.5 datasets and stages. Runs download and extraction automatically.
  • pretrain.py: Pretraining script for Prismatic VLM in native PyTorch, using Fully-Sharded Data Parallel (FSDP) for distributed training across GPUs. Assumes CUDA toolkit >= 11.0 for BF16 mixed precision.

additional-datasets/

  • lrv_instruct.py: Preprocessing script for LRV-Instruct data. Requires prior download of images and JSON files.
  • lvis_instruct_4v.py: Preprocessing script for LVIS-Instruct4v (language/chat) data. Requires prior download of images.

extern/

  • convert_openvla_weights_to_hf.py: Converts weights from a Prismatic VLM model to HuggingFace format for compatibility with the transformers library.
  • verify_openvla.py: Verifies functionality of a Prismatic model exported to HuggingFace using AutoClasses.

vla-scripts/

  • deploy.py: Sets up a server for deploying OpenVLA models via a REST API.
  • finetune.py: Fine-tunes a model given a finetune model, dataset directory, name of fine-tuning dataset, and hyperparameters.
  • train.py: Configures and automates training for OpenVLA models.

extern/

  • convert_openvla_weights_to_hf.py: Utility script for converting OpenVLA weights to HuggingFace-compatible weights.
  • verify_openvla.py: Checks functionality of a HuggingFace-exported OpenVLA model.