Instructions to use gulucaptain/pi05-base-pytorch with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use gulucaptain/pi05-base-pytorch with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("gulucaptain/pi05-base-pytorch", device_map="auto") - Notebooks
- Google Colab
- Kaggle
OpenPI pi05_base β BF16 PyTorch conversion
This is the base checkpoint used to initialize gulucaptain/pi05-icl3000-demo, before any Demo or LoRA training.
Original weights: gs://openpi-assets/checkpoints/pi05_base, published by Physical Intelligence / OpenPI.
Converted using the experiment's OpenPI JAX-to-PyTorch converter with configuration pi05_aloha and precision bfloat16. The configuration selects the pi05 architecture; these are pi05_base weights, not ALOHA fine-tuned weights.
Files and identity
model.safetensors: 7,233,650,408 bytes.- SHA256:
abecb1fa2c618368b5aecbc0e10c546f64927077d88fea6afc7c564371781270. config.json: conversion architecture settings, not a complete Transformers AutoModel configuration.base_identity.json: source URI, conversion config and converter SHA256.convert_jax_model_to_pytorch.py: exact converter used (OpenPI-derived, with completeness checks).OPENPI_CODE_LICENSE: license for the OpenPI-derived converter code; this does not replace upstream model terms.
Verification
All 29 cached official source objects matched GCS sizes and MD5/CRC32C checksums. All 811 mapped official parameter tensors, including tied token embeddings, were compared on CPU and exactly matched after BF16 conversion. The tied embedding is saved as paligemma_with_expert.paligemma.lm_head.weight; safetensors metadata records the alias. The serialized model also contains an unused action-expert lm_head without a corresponding JAX checkpoint tensor; it is not used by the policy forward path.
Usage
This is a native OpenPI-style checkpoint, not a standalone robot controller or a Transformers AutoModel export. Use the matching training/inference implementation and its required Transformers replacement. When loading a Demo LoRA checkpoint, pass the downloaded model.safetensors as base_checkpoint. The inference code checks its SHA256 before loading the adapter. This repository contains no Demo/LoRA updates or optimizer state.
Use remains subject to the upstream OpenPI, PaliGemma/Gemma and model terms. This conversion does not assert a new license grant for the original model weights.
- Downloads last month
- 10