README.md

Coda-Robotics/OpenVLA-ER-Select-Book-LoRA

Model Description

This is a LoRA adapter weights only (requires base OpenVLA model) of OpenVLA, fine-tuned on the select_book dataset.

Training Details

• Dataset: select_book
• Number of Episodes: 479
• Batch Size: 8
• Training Steps: 20000
• Learning Rate: 2e-5
• LoRA Configuration:
  • Rank: 32
  • Dropout: 0.0
  • Target Modules: all-linear

Usage

from transformers import AutoProcessor, AutoModelForVision2Seq

# Load the model and processor
processor = AutoProcessor.from_pretrained("Coda-Robotics/OpenVLA-ER-Select-Book-LoRA")
model = AutoModelForVision2Seq.from_pretrained("Coda-Robotics/OpenVLA-ER-Select-Book-LoRA")

# Process an image
image = ...  # Load your image
inputs = processor(images=image, return_tensors="pt")
outputs = model.generate(**inputs)
text = processor.decode(outputs[0], skip_special_tokens=True)

Using with PEFT

To use this adapter with the base OpenVLA model:

from transformers import AutoProcessor, AutoModelForVision2Seq
from peft import PeftModel, PeftConfig

# Load the base model
base_model = AutoModelForVision2Seq.from_pretrained("openvla/openvla-7b")

# Load the LoRA adapter
adapter_model = PeftModel.from_pretrained(base_model, "{model_name}")

# Merge weights for faster inference (optional)
merged_model = adapter_model.merge_and_unload()