PaliGemma2-3b-VQAv2

This model is a fine-tuned version of google/paligemma2-3b-pt-448 on half of the VQAv2 validation split, for task conditioning. Fine-tuning script is here which also comes in notebook form here. Make sure you install transformers in main branch before using this or running fine-tuning.

How to Use

Below is the code to use this model. Also see inference notebook.

from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
from PIL import Image
import requests

model_id = "merve/paligemma2-3b-vqav2"
model = PaliGemmaForConditionalGeneration.from_pretrained(model_id)
processor = AutoProcessor.from_pretrained("google/paligemma2-3b-pt-224")

prompt = "What is behind the cat?"
image_file = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cat.png?download=true"
raw_image = Image.open(requests.get(image_file, stream=True).raw)

inputs = processor(prompt, raw_image.convert("RGB"), return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=20)

print(processor.decode(output[0], skip_special_tokens=True)[len(prompt):])
# gramophone

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2
• num_epochs: 2

Framework versions

• Transformers (main as of Dec 5)
• PEFT 0.13.2