VisAble diagram captioner

BLIP fine-tuned on AI2D-Caption diagrams to produce part-to-whole descriptions for blind and low-vision students. The vision encoder was frozen and only the text decoder was trained, using a cross entropy loss that up-weights AI2D entity labels by 5.0x.

Trained for 10 epochs, batch size 8, learning rate 5e-05.

from transformers import BlipForConditionalGeneration, BlipProcessor
from PIL import Image

processor = BlipProcessor.from_pretrained("hkondle/CapstoneMainModel-10ep")
model = BlipForConditionalGeneration.from_pretrained("hkondle/CapstoneMainModel-10ep")

image = Image.open("diagram.png").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
print(processor.decode(model.generate(**inputs, max_length=256, num_beams=4)[0], skip_special_tokens=True))