caption.py

import gradio as gr
from transformers import AutoProcessor, AutoModelForCausalLM
import torch
from PIL import Image

from caption import predict_step

with gr.Blocks() as demo:
    image = gr.Image(type='pil', label='Image')
    label = gr.Text(label='Generated Caption')
    image.upload(
        predict_step,
        [image],
        [label]
    )

model = AutoModelForCausalLM.from_pretrained("Chesscorner/git-chess-v3")
processor = AutoProcessor.from_pretrained("Chesscorner/git-chess-v3")

# Set up device and move model to it
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# Enable mixed precision if on GPU
use_fp16 = device.type == "cuda"
if use_fp16:
    model.half()

# Set generation parameters
gen_kwargs = {'max_length': 100, 'num_beams': 2}  # Adjust num_beams if needed


# Prediction function
def predict_step(image):
    # Preprocess the image
    pixel_values = processor(images=image, return_tensors="pt").pixel_values.to(device)

    # Generate predictions with no_grad for efficiency
    with torch.no_grad():
        output_ids = model.generate(pixel_values=pixel_values, **gen_kwargs)

    # Decode predictions
    preds = processor.batch_decode(output_ids, skip_special_tokens=True)
    return preds[0].strip()

if __name__ == '__main__':
    demo.launch(share=True)