import gradio as gr
from PIL import Image
import torch
from transformers import BlipProcessor, BlipForConditionalGeneration

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

processor = BlipProcessor.from_pretrained("noamrot/FuseCap_Image_Captioning")
model = BlipForConditionalGeneration.from_pretrained("noamrot/FuseCap_Image_Captioning").to(device)

def inference(raw_image):
    text = "a picture of "
    inputs = processor(raw_image, text, return_tensors="pt").to(device)
    out = model.generate(**inputs)
    caption = processor.decode(out[0], skip_special_tokens=True)
    return caption


inputs = [gr.Image(type='pil', interactive=False),]
# outputs = gr.outputs.Textbox(label="Caption")
outputs = gr.Textbox(label="Caption")

description = "Gradio demo for FuseCap: Leveraging Large Language Models for Enriched Fused Image Captions. This demo features a BLIP-based model, trained using FuseCap."
examples = [["surfer.jpg"], ["bike.jpg"]]
article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2305.17718' target='_blank'>FuseCap: Leveraging Large Language Models for Enriched Fused Image Captions</a>"


iface = gr.Interface(fn=inference, 
                    inputs="image",
                    outputs="text", 
                    title="FuseCap",
                    description=description,
                    article=article,
                    examples=examples,
                    # enable_queue=True
                    )

iface.queue()


iface.launch()