from transformers import BlipProcessor, BlipForConditionalGeneration
import os
import gradio as gr

caption_model = 'Salesforce/blip-image-captioning-base'
processor = BlipProcessor.from_pretrained(caption_model)
model = BlipForConditionalGeneration.from_pretrained(caption_model)

def launch(input):
  # proprocess text to prepare for the model
  inputs = processor(input, return_tensors="pt")
  output = model.generate(**inputs)
  cond_caption = processor.decode(output[0], skip_special_tokens=True)
  return cond_caption

iface = gr.Interface(launch,
                     inputs=gr.Image(type='pil'),
                     outputs="text")

# launch demo
iface.launch()