app.py

from PIL import Image
import os
import requests
import torch
from torchvision import transforms
from torchvision.transforms.functional import InterpolationMode

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


import gradio as gr

from models.blip_vqa import blip_vqa

image_size_vq = 480
transform_vq = transforms.Compose([
    transforms.Resize((image_size_vq,image_size_vq),interpolation=InterpolationMode.BICUBIC),
    transforms.ToTensor(),
    transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
    ]) 

model_url_vq = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model*_vqa.pth'
    
model_vq = blip_vqa(pretrained=model_url_vq, image_size=480, vit='base')
model_vq.eval()
model_vq = model_vq.to(device)



def inference(raw_image, question, mnlen, mxlen, token):
    if token != os.environ["TOKEN"]:
        return "Rong token"
    image_vq = transform_vq(raw_image).unsqueeze(0).to(device)  
    with torch.no_grad():
        answer = model_vq(image_vq, question, train=False, inference='generate', mina_len=mnlen, maxa_len=mxlen) 
    return  'answer: '+answer[0]
    
inputs = [gr.Image(type='pil'),
gr.Textbox(lines=2, label="Question"),
gr.Number(value=1, label="Min length", precision=0),
gr.Number(value=10, label="Max length", precision=0),
gr.Textbox(lines=1, label="Auth token")]

outputs = gr.outputs.Textbox(label="Output")

title = "BLIP"

description = "Gradio endpoint for spuun's BLIP (Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation (Salesforce Research)). To use it you need to obtain a token from me :) Read more at the links below."

article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2201.12086' target='_blank'>BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation</a> | <a href='https://github.com/salesforce/BLIP' target='_blank'>Github Repo</a></p>"

gr.Interface(inference, inputs, outputs, title=title, description=description, article=article).launch()