app.py

# --------------------------------------------------------
# X-Decoder -- Generalized Decoding for Pixel, Image, and Language
# Copyright (c) 2022 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Written by Jianwei Yang (jianwyan@microsoft.com), Xueyan Zou (xueyan@cs.wisc.edu)
# --------------------------------------------------------

import os
os.system("python -m pip install git+https://github.com/MaureenZOU/detectron2-xyz.git")

import gradio as gr
import torch
import argparse

from xdecoder.BaseModel import BaseModel
from xdecoder import build_model
from utils.distributed import init_distributed
from utils.arguments import load_opt_from_config_files

from tasks import *

def parse_option():
    parser = argparse.ArgumentParser('X-Decoder All-in-One Demo', add_help=False)
    parser.add_argument('--conf_files', default="configs/xdecoder/svlp_focalt_lang.yaml", metavar="FILE", help='path to config file', )
    args = parser.parse_args()

    return args

'''
build args
'''
args = parse_option()
opt = load_opt_from_config_files(args.conf_files)
opt = init_distributed(opt)

# META DATA
pretrained_pth_last = os.path.join("xdecoder_focalt_last.pt")
pretrained_pth_novg = os.path.join("xdecoder_focalt_last_novg.pt")

if not os.path.exists(pretrained_pth_last):
    os.system("wget {}".format("https://projects4jw.blob.core.windows.net/x-decoder/release/xdecoder_focalt_last.pt"))

if not os.path.exists(pretrained_pth_novg):
    os.system("wget {}".format("https://projects4jw.blob.core.windows.net/x-decoder/release/xdecoder_focalt_last_novg.pt"))


'''
build model
'''
model_last = BaseModel(opt, build_model(opt)).from_pretrained(pretrained_pth_last).eval().cuda()
model_cap = BaseModel(opt, build_model(opt)).from_pretrained(pretrained_pth_novg).eval().cuda()

with torch.no_grad():
    model_last.model.sem_seg_head.predictor.lang_encoder.get_text_embeddings(["background", "background"], is_eval=True)
    model_cap.model.sem_seg_head.predictor.lang_encoder.get_text_embeddings(["background", "background"], is_eval=True)

'''
inference model
'''

@torch.no_grad()
def inference(image, instruction, *args, **kwargs):
    image = image.convert("RGB")
    with torch.autocast(device_type='cuda', dtype=torch.float16):
        return referring_inpainting_gpt3(model_last, image, instruction, *args, **kwargs)

'''
launch app
'''
title = "Instructional Image Editing"
description = "<p style='text-align: center'> <a href='https://x-decoder-vl.github.io/' target='_blank'>Project Page</a> | <a href='https://arxiv.org/pdf/2212.11270.pdf' target='_blank'>Paper</a> | <a href='https://github.com/microsoft/X-Decoder' target='_blank'>Github Repo</a> | <a href='https://youtu.be/wYp6vmyolqE' target='_blank'>Video</a> </p>"

article = "The Demo is Run on X-Decoder (Focal-T)."

inputs = [gr.inputs.Image(type='pil'), gr.Textbox(label="instruction")]
gr.Interface(
    fn=inference,
    inputs=inputs,
    outputs=[
        gr.outputs.Image(
        type="pil",
        label="edit result"),
    ],
    examples=[
    ["./images/apples.jpg", "change green apple to a red apple"],
    ["./images/girl_and_two_boys.png", "remove the boy with blue backbag"],
    ["./images/dog.png", "remove the dog"],
    ],
    title=title,
    description=description,
    article=article,
    allow_flagging='never',
    cache_examples=True,
).launch()