Instruct-X-Decoder / tasks /
jw2yang's picture
Update tasks/
# --------------------------------------------------------
# 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 (
# --------------------------------------------------------
import os
import openai
import torch
import numpy as np
from scipy import ndimage
from PIL import Image
from utils.inpainting import pad_image, crop_image
from torchvision import transforms
from utils.visualizer import Visualizer
from diffusers import StableDiffusionInpaintPipeline
from detectron2.utils.colormap import random_color
from import MetadataCatalog
t = []
t.append(transforms.Resize(512, interpolation=Image.BICUBIC))
transform = transforms.Compose(t)
metadata = MetadataCatalog.get('ade20k_panoptic_train')
pipe = StableDiffusionInpaintPipeline.from_pretrained(
# "stabilityai/stable-diffusion-2-inpainting",
prompts = []
prompts.append("instruction: remove the person, task: (referring editing), source: [person], target:<clean and empty scene>.")
prompts.append("instruction: remove the person in the middle, task: (referring editing), source: [person in the middle], target:<clean and empty scene>.")
prompts.append("instruction: remove the dog on the left side, task: (referring editing), source: [dog on the left side], target:<clean and empty scene>.")
prompts.append("instruction: change the apple to a pear, task: (referring editing), source: [apple], target: <pear>.")
prompts.append("instruction: change the red apple to a green one, task: (referring editing), source: [red apple], target: <green apple>.")
prompts.append("instruction: change the color of bird's feathers from white to blue, task: (referring editing), source: [white bird], target: <blue bird>.")
prompts.append("instruction: replace the dog with a cat, task: (referring editing), source: [dot], target: <cat>.")
prompts.append("instruction: replace the red apple with a green one, task: (referring editing), source: [red apple], target: <green apple>.")
openai.api_type = "azure"
openai.api_base = ""
openai.api_version = "2022-12-01"
openai.api_key = os.environ["OPENAI_API_KEY"]
def get_gpt3_response(prompt):
response = openai.Completion.create(
return response
def referring_inpainting_gpt3(model, image, instruction, *args, **kwargs):
# convert instruction to source and target
instruction = instruction.replace('.', '')
resp = get_gpt3_response(' '.join(prompts) + ' instruction: ' + instruction + ',')
resp_text = resp['choices'][0]['text']
ref_text = resp_text[resp_text.find('[')+1:resp_text.find(']')]
inp_text = resp_text[resp_text.find('<')+1:resp_text.find('>')]
model.model.metadata = metadata
texts = [[ref_text if ref_text.strip().endswith('.') else (ref_text.strip() + '.')]]
image_ori = crop_image(transform(image))
with torch.no_grad():
width = image_ori.size[0]
height = image_ori.size[1]
image = np.asarray(image_ori)
image_ori_np = np.asarray(image_ori)
images = torch.from_numpy(image.copy()).permute(2,0,1).cuda()
batch_inputs = [{'image': images, 'height': height, 'width': width, 'groundings': {'texts': texts}}]
outputs = model.model.evaluate_grounding(batch_inputs, None)
visual = Visualizer(image_ori_np, metadata=metadata)
grd_mask = (outputs[0]['grounding_mask'] > 0).float().cpu().numpy()
for idx, mask in enumerate(grd_mask):
color = random_color(rgb=True, maximum=1).astype(np.int32).tolist()
demo = visual.draw_binary_mask(mask, color=color, text=texts[idx])
res = demo.get_image()
if inp_text not in ['no', '']:
image_crop = image_ori
struct2 = ndimage.generate_binary_structure(2, 2)
mask_dilated = ndimage.binary_dilation(grd_mask[0], structure=struct2, iterations=3).astype(grd_mask[0].dtype)
mask = Image.fromarray(mask_dilated * 255).convert('RGB')
image_and_mask = {
"image": image_crop,
"mask": mask,
# images_inpainting = inpainting(inpainting_model, image_and_mask, inp_text, ddim_steps, num_samples, scale, seed)
width = image_ori.size[0]; height = image_ori.size[1]
images_inpainting = pipe(prompt = inp_text.strip(), image=image_and_mask['image'], mask_image=image_and_mask['mask'], height=height, width=width).images
return images_inpainting[0]
return Image.fromarray(res)