import os; os.system("pip3 install transformers gradio fire accelerate bitsandbytes > /dev/null")
from transformers import AutoModel, AutoTokenizer, AutoImageProcessor
import torch

model_id = "jxu124/TiO"
model = AutoModel.from_pretrained(model_id, trust_remote_code=True, torch_dtype=torch.float16).cuda()
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False)
image_processor = AutoImageProcessor.from_pretrained(model_id)

# ---- gradio demo ----
model.get_gradio_demo(tokenizer, image_processor).queue(max_size=20).launch(server_name="0.0.0.0", server_port=7860)

Mini-Example

python

import os; os.system("pip3 install transformers accelerate bitsandbytes gradio fire")
from transformers import AutoModel, AutoTokenizer, AutoImageProcessor
import torch

model_id = "jxu124/TiO"
model = AutoModel.from_pretrained(model_id, trust_remote_code=True, torch_dtype=torch.float16).cuda()
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False)
image_processor = AutoImageProcessor.from_pretrained(model_id)

# ---- mini example ----
from PIL import Image
from io import BytesIO
import requests

# Prepare example
image = Image.open(BytesIO(requests.get("http://images.cocodataset.org/val2014/COCO_val2014_000000429913.jpg").content))
text = """\
#instruction: can you specify which region the context describes?
#context:
human: look that man in white!"""

# Inference
with torch.no_grad():
    pt_txt = tokenizer([text], return_tensors="pt").input_ids.cuda()
    pt_img = image_processor([image], return_tensors="pt").pixel_values.to(torch.float16).cuda()
    gen = model.generate(pt_txt, patch_images=pt_img, top_p=0.5, do_sample=True, no_repeat_ngram_size=3, max_length=256)
print(tokenizer.batch_decode(gen, skip_special_tokens=True)[0].replace("not yet.", ""))
# e.g. [' is he the one who just threw the ball?']  # Due to the generator, different results may be output

Other Examples (text)

Guesser(grounding):

python

text = """\
#instruction: which region does the context describe?
#context:
human: look that man in white! 
agent: is he the one who just threw the ball? 
human: yes. I mean the pitcher."""

Questioner(question generation):

python

text = """\
#instruction: guess what I want?
#context:
human: look that man in white!"""

Oracle(answering):

python

text = """\
#instruction: answer the question based on the region.
#context:
agent: look that man in white! 
human: is he the one who just threw the ball?
#region: <bin_847> <bin_319> <bin_923> <bin_467>"""