LLaVA-Next-Inst-It-Qwen2-7B

Homepage | Code | Paper | arXiv

LLaVA-Next-Inst-It-Qwen2-7B is a multimodal model that excels at instance-level understanding, which is introduced in the paper Inst-IT: Boosting Multimodal Instance Understanding via Explicit Visual Prompt Instruction Tuning

• Architecture: siglip-so400m-patch14-384 + Qwen2-7B
• Data: LLaVA-NeXT-Data / Inst-IT-Dataset
• Precision: bfloat16

Quick Start

Install

Our code is based on LLaVA-NeXT, before running, please install the LLaVA-NeXT to prepare the environment:

pip install git+https://github.com/LLaVA-VL/LLaVA-NeXT.git

Error Handling

Click to unfold

• Common error case 1:

  Exception: data did not match any variant of untagged enum ModelWrapper at line 757272 column 3

This is caused by the version of transformers, try to update it:

pip install -U transformers

• Common error case 2:

RuntimeError: Error(s) in loading state_dict for CLIPVisionModel:
    size mismatch for vision_model.embeddings.position_embedding.weight: copying a param with shape torch.Size([729, 1152]) from checkpoint, the shape in current model is torch.Size([730, 1152]).
    You may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method.

This is a logical error encountered when loading the vision tower from the local path. To fix this issue, you can prepare the environment in any of the following ways.

Option 1: Install from our fork of LLaVA-NeXT:

pip install git+https://github.com/inst-it/LLaVA-NeXT.git

Option 2: Install from LLaVA-NeXT and manually modify its code:

• step 1: clone source code

git clone https://github.com/LLaVA-VL/LLaVA-NeXT.git

• step 2: before installing LLaVA-NeXT, you need to modify line 17 of llava/model/multimodal_encoder/builder.py.

# Before modification:
if is_absolute_path_exists or vision_tower.startswith("openai") or vision_tower.startswith("laion") or "ShareGPT4V" in vision_tower:

# After modification:
if "clip" in vision_tower or vision_tower.startswith("openai") or vision_tower.startswith("laion") or "ShareGPT4V" in vision_tower:

• step 3: install LLaVA-NeXT from source:

cd LLaVA-NeXT
pip install --upgrade pip  # Enable PEP 660 support.
pip install -e ".[train]"

We recommend the first way because it is simple.

Load Model

from llava.model.builder import load_pretrained_model
from llava.constants import DEFAULT_IMAGE_TOKEN

from llava.mm_utils import (
    KeywordsStoppingCriteria,
    get_model_name_from_path,
    tokenizer_image_token,
    process_images
)
from llava.conversation import SeparatorStyle, conv_templates
from llava.eval.model_vqa import preprocess_qwen

overwrite_config = {}
overwrite_config["mm_spatial_pool_stride"] = 2
overwrite_config["mm_spatial_pool_mode"] = 'bilinear'
overwrite_config["mm_pooling_position"] = 'after'
overwrite_config["mm_newline_position"] = 'no_token'

model_path = "Inst-IT/LLaVA-Next-Inst-It-Qwen2-7B"
model_name = get_model_name_from_path(model_path)

tokenizer, model, image_processor, max_length = load_pretrained_model(
            model_path=model_path, 
            model_base=None, 
            model_name=model_name,
            device_map="auto", 
            torch_dtype='bfloat16', 
            overwrite_config=overwrite_config,
            attn_implementation='sdpa')

Image Inference

Inference without SoMs

Our model can perform inference on images without Set-of-Marks visual prompts, in this case, it can be used in the same way as its base mode LLaVA-NeXT.

import torch
import requests
from PIL import Image

img_url = "https://github.com/inst-it/inst-it/blob/main/assets/demo/image.jpg?raw=true"
image = Image.open(requests.get(img_url, stream=True).raw)
image_tensor = process_images([image], image_processor, model.config).bfloat16()
image_sizes = [image.size]

question = "Describe this image."
question = DEFAULT_IMAGE_TOKEN + "\n" + question

conv_template = 'qwen_1_5'
conv = conv_templates[conv_template].copy()
conv.append_message(conv.roles[0], question)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()

input_ids = preprocess_qwen([{'from': 'human','value': question},{'from': 'gpt','value': None}], tokenizer, has_image=True).cuda()

pad_token_ids = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
attention_masks = input_ids.ne(pad_token_ids).long().cuda()

stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
keywords = [stop_str]
stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)

with torch.inference_mode():
    output_ids = model.generate(
        inputs=input_ids,
        images=image_tensor,
        attention_mask=attention_masks,
        modalities="image",
        image_sizes=image_sizes,
        use_cache=True,
        stopping_criteria=[stopping_criteria],
        max_new_tokens=4096
    )

pred = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(pred)

Inference with SoMs

Our model performs more fine-grained understanding when Set-of-Marks visual prompts are provided. You can refer to the instances that you are interested in using their IDs. Compared to the previous inference code, the following code has no modifications except for the input image, which is visual prompted with Set-of-Marks. Refer to this link to learn how to generate SoMs for an image.

import torch
import requests
from PIL import Image

img_url = "https://github.com/inst-it/inst-it/blob/main/assets/demo/image_som.jpg?raw=true"
image = Image.open(requests.get(img_url, stream=True).raw)
image_tensor = process_images([image], image_processor, model.config).bfloat16()
image_sizes = [image.size]

# You can use [id] to refer to the instances that you are interested in
question = "Describe [8] in detail."
question = DEFAULT_IMAGE_TOKEN + "\n" + question

conv_template = 'qwen_1_5'
conv = conv_templates[conv_template].copy()
conv.append_message(conv.roles[0], question)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()

input_ids = preprocess_qwen([{'from': 'human','value': question},{'from': 'gpt','value': None}], tokenizer, has_image=True).cuda()

pad_token_ids = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
attention_masks = input_ids.ne(pad_token_ids).long().cuda()

stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
keywords = [stop_str]
stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)

with torch.inference_mode():
    output_ids = model.generate(
        inputs=input_ids,
        images=image_tensor,
        attention_mask=attention_masks,
        modalities="image",
        image_sizes=image_sizes,
        use_cache=True,
        stopping_criteria=[stopping_criteria],
        max_new_tokens=4096
    )

pred = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(pred)

Video Inference

For the video, we organize each frame into a list. You can use the format <t> to refer to a specific timestamp (e.g. <1>).

Inference without SoMs

Our model can perform inference on videos without Set-of-Marks visual prompts, in this case, it can be used in the same way as its base mode LLaVA-NeXT.

import torch
import requests
from PIL import Image

frame_urls = [
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/frame_1.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/frame_2.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/frame_3.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/frame_4.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/frame_5.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/frame_6.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/frame_7.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/frame_8.jpg?raw=true"
]
video = [Image.open(requests.get(frame_url, stream=True).raw) for frame_url in frame_urls]
video = image_processor.preprocess(video, return_tensors="pt")["pixel_values"].cuda()
video = video.bfloat16()
videos = [video]

question = "Describe the video."  # overall video caption
question = "What happens at frame <1>?"  # caption a specific moment
question = DEFAULT_IMAGE_TOKEN + "\n" + question

conv_template = 'qwen_1_5'
conv = conv_templates[conv_template].copy()
conv.append_message(conv.roles[0], question)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()

input_ids = preprocess_qwen([{'from': 'human','value': question},{'from': 'gpt','value': None}], tokenizer, has_image=True).cuda()

pad_token_ids = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
attention_masks = input_ids.ne(pad_token_ids).long().cuda()

stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
keywords = [stop_str]
stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)

with torch.inference_mode():
    output_ids = model.generate(
        inputs=input_ids,
        images=videos,
        attention_mask=attention_masks,
        modalities="video",
        use_cache=True,
        stopping_criteria=[stopping_criteria],
        max_new_tokens=4096
    )

pred = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(pred)

Inference with SoMs

Our model performs more fine-grained understanding when Set-of-Marks visual prompts are provided. You can refer to the instances that you are interested in using their IDs. Compared to the previous inference code, the following code has no modifications except for the input video, which is visual prompted with Set-of-Marks. Refer to SAM2 and SoM to learn how to generate SoMs for a video.

import torch
import requests
from PIL import Image

frame_urls = [
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/som_frame_1.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/som_frame_2.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/som_frame_3.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/som_frame_4.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/som_frame_5.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/som_frame_6.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/som_frame_7.jpg?raw=true",
    "https://github.com/inst-it/inst-it/blob/main/assets/demo/som_frame_8.jpg?raw=true"
]
video = [Image.open(requests.get(frame_url, stream=True).raw) for frame_url in frame_urls]
video = image_processor.preprocess(video, return_tensors="pt")["pixel_values"].cuda()
video = video.bfloat16()
videos = [video]

# You can use [id] to refer to the instances that you are interested in
question = "Is [3] visible at <1>?"
question = DEFAULT_IMAGE_TOKEN + "\n" + question

conv_template = 'qwen_1_5'
conv = conv_templates[conv_template].copy()
conv.append_message(conv.roles[0], question)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()

input_ids = preprocess_qwen([{'from': 'human','value': question},{'from': 'gpt','value': None}], tokenizer, has_image=True).cuda()

pad_token_ids = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
attention_masks = input_ids.ne(pad_token_ids).long().cuda()

stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
keywords = [stop_str]
stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)

with torch.inference_mode():
    output_ids = model.generate(
        inputs=input_ids,
        images=videos,
        attention_mask=attention_masks,
        modalities="video",
        use_cache=True,
        stopping_criteria=[stopping_criteria],
        max_new_tokens=4096
    )

pred = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(pred)

Contact

Feel free to contact us if you have any questions or suggestions

• Email (Wujian Peng): [email protected]
• Email (Lingchen Meng): [email protected]

Citation

@article{peng2024inst,
  title={Inst-IT: Boosting Multimodal Instance Understanding via Explicit Visual Prompt Instruction Tuning},
  author={Peng, Wujian and Meng, Lingchen and Chen, Yitong and Xie, Yiweng and Liu, Yang and Gui, Tao and Xu, Hang and Qiu, Xipeng and Wu, Zuxuan and Jiang, Yu-Gang},
  journal={arXiv preprint arXiv:2412.03565},
  year={2024}
}