This is a
finetuned version of
LLaVA-OneVision-Qwen2-7B-OV for egocentric vision-language tasks.
This model is a
finetuned version of LLaVA-OneVision-Qwen2-7B-OV, fine-tuned on
EgoIT-99K and Ego4D-like datasets for egocentric video question answering tasks. The base model is a 7B parameter multimodal model based on Qwen2 language model with a context window of 32K tokens, capable of understanding images, multi-image, and videos.
This model is
finetuned on
EgoIT-99K and Ego4D datasets for egocentric vision-language understanding tasks, particularly video question answering from first-person perspective. The model inherits the base model's ability to interact with images, multi-image and videos, with enhanced capabilities for egocentric video understanding.
We provide the simple generation process for using our model. For more details, you could refer to
Github.
1# pip install git+https://github.com/LLaVA-VL/LLaVA-NeXT.git
2from llava.model.builder import load_pretrained_model
3from llava.mm_utils import get_model_name_from_path, process_images, tokenizer_image_token
4from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, IGNORE_INDEX
5from llava.conversation import conv_templates, SeparatorStyle
6
7from PIL import Image
8import requests
9import copy
10import torch
11
12import sys
13import warnings
14
15warnings.filterwarnings("ignore")
16pretrained = "sunidhitandel/hpml-egoqa-baseline" # Finetuned model
17model_name = "llava_qwen"
18device = "cuda"
19device_map = "auto"
20tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, device_map=device_map) # Add any other thing you want to pass in llava_model_args
21
22model.eval()
23
24url = "https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true"
25image = Image.open(requests.get(url, stream=True).raw)
26image_tensor = process_images([image], image_processor, model.config)
27image_tensor = [_image.to(dtype=torch.float16, device=device) for _image in image_tensor]
28
29conv_template = "qwen_1_5" # Make sure you use correct chat template for different models
30question = DEFAULT_IMAGE_TOKEN + "\nWhat is shown in this image?"
31conv = copy.deepcopy(conv_templates[conv_template])
32conv.append_message(conv.roles[0], question)
33conv.append_message(conv.roles[1], None)
34prompt_question = conv.get_prompt()
35
36input_ids = tokenizer_image_token(prompt_question, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(device)
37image_sizes = [image.size]
38
39
40cont = model.generate(
41 input_ids,
42 images=image_tensor,
43 image_sizes=image_sizes,
44 do_sample=False,
45 temperature=0,
46 max_new_tokens=4096,
47)
48text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)
49print(text_outputs)
This model is finetuned from
LLaVA-OneVision-Qwen2-7B-OV, which was trained on:
1@article{li2024llavaonevision,
2 title={LLaVA-OneVision},
3 author={Li, Bo and others},
4 journal={arXiv preprint arXiv:2408.03326},
5 year={2024}
6}
7
8@misc{hpml-egoqa-baseline,
9 title={HPML-EgoQA-Baseline: Finetuned LLaVA-OneVision for Egocentric Video QA},
10 author={Tandel, Sunidhi and Rahil and HPML Project Team},
11 year={2024},
12 howpublished={\url{https://huggingface.co/sunidhitandel/hpml-egoqa-baseline}},
13 note={HPML Project - High-Performance Machine Learning for Egocentric Vision}
14}
This work is part of the HPML (High-Performance Machine Learning) Project. We thank the LLaVA-OneVision team for providing the base model and the EgoIT-99K dataset contributors.
