--- license: apache-2.0 library_name: transformers base_model: OpenGVLab/InternVL2-4B pipeline_tag: image-text-to-text --- # OS-Genesis: Automating GUI Agent Trajectory Construction via Reverse Task Synthesis
[\[🏠Homepage\]](https://qiushisun.github.io/OS-Genesis-Home/) [\[💻Code\]](https://github.com/OS-Copilot/OS-Genesis) [\[📝Paper\]](https://arxiv.org/abs/2412.19723) [\[🤗Models\]](https://huggingface.co/collections/OS-Copilot/os-genesis-6768d4b6fffc431dbf624c2d)[\[🤗Data\]](https://huggingface.co/collections/OS-Copilot/os-genesis-6768d4b6fffc431dbf624c2d)
## Overview ![os-genesis](https://cdn-uploads.huggingface.co/production/uploads/6064a0eeb1703ddba0d458b9/XvcAh92uvJQglmIu_L_nK.png) We introduce OS-Genesis, an interaction-driven pipeline that synthesizes high-quality and diverse GUI agent trajectory data without human supervision. By leveraging reverse task synthesis, OS-Genesis enables effective training of GUI agents to achieve superior performance on dynamic benchmarks such as AndroidWorld and WebArena. ## Quick Start OS-Genesis-8B-AC is a mobile action model finetuned from [InternVL2-8B](https://huggingface.co/OpenGVLab/InternVL2-8B). ### OS-Genesis AC Family Models In the following table, we provide an overview of the OS-Genesis AC Family Models used for evaluating the AndroidControl Benchmark. | Model Name | Base Model | Training Data | HF Link | | :-------------: | :-------------------------------------------------------------------------------------: | :----------------------------------------------------------------------------: | :---------------------------------------------------------: | | OS-Genesis-4B-AC | [InternVL2-4B](https://huggingface.co/OpenGVLab/InternVL2-4B) | [OS-Genesis-ac-training-data](https://huggingface.co/datasets/OS-Copilot/OS-Genesis-mobile-data/blob/main/os_genesis_ac_training_data.jsonl) | [🤗 link](https://huggingface.co/OS-Copilot/OS-Genesis-4B-AC) | | OS-Genesis-7B-AC | [Qwen2-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct) | [OS-Genesis-ac-training-data](https://huggingface.co/datasets/OS-Copilot/OS-Genesis-mobile-data/blob/main/os_genesis_ac_training_data.jsonl) | [🤗 link](https://huggingface.co/OS-Copilot/OS-Genesis-7B-AC) | | OS-Genesis-8B-AC | [InternVL2-8B](https://huggingface.co/OpenGVLab/InternVL2-8B) | [OS-Genesis-ac-training-data](https://huggingface.co/datasets/OS-Copilot/OS-Genesis-mobile-data/blob/main/os_genesis_ac_training_data.jsonl) | [🤗 link](https://huggingface.co/OS-Copilot/OS-Genesis-8B-AC) | ### Inference Example First, install the `transformers` library: ``` pip install transformers ``` For additional dependencies, please refer to the [InternVL2 documentation](https://internvl.readthedocs.io/en/latest/get_started/installation.html). For evaluating the AndroidControl Benchmark, please refer to the [**evaluation code**](https://github.com/OS-Copilot/OS-Genesis/tree/main/evaluation/android_control). Inference code example: ```python import numpy as np import torch import torchvision.transforms as T from PIL import Image from torchvision.transforms.functional import InterpolationMode from transformers import AutoModel, AutoTokenizer IMAGENET_MEAN = (0.485, 0.456, 0.406) IMAGENET_STD = (0.229, 0.224, 0.225) def build_transform(input_size): MEAN, STD = IMAGENET_MEAN, IMAGENET_STD transform = T.Compose([ T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img), T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC), T.ToTensor(), T.Normalize(mean=MEAN, std=STD) ]) return transform def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size): best_ratio_diff = float('inf') best_ratio = (1, 1) area = width * height for ratio in target_ratios: target_aspect_ratio = ratio[0] / ratio[1] ratio_diff = abs(aspect_ratio - target_aspect_ratio) if ratio_diff < best_ratio_diff: best_ratio_diff = ratio_diff best_ratio = ratio elif ratio_diff == best_ratio_diff: if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]: best_ratio = ratio return best_ratio def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False): orig_width, orig_height = image.size aspect_ratio = orig_width / orig_height # calculate the existing image aspect ratio target_ratios = set( (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if i * j <= max_num and i * j >= min_num) target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1]) # find the closest aspect ratio to the target target_aspect_ratio = find_closest_aspect_ratio( aspect_ratio, target_ratios, orig_width, orig_height, image_size) # calculate the target width and height target_width = image_size * target_aspect_ratio[0] target_height = image_size * target_aspect_ratio[1] blocks = target_aspect_ratio[0] * target_aspect_ratio[1] # resize the image resized_img = image.resize((target_width, target_height)) processed_images = [] for i in range(blocks): box = ( (i % (target_width // image_size)) * image_size, (i // (target_width // image_size)) * image_size, ((i % (target_width // image_size)) + 1) * image_size, ((i // (target_width // image_size)) + 1) * image_size ) # split the image split_img = resized_img.crop(box) processed_images.append(split_img) assert len(processed_images) == blocks if use_thumbnail and len(processed_images) != 1: thumbnail_img = image.resize((image_size, image_size)) processed_images.append(thumbnail_img) return processed_images def load_image(image_file, input_size=448, max_num=12): image = Image.open(image_file).convert('RGB') transform = build_transform(input_size=input_size) images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num) pixel_values = [transform(image) for image in images] pixel_values = torch.stack(pixel_values) return pixel_values # If you want to load a model using multiple GPUs, please refer to the `Multiple GPUs` section. path = 'OS-Copilot/OS-Genesis-8B-AC' model = AutoModel.from_pretrained( path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, trust_remote_code=True).eval().cuda() tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False) # set the max number of tiles in `max_num` pixel_values = load_image('./web_dfacd48d-d2c2-492f-b94c-41e6a34ea99f.png', max_num=6).to(torch.bfloat16).cuda() generation_config = dict(max_new_tokens=1024, do_sample=True) question = "\nYou are a GUI task expert, I will provide you with a high-level instruction, an action history, a screenshot with its corresponding accessibility tree.\n High-level instruction: {high_level_instruction}\n Action history: {action_history}\n Accessibility tree: {a11y_tree}\n Please generate the low-level thought and action for the next step." response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True) print(f'User: {question}\nAssistant: {response}') ``` ## Citation If you find this repository helpful, feel free to cite our paper: ```bibtex @article{sun2024genesis, title={OS-Genesis: Automating GUI Agent Trajectory Construction via Reverse Task Synthesis}, author={Sun, Qiushi and Cheng, Kanzhi and Ding, Zichen and Jin, Chuanyang and Wang, Yian and Xu, Fangzhi and Wu, Zhenyu and Jia, Chengyou and Chen, Liheng and Liu, Zhoumianze and others}, journal={arXiv preprint arXiv:2412.19723}, year={2024} } ```