📗 中文版README

🌟 Jump to detailed introduction: Introduction to CogVLM， 🆕 Introduction to CogAgent

📔 For more detailed usage information, please refer to: CogVLM & CogAgent's technical documentation (in Chinese)

Table of Contents

• CogVLM & CogAgent
  • Release
  • Get Started
    • Option 1: Inference Using Web Demo.
    • Option 2：Deploy CogVLM / CogAgent by yourself
      • Situation 2.1 CLI (SAT version)
      • Situation 2.2 CLI (Huggingface version)
      • Situation 2.3 Web Demo
    • Option 3：Finetuning CogAgent / CogVLM
    • Option 4: OpenAI Vision format
    • Hardware requirement
    • Model checkpoints
  • Introduction to CogVLM
    • Examples
  • Introduction to CogAgent
    • GUI Agent Examples
  • Cookbook
    • Task Prompts
    • Which --version to use
    • FAQ
  • License
  • Citation & Acknowledgements

• 🔥🔥🔥 News: 2024/5/20: We released the next generation of model, CogVLM2, which is based on llama3-8b and on the par of (or better than) GPT-4V in most cases! DOWNLOAD and TRY!

• 🔥🔥 News: 2024/4/5: CogAgent was selected as a CVPR 2024 Highlights!

• 🔥 News: 2023/12/26: We have released the CogVLM-SFT-311K dataset, which contains over 150,000 pieces of data that we used for CogVLM v1.0 only training. Welcome to follow and use.

• News: 2023/12/18: New Web UI Launched! We have launched a new web UI based on Streamlit, users can painlessly talk to CogVLM, CogAgent in our UI. Have a better user experience.

• News: 2023/12/15: CogAgent Officially Launched! CogAgent is an image understanding model developed based on CogVLM. It features visual-based GUI Agent capabilities and has further enhancements in image understanding. It supports image input with a resolution of 1120*1120, and possesses multiple abilities including multi-turn dialogue with images, GUI Agent, Grounding, and more.

• News: 2023/12/8 We have updated the checkpoint of cogvlm-grounding-generalist to cogvlm-grounding-generalist-v1.1, with image augmentation during training, therefore more robust. See details.

• News: 2023/12/7 CogVLM supports 4-bit quantization now! You can inference with just 11GB GPU memory!

• News: 2023/11/20 We have updated the checkpoint of cogvlm-chat to cogvlm-chat-v1.1, unified the versions of chat and VQA, and refreshed the SOTA on various datasets. See details

• News: 2023/11/20 We release cogvlm-chat, cogvlm-grounding-generalist/base, cogvlm-base-490/224 on 🤗Huggingface. you can infer with transformers in a few lines of codenow!

• 2023/10/27 CogVLM bilingual version is available online! Welcome to try it out!

• 2023/10/5 CogVLM-17B released。

• Click here to enter CogVLM2 Demo。

If you need to use Agent and Grounding functions, please refer to Cookbook - Task Prompts

We support two GUIs for model inference, CLI and web demo . If you want to use it in your python code, it is easy to modify the CLI scripts for your case.

First, we need to install the dependencies.

# CUDA >= 11.8
pip install -r requirements.txt
python -m spacy download en_core_web_sm

All code for inference is located under the basic_demo/ directory. Please switch to this directory first before proceeding with further operations.

Run CLI demo via:

# CogAgent
python cli_demo_sat.py --from_pretrained cogagent-chat --version chat --bf16  --stream_chat
python cli_demo_sat.py --from_pretrained cogagent-vqa --version chat_old --bf16  --stream_chat

# CogVLM
python cli_demo_sat.py --from_pretrained cogvlm-chat --version chat_old --bf16  --stream_chat
python cli_demo_sat.py --from_pretrained cogvlm-grounding-generalist --version base --bf16  --stream_chat

The program will automatically download the sat model and interact in the command line. You can generate replies by entering instructions and pressing enter. Enter clear to clear the conversation history and stop to stop the program.

We also support model parallel inference, which splits model to multiple (2/4/8) GPUs. --nproc-per-node=[n] in the following command controls the number of used GPUs.

torchrun --standalone --nnodes=1 --nproc-per-node=2 cli_demo_sat.py --from_pretrained cogagent-chat --version chat --bf16

• If you want to manually download the weights, you can replace the path after --from_pretrained with the model path.

• Our model supports SAT's 4-bit quantization and 8-bit quantization. You can change --bf16 to --fp16, or --fp16 --quant 4, or --fp16 --quant 8.

  For example

  python cli_demo_sat.py --from_pretrained cogagent-chat --fp16 --quant 8 --stream_chat
  python cli_demo_sat.py --from_pretrained cogvlm-chat-v1.1 --fp16 --quant 4 --stream_chat
  # In SAT version，--quant should be used with --fp16

• The program provides the following hyperparameters to control the generation process:

  usage: cli_demo_sat.py [-h] [--max_length MAX_LENGTH] [--top_p TOP_P] [--top_k TOP_K] [--temperature TEMPERATURE]
  
  optional arguments:
  -h, --help            show this help message and exit
  --max_length MAX_LENGTH
                          max length of the total sequence
  --top_p TOP_P         top p for nucleus sampling
  --top_k TOP_K         top k for top k sampling
  --temperature TEMPERATURE
                          temperature for sampling
  

• Click here to view the correspondence between different models and the --version parameter.

Run CLI demo via:

# CogAgent
python cli_demo_hf.py --from_pretrained THUDM/cogagent-chat-hf --bf16
python cli_demo_hf.py --from_pretrained THUDM/cogagent-vqa-hf --bf16

# CogVLM
python cli_demo_hf.py --from_pretrained THUDM/cogvlm-chat-hf --bf16
python cli_demo_hf.py --from_pretrained THUDM/cogvlm-grounding-generalist-hf --bf16

• If you want to manually download the weights, you can replace the path after --from_pretrained with the model path.

• You can change --bf16 to --fp16, or --quant 4. For example, our model supports Huggingface's 4-bit quantization:

  python cli_demo_hf.py --from_pretrained THUDM/cogvlm-chat-hf --quant 4

We also offer a local web demo based on Gradio. First, install Gradio by running: pip install gradio. Then download and enter this repository and run web_demo.py. See the next section for detailed usage:

python web_demo.py --from_pretrained cogagent-chat --version chat --bf16
python web_demo.py --from_pretrained cogagent-vqa --version chat_old --bf16
python web_demo.py --from_pretrained cogvlm-chat-v1.1 --version chat_old --bf16
python web_demo.py --from_pretrained cogvlm-grounding-generalist --version base --bf16

The GUI of the web demo looks like:

You may want to use CogVLM in your own task, which needs a different output style or domain knowledge. All code for finetuning is located under the finetune_demo/ directory.

We here provide a finetuning example for Captcha Recognition using lora.

1. Start by downloading the Captcha Images dataset. Once downloaded, extract the contents of the ZIP file.

2. To create a train/validation/test split in the ratio of 80/5/15, execute the following:

   python utils/split_dataset.py

3. Start the fine-tuning process with this command:

   bash finetune_demo/finetune_(cogagent/cogvlm)_lora.sh

4. Merge the model to model_parallel_size=1: (replace the 4 below with your training MP_SIZE)

   torchrun --standalone --nnodes=1 --nproc-per-node=4 utils/merge_model.py --version base --bf16 --from_pretrained ./checkpoints/merged_lora_(cogagent/cogvlm490/cogvlm224)

5. Evaluate the performance of your model.

   bash finetune_demo/evaluate_(cogagent/cogvlm).sh

We provide the same API examples as GPT-4V, which you can view in openai_demo.

1. First, start the node

python openai_demo/openai_api.py

2. Next, run the request example node, which is an example of a continuous dialogue

python openai_demo/openai_api_request.py

3. You will get output similar to the following

This image showcases a tranquil natural scene with a wooden pathway leading through a field of lush green grass. In the distance, there are trees and some scattered structures, possibly houses or small buildings. The sky is clear with a few scattered clouds, suggesting a bright and sunny day.

• Model Inference:

  For INT4 quantization: 1 * RTX 3090(24G) (CogAgent takes ~ 12.6GB, CogVLM takes ~ 11GB)

  For FP16: 1 * A100(80G) or 2 * RTX 3090(24G)

• Finetuning:

  For FP16: 4 * A100(80G) [Recommend] or 8* RTX 3090(24G).

If you run the basic_demo/cli_demo*.py from the code repository, it will automatically download SAT or Hugging Face weights. Alternatively, you can choose to manually download the necessary weights.

• CogAgent

  Model name	Input resolution	Introduction	Huggingface model	SAT model
  cogagent-chat	1120	Chat version of CogAgent. Supports GUI Agent, multiple-round chat and visual grounding.	HF link OpenXLab link	HF link OpenXLab link
  cogagent-vqa	1120	VQA version of CogAgent. Has stronger capabilities in single-turn visual dialogue. Recommended for VQA benchmarks.	HF link OpenXLab link	HF link OpenXLab link

c

• CogVLM

  Model name	Input resolution	Introduction	Huggingface model	SAT model
  cogvlm-chat-v1.1	490	Supports multiple rounds of chat and vqa simultaneously, with different prompts.	HF link OpenXLab link	HF link OpenXLab link
  cogvlm-base-224	224	The original checkpoint after text-image pretraining.	HF link OpenXLab link	HF link OpenXLab link
  cogvlm-base-490	490	Amplify the resolution to 490 through position encoding interpolation from cogvlm-base-224.	HF link OpenXLab link	HF link OpenXLab link
  cogvlm-grounding-generalist	490	This checkpoint supports different visual grounding tasks, e.g. REC, Grounding Captioning, etc.	HF link OpenXLab link	HF link OpenXLab link

• CogVLM is a powerful open-source visual language model (VLM). CogVLM-17B has 10 billion vision parameters and 7 billion language parameters.

• CogVLM-17B achieves state-of-the-art performance on 10 classic cross-modal benchmarks, including NoCaps, Flicker30k captioning, RefCOCO, RefCOCO+, RefCOCOg, Visual7W, GQA, ScienceQA, VizWiz VQA and TDIUC, and rank the 2nd on VQAv2, OKVQA, TextVQA, COCO captioning, etc., surpassing or matching PaLI-X 55B. CogVLM can also chat with you about images.

• CogVLM can accurately describe images in details with very few hallucinations.

  Click for comparison with LLAVA-1.5 and MiniGPT-4.

  
  
  

• CogVLM can understand and answer various types of questions, and has a visual grounding version.

• CogVLM sometimes captures more detailed content than GPT-4V(ision).

CogAgent is an open-source visual language model improved based on CogVLM. CogAgent-18B has 11 billion visual parameters and 7 billion language parameters

CogAgent-18B achieves state-of-the-art generalist performance on 9 classic cross-modal benchmarks, including VQAv2, OK-VQ, TextVQA, ST-VQA, ChartQA, infoVQA, DocVQA, MM-Vet, and POPE. It significantly surpasses existing models on GUI operation datasets such as AITW and Mind2Web.

In addition to all the features already present in CogVLM (visual multi-round dialogue, visual grounding), CogAgent:

1. Supports higher resolution visual input and dialogue question-answering. It supports ultra-high-resolution image inputs of 1120x1120.

2. Possesses the capabilities of a visual Agent, being able to return a plan, next action, and specific operations with coordinates for any given task on any GUI screenshot.

3. Enhanced GUI-related question-answering capabilities, allowing it to handle questions about any GUI screenshot, such as web pages, PC apps, mobile applications, etc.

4. Enhanced capabilities in OCR-related tasks through improved pre-training and fine-tuning.

1. General Multi-Round Dialogue: Say whatever you want.

2. GUI Agent Task: Use the Agent template and replace <TASK> with the task instruction enclosed in double quotes. This query can make CogAgent infer Plan and Next Action. If adding (with grounding) at the end of the query, the model will return a formalized action representation with coordinates.

For example, to ask the model how to complete the task "Search for CogVLM" on a current GUI screenshot, follow these steps:

1. Randomly select a template from the Agent template. Here, we choose What steps do I need to take to <TASK>?.

2. Replace with the task instruction enclosed in double quotes, for example, What steps do I need to take to "Search for CogVLM"? . Inputting this to the model yields:

Plan: 1. Type 'CogVLM' into the Google search bar. 2. Review the search results that appear. 3. Click on a relevant result to read more about CogVLM or access further resources.

Next Action: Move the cursor to the Google search bar, and type 'CogVLM' into it.

3. If adding (with grounding) at the end, i.e. changing the input to What steps do I need to take to "Search for CogVLM"?(with grounding), the output of CogAgent would be:

Plan: 1. Type 'CogVLM' into the Google search bar. 2. Review the search results that appear. 3. Click on a relevant result to read more about CogVLM or access further resources.

Next Action: Move the cursor to the Google search bar, and type 'CogVLM' into it. Grounded Operation:[combobox] Search -> TYPE: CogVLM at the box [[212,498,787,564]]

Tip: For GUI Agent tasks, it is recommended to conduct only single-round dialogues for each image for better results.

3. Visual Grounding. Three modes of grounding are supported:

   • Image description with grounding coordinates (bounding box). Use any template from caption_with_box template as model input. For example:

   Can you provide a description of the image and include the coordinates [[x0,y0,x1,y1]] for each mentioned object?

   • Returning grounding coordinates (bounding box) based on the description of objects. Use any template from caption2box template, replacing <expr> with the object's description. For example:

   Can you point out children in blue T-shirts in the image and provide the bounding boxes of their location?

   • Providing a description based on bounding box coordinates. Use a template from box2caption template, replacing <objs> with the position coordinates. For example:

   Tell me what you see within the designated area [[086,540,400,760]] in the picture.

Format of coordination: The bounding box coordinates in the model's input and output use the format [[x1, y1, x2, y2]], with the origin at the top left corner, the x-axis to the right, and the y-axis downward. (x1, y1) and (x2, y2) are the top-left and bottom-right corners, respectively, with values as relative coordinates multiplied by 1000 (prefixed with zeros to three digits).

Due to differences in model functionalities, different model versions may have distinct --version specifications for the text processor, meaning the format of the prompts used varies.

• If you have trouble in accessing huggingface.co, you can add --local_tokenizer /path/to/vicuna-7b-v1.5 to load the tokenizer.
• If you have trouble in automatically downloading model with 🔨SAT, try downloading from 🤖modelscope or 🤗huggingface or 💡wisemodel manually.
• Download model using 🔨SAT, the model will be saved to the default location ~/.sat_models. Change the default location by setting the environment variable SAT_HOME. For example, if you want to save the model to /path/to/my/models, you can run export SAT_HOME=/path/to/my/models before running the python command.

The code in this repository is open source under the Apache-2.0 license, while the use of the CogVLM model weights must comply with the Model License.

If you find our work helpful, please consider citing the following papers

@misc{wang2023cogvlm,
      title={CogVLM: Visual Expert for Pretrained Language Models}, 
      author={Weihan Wang and Qingsong Lv and Wenmeng Yu and Wenyi Hong and Ji Qi and Yan Wang and Junhui Ji and Zhuoyi Yang and Lei Zhao and Xixuan Song and Jiazheng Xu and Bin Xu and Juanzi Li and Yuxiao Dong and Ming Ding and Jie Tang},
      year={2023},
      eprint={2311.03079},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{hong2023cogagent,
      title={CogAgent: A Visual Language Model for GUI Agents}, 
      author={Wenyi Hong and Weihan Wang and Qingsong Lv and Jiazheng Xu and Wenmeng Yu and Junhui Ji and Yan Wang and Zihan Wang and Yuxiao Dong and Ming Ding and Jie Tang},
      year={2023},
      eprint={2312.08914},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

In the instruction fine-tuning phase of the CogVLM, there are some English image-text data from the MiniGPT-4, LLAVA, LRV-Instruction, LLaVAR and Shikra projects, as well as many classic cross-modal work datasets. We sincerely thank them for their contributions.