Merge pull request #15 from togethercomputer/rahul_dev

New Dragonfly Architecture Merge

README.md

@@ -1,32 +1,37 @@
 <div align="center">
   <img src="assets/dragonfly_icon.png" alt="Dragonfly" style="width: 150px; display: block; margin-left: auto; margin-right: auto;" />
-  <h1>Dragonfly: Multi-Resolution Zoom Supercharges Large Visual-Language Model</h1>
+  <h1>Dragonfly: Multi-Resolution Zoom-In Encoding Enhances Vision-Language Models</h1>
 </div>
 
 ## 🔥 News
+- **Note**: We updated our codebase and arxiv paper with improved version of Dragonfly architecture. If you still want to use the old version of the code, it is still in [github branch](https://github.com/togethercomputer/Dragonfly/tree/dragonfly-v1).
 - [Our paper](https://arxiv.org/abs/2406.00977) is out on arxiv.
-- Check out [our blogpost](https://www.together.ai/blog/dragonfly-v1).
 - Our model checkpoints are out on huggingface 🤗 🚀: 
-    - General: [`togethercomputer/Llama-3-8B-Dragonfly-v1`](https://huggingface.co/togethercomputer/Llama-3-8B-Dragonfly-v1) 
-    - Biomed: [`togethercomputer/Llama-3-8B-Dragonfly-Med-v1`](https://huggingface.co/togethercomputer/Llama-3-8B-Dragonfly-Med-v1)
+    - General: [`togethercomputer/Llama-3.1-8B-Dragonfly-v2`](https://huggingface.co/togethercomputer/Llama-3.1-8B-Dragonfly-v2) 
+    - Biomed: [`togethercomputer/Llama-3.1-8B-Dragonfly-Med-v2`](https://huggingface.co/togethercomputer/Llama-3.1-8B-Dragonfly-Med-v2)
 
 
 ## 📖 Introduction
 
 ![Dragonfly framework](assets/model_overview.png)
 
-Recent advances in large multimodal models (LMMs) suggest that higher image resolution enhances the fine-grained understanding of image details, crucial for tasks such as visual commonsense reasoning and analyzing biomedical images. However, increasing input resolution poses two main challenges: 1) It extends the context length required by the language model, leading to inefficiencies and hitting the model's context limit; 2) It increases the complexity of visual features, necessitating more training data or more complex architecture. We introduce Dragonfly, a new LMM architecture that enhances fine-grained visual understanding and reasoning about image regions to address these challenges. Dragonfly employs two key strategies: multi-resolution visual encoding and zoom-in patch selection. These strategies allow the model to process high-resolution images efficiently while maintaining reasonable context length. Our experiments on eight popular benchmarks demonstrate that Dragonfly achieves competitive or better performance compared to other architectures, highlighting the effectiveness of our design. Additionally, we finetuned Dragonfly on biomedical instructions, achieving state-of-the-art results on multiple biomedical tasks requiring fine-grained visual understanding, including 92.3% accuracy on the Path-VQA dataset (compared to 83.3% for Med-Gemini) and the highest reported results on biomedical image captioning. To support model training, we curated a visual instruction-tuning dataset with 5.5 million image-instruction samples in the general domain and 1.4 million samples in the biomedical domain. We also conducted ablation studies to characterize the impact of various architectural designs and image resolutions, providing insights for future research on visual instruction alignment.
+Recent advances in vision-language models (VLMs) have demonstrated the advantages of processing images at higher resolutions and utilizing multi-crop features to preserve native resolution details. However, despite these improvements, existing vision transformers (ViTs) still struggle to capture fine-grained details from less prominent objects, charts, and embedded text, limiting their effectiveness in certain tasks. In this paper, we go beyond recent high-resolution and multi-crop techniques by not only preserving the native resolution, but zooming in beyond it and extracting features from a large number of image sub-crops. This enhancement allows our model to better capture fine-grained details, overcoming the limitations of current ViTs. To manage the increased token count and computational complexity, we demonstrate that a simple mean-pooling aggregation over tokens is effective. Our model, Dragonfly, achieves competitive performance on general-domain tasks such as ScienceQA and AI2D, and excels in tasks requiring fine-grained image understanding, including TextVQA and ChartQA. Among models in the 7-8B parameter range, Dragonfly consistently ranks at the top across ten general-domain benchmarks, achieving the highest or second-highest scores in most cases, outperforming models that are significantly larger or trained on larger datasets. Our biomedical version, Dragonfly-Med, sets new benchmarks on several medical tasks, achieving 91.6% accuracy on SLAKE (compared to 84.8% for Med-Gemini), 67.1% token F1 score on Path-VQA (compared to 62.7% for Med-PaLM M), and attains state-of-the-art results across the majority of performance metrics. Overall, our work highlights the persistent challenge of engineering visual representations with fixed-resolution ViTs, and proposes a simple yet effective solution to address this issue and boost performance in both general and specialized domains. 
 
+![Example Generations](assets/examples.png)
 
-# 📖 Table of Contents
-1. [Installation](#installation)
-2. [Checkpoint](#checkpoint)
-5. [Inference](#inference)
-3. [Dataset](#dataset)
-4. [Training](#training)
-6. [BibTeX](#bibtex)
-7. [Licence](#license)
 
+# 📖 Table of Contents
+- [📖 Table of Contents](#-table-of-contents)
+  - [💿 Installation](#-installation)
+  - [🏁 Checkpoint](#-checkpoint)
+  - [🧠 Inference](#-inference)
+  - [📊 Dataset](#-dataset)
+  - [🏋️‍♂️ Training](#️️-training)
+    - [Stage 1](#stage-1)
+    - [Stage 2](#stage-2)
+  - [🏆 Credits](#-credits)
+  - [📚 BibTeX](#-bibtex)
+  - [🪪 License](#-license)
 
 <a name="installation"/>
 
@@ -52,21 +57,21 @@ pip install --upgrade -e .
 
 ## 🏁 Checkpoint
 
-*Note: These models are released under [Llama 3 Community License Agreement](LICENSE)*
+*Note: These models are released under [Llama 3.1 Community License Agreement](LICENSE)*
 
-We release two huggingface model checkpoints: [`togethercomputer/Llama-3-8B-Dragonfly-v1`](https://huggingface.co/togethercomputer/Llama-3-8B-Dragonfly-v1) and [`togethercomputer/Llama-3-8B-Dragonfly-Med-v1`](https://huggingface.co/togethercomputer/Llama-3-8B-Dragonfly-Med-v1). Please follow the script [`test_dragonfly.py`](test_dragonfly.py) for more details. We provide a brief description on how to use them below.
+We release two huggingface model checkpoints: [`togethercomputer/Llama-3.1-8B-Dragonfly-v2`](https://huggingface.co/togethercomputer/Llama-3.1-8B-Dragonfly-v2) and [`togethercomputer/Llama-3.1-8B-Dragonfly-Med-v2`](https://huggingface.co/togethercomputer/Llama-3.1-8B-Dragonfly-Med-v2). Please follow the script [`test_dragonfly.py`](test_dragonfly.py) for more details. We provide a brief description on how to use them below.
 
 <a name="inference"/>
 
 ## 🧠 Inference
 
 If you have successfully completed the [Installation](#installation) process, then you should be able to follow the steps below. 
 
-We provide two test examples inside [`test_images`](test_images). 
+We provide two test examples inside [`assets`](assets). 
 
-Question: Summarize the visual content of the image.
+Question: What is so funny about this image?
 
-![Skateboard](test_images/skateboard.png)
+![Monalisa Dog](assets/monalisa_dog.jpg)
 
 Load necessary packages
 ```python
@@ -83,26 +88,26 @@ Instantiate the tokenizer, processor, and model.
 ```python
 device = torch.device("cuda:0")
 
-tokenizer = AutoTokenizer.from_pretrained("togethercomputer/Llama-3-8B-Dragonfly-v1")
-clip_processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+tokenizer = AutoTokenizer.from_pretrained("togethercomputer/Llama-3.1-8B-Dragonfly-v2")
+clip_processor = AutoProcessor.from_pretrained("openai/clip-vit-large-patch14-336")
 image_processor = clip_processor.image_processor
 processor = DragonflyProcessor(image_processor=image_processor, tokenizer=tokenizer, image_encoding_style="llava-hd")
 
-model = DragonflyForCausalLM.from_pretrained("togethercomputer/Llama-3-8B-Dragonfly-v1")
+model = DragonflyForCausalLM.from_pretrained("togethercomputer/Llama-3.1-8B-Dragonfly-v2")
 model = model.to(torch.bfloat16)
 model = model.to(device)
 ```
 
 Now, lets load the image and process them.
 ```python
-image = Image.open("./test_images/skateboard.png")
+image = Image.open("./assets/monalisa_dog.jpg")
 image = image.convert("RGB")
 images = [image]
 # images = [None] # if you do not want to pass any images
 
-text_prompt = "<|start_header_id|>user<|end_header_id|>\n\nSummarize the visual content of the image.<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+text_prompt = "<|start_header_id|>user<|end_header_id|>\n\nWhat is so funny about this image?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
 
-inputs = processor(text=[text_prompt], images=images, max_length=2048, return_tensors="pt", is_generate=True)
+inputs = processor(text=[text_prompt], images=images, max_length=4096, return_tensors="pt", is_generate=True)
 inputs = inputs.to(device)
 ```
 
@@ -118,11 +123,7 @@ generation_text = processor.batch_decode(generation_output, skip_special_tokens=
 
 An example response.
 ```plaintext
-In the heart of a vibrant skatepark, a skateboarder is caught in a moment of pure exhilaration. The skateboarder, dressed in a black t-shirt adorned with a yellow graphic and black pants, is suspended in mid-air, performing an impressive trick on a concrete ramp. The skateboarder's arms are outstretched, adding balance to the daring stunt.
-
-The skatepark itself is a concrete playground, with the skateboarder's ramp being the main focus. In the background, palm trees sway gently, adding a touch of nature to the urban setting. A few spectators can be seen in the distance, their attention riveted on the airborne skateboarder.
-
-The image captures not just a moment, but a story of skill, courage, and the joy of skateboarding.<|eot_id|>
+The humor in this image comes from the surreal juxtaposition of a dog's face with the body of the Mona Lisa, a famous painting by Leonardo da Vinci. The Mona Lisa is known for her enigmatic smile and is often considered one of the most famous paintings in the world. By combining the dog's face with the body of the Mona Lisa, the artist has created a whimsical and amusing image that plays on the viewer 's expectations and familiarity with the original paintings. The contrast between the dog's natural, expressive features and the serene, mysterious expression of the Mona Lisa creates a humerous effect that is likely to elicit laughter<|eot_id|>
 ```
 
 <a name="dataset"/>
@@ -179,7 +180,7 @@ Describe the content in the image.<|eot_id|><|start_header_id|>assistant<|end_he
 We would like to acknowledge the following resources that were instrumental in the development of Dragonfly:
 
 - [Meta Llama 3](https://huggingface.co/meta-llama/Meta-Llama-3-8B): We utilized the Llama 3 model as our foundational language model.
-- [CLIP](https://huggingface.co/openai/clip-vit-base-patch32): Our vision backbone is CLIP model from OpenAI. 
+- [CLIP](https://huggingface.co/openai/clip-vit-large-patch14-336): Our vision backbone is CLIP model from OpenAI. 
 - Our codebase is built upon the following two codebases:
   - [Otter: A Multi-Modal Model with In-Context Instruction Tuning](https://github.com/Luodian/Otter)
   - [LLaVA-UHD: an LMM Perceiving Any Aspect Ratio and High-Resolution Images](https://github.com/thunlp/LLaVA-UHD)
@@ -190,8 +191,8 @@ We would like to acknowledge the following resources that were instrumental in t
 
 ```bibtex
 @misc{chen2024dragonfly,
-      title={Dragonfly: Multi-Resolution Zoom Supercharges Large Visual-Language Model}, 
-      author={Kezhen Chen and Rahul Thapa and Rahul Chalamala and Ben Athiwaratkun and Shuaiwen Leon Song and James Zou},
+      title={Dragonfly: Multi-Resolution Zoom-In Encoding Enhances Vision-Language Models}, 
+      author={Rahul Thapa and Kezhen Chen and Ian Covert and Rahul Chalamala and Ben Athiwaratkun and Shuaiwen Leon Song and James Zou},
       year={2024},
       eprint={2406.00977},
       archivePrefix={arXiv},

environment.yml

@@ -2,8 +2,7 @@ name: dragonfly_env
 channels:
   - defaults
 dependencies:
-  - python=3.9
-  - conda-forge::openjdk
+  - python=3.10
   - pip
   - pip:
     - -r requirements.txt

pipeline/data_utils/data.py

@@ -33,13 +33,6 @@
 USER_AGENT = get_datasets_user_agent()
 
 Image.MAX_IMAGE_PIXELS = 1000000000
-MAX_NUM_TOKENS = 256
-MAX_NUM_IMAGES = 5
-TINY_IMAGE_SIZE_THRESHOLD = 1
-NUM_BACKUP_SPLIT = 5000
-N_CHANNELS = 3
-INTERLEAVED_IMAGE_SIZE = 224
-MIN_KB = 10
 
 IMAGE_CAP_INSTRUCT = [
     "Analyze the image in a comprehensive and detailed manner.",

pipeline/train/training.py

@@ -325,6 +325,7 @@ def train_one_epoch(
         total=total_training_steps,
         initial=current_global_steps,
     ):
+
         data_time_m.update(time.time() - end)
         global_step = num_steps + current_global_steps
 
@@ -423,6 +424,8 @@ def mask_embedding(m):
             lr_scheduler.step()
             optimizer.zero_grad()
 
+            # print(f"Step 3: Beginning Step: {num_steps}; Global Step: {global_step}")
+
             # step time and reset end outside of rank 0
             step_time_m.update(time.time() - end)
             end = time.time()

requirements.txt

@@ -1,40 +1,11 @@
-accelerate>=0.19.0
-braceexpand>=0.1.7
-einops>=0.6.1
-einops_exts>=0.0.4
-fastapi>=0.95.2
-gradio>=3.33.1
-huggingface_hub>=0.13.3
-importlib_metadata>=6.6.0
-inflection>=0.5.1
-markdown2>=2.4.8
-more_itertools>=9.1.0
-nltk>=3.8.1
-numpy>=1.23.5
-open_clip_torch>=2.16.0
-openai>=1.1.1
-opencv_python_headless>=4.5.5.64
-Pillow>=9.5.0
-pycocoevalcap>=1
-pycocotools>=2.0.6
-Requests>=2.31.0
-scipy>=1.10.1
-timm>=0.9.2
-tqdm>=4.65.0
-transformers==4.35.1
-uvicorn>=0.22.0
-webdataset>=0.2.48
-natsort>=8.4.0
-peft>=0.4.0
-ijson>=3.2.3
-yajl>=0.3.5
-deepspeed>=0.10.0
-wandb>=0.15.8
-trl>=0.5.0
-cffi>=1.15.1
-pyyaml>=6.0.1
-pytest>=7.4.2
-prettytable>=3.9.0
-datasets
+torch==2.4.1
+transformers==4.45.2
+datasets==3.0.1
+accelerate==1.0.1
+deepspeed==0.15.2
 packaging
-ninja
+ninja
+tqdm
+wandb
+numpy
+Pillow

setup.py

@@ -7,13 +7,13 @@
 
 setup(
     name="dragonfly",
-    version="0.1.0",
+    version="0.1.1",
     packages=find_packages(where="src"),
     package_dir={"": "src"},
     install_requires=requirements,
     author="Together AI",
-    author_email="kezhen@together.ai",
-    description="Dragonfly: Multi-Resolution Zoom Supercharges Large Visual-Language Model",
+    author_email="rthapa84@stanford.edu",
+    description="Dragonfly: Multi-Resolution Zoom-In Encoding Enhances Vision-Language Models",
     long_description=open("README.md").read(),
     long_description_content_type="text/markdown",
     url="https://github.com/togethercomputer/Dragonfly",