Florence-2-large-safetensors

Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks

Changes to Original

This is a modified version of Florence-2-large-ft. The original failed the conversion to safetensors using the Huggingface Space Safetensors/convert due to inconsistencies with some of the data pointers. The convert_to_safetensors.py script does the minimal steps to enable conversion to safetensors. It does check that the resulting tensors are equal and validates against the single image listed below to ensure both .bin and .safetensors provide the same output.

Only the modeling_florence2.py file has been modified:

• Added Florence2LanguageForConditionalGeneration._tie_weights() which was missing
• Added GenerationMixin as a parent class to Florence2LanguageForConditionalGeneration and Florence2ForConditionalGeneration to stop the deprecation warning that PreTrainedModel will NOT inherit from GenerationMixin from v4.50 onwards.
• Added @torch.no_grad() decorator to the generate() function to follow standard transformers usage of turning off gradient accumulation. Without this, VRAM usage when up significantly each time the batch size would increase by 1.
• Fixed batch attention masking following pawlowskipawel's PR
• Changed importing of timm following deprecation warnings.
• Thanks to @mrhendrey for reference

Model Summary

This is a continued pretrained version of Florence-2-large model with 4k context length, only 0.1B samples are used for continue pretraining, thus it might not be trained well. In addition, OCR task has been updated with line separator ('\n'). COCO OD AP 39.8

This Hub repository contains a HuggingFace's transformers implementation of Florence-2 model from Microsoft.

Florence-2 is an advanced vision foundation model that uses a prompt-based approach to handle a wide range of vision and vision-language tasks. Florence-2 can interpret simple text prompts to perform tasks like captioning, object detection, and segmentation. It leverages our FLD-5B dataset, containing 5.4 billion annotations across 126 million images, to master multi-task learning. The model's sequence-to-sequence architecture enables it to excel in both zero-shot and fine-tuned settings, proving to be a competitive vision foundation model.

Resources and Technical Documentation:

• Florence-2 technical report.
• Jupyter Notebook for inference and visualization of Florence-2-large

How to Get Started with the Model

Use the code below to get started with the model. All models are trained with float16.

import requests

import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM 


device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model = AutoModelForCausalLM.from_pretrained("microsoft/Florence-2-large", torch_dtype=torch_dtype, trust_remote_code=True).to(device)
processor = AutoProcessor.from_pretrained("microsoft/Florence-2-large", trust_remote_code=True)

prompt = "<OD>"

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true"
image = Image.open(requests.get(url, stream=True).raw)

inputs = processor(text=prompt, images=image, return_tensors="pt").to(device, torch_dtype)

generated_ids = model.generate(
    input_ids=inputs["input_ids"],
    pixel_values=inputs["pixel_values"],
    max_new_tokens=4096,
    num_beams=3,
    do_sample=False
)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]

parsed_answer = processor.post_process_generation(generated_text, task="<OD>", image_size=(image.width, image.height))

print(parsed_answer)

Tasks

This model is capable of performing different tasks through changing the prompts.

First, let's define a function to run a prompt.

<details> <summary> Click to expand </summary>

import requests

import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM 

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model = AutoModelForCausalLM.from_pretrained("microsoft/Florence-2-large", torch_dtype=torch_dtype, trust_remote_code=True).to(device)
processor = AutoProcessor.from_pretrained("microsoft/Florence-2-large", trust_remote_code=True)

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true"
image = Image.open(requests.get(url, stream=True).raw)

def run_example(task_prompt, text_input=None):
    if text_input is None:
        prompt = task_prompt
    else:
        prompt = task_prompt + text_input
    inputs = processor(text=prompt, images=image, return_tensors="pt").to(device, torch_dtype)
    generated_ids = model.generate(
      input_ids=inputs["input_ids"],
      pixel_values=inputs["pixel_values"],
      max_new_tokens=1024,
      num_beams=3
    )
    generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]

    parsed_answer = processor.post_process_generation(generated_text, task=task_prompt, image_size=(image.width, image.height))

    print(parsed_answer)

</details>

Here are the tasks Florence-2 could perform:

<details> <summary> Click to expand </summary>

Caption

prompt = "<CAPTION>"
run_example(prompt)

Detailed Caption

prompt = "<DETAILED_CAPTION>"
run_example(prompt)

More Detailed Caption

prompt = "<MORE_DETAILED_CAPTION>"
run_example(prompt)

Caption to Phrase Grounding

caption to phrase grounding task requires additional text input, i.e. caption.

Caption to phrase grounding results format: {'<CAPTION_TO_PHRASE_GROUNDING>': {'bboxes': [[x1, y1, x2, y2], ...], 'labels': ['', '', ...]}}

task_prompt = "<CAPTION_TO_PHRASE_GROUNDING>"
results = run_example(task_prompt, text_input="A green car parked in front of a yellow building.")

Object Detection

OD results format: {'<OD>': {'bboxes': [[x1, y1, x2, y2], ...], 'labels': ['label1', 'label2', ...]} }

prompt = "<OD>"
run_example(prompt)

Dense Region Caption

Dense region caption results format: {'<DENSE_REGION_CAPTION>' : {'bboxes': [[x1, y1, x2, y2], ...], 'labels': ['label1', 'label2', ...]} }

prompt = "<DENSE_REGION_CAPTION>"
run_example(prompt)

Region proposal

Dense region caption results format: {'<REGION_PROPOSAL>': {'bboxes': [[x1, y1, x2, y2], ...], 'labels': ['', '', ...]}}

prompt = "<REGION_PROPOSAL>"
run_example(prompt)

OCR

prompt = "<OCR>"
run_example(prompt)

OCR with Region

OCR with region output format: {'<OCR_WITH_REGION>': {'quad_boxes': [[x1, y1, x2, y2, x3, y3, x4, y4], ...], 'labels': ['text1', ...]}}

prompt = "<OCR_WITH_REGION>"
run_example(prompt)

Output confidence score with Object Detection

def run_example_with_score(task_prompt, text_input=None):
    if text_input is None:
        prompt = task_prompt
    else:
        prompt = task_prompt + text_input
    inputs = processor(text=prompt, images=image, return_tensors="pt").to(device, torch_dtype)
    generated_ids = model.generate(
      input_ids=inputs["input_ids"],
      pixel_values=inputs["pixel_values"],
      max_new_tokens=1024,
      num_beams=3,
      return_dict_in_generate=True,
      output_scores=True,
    )
    generated_text = processor.batch_decode(generated_ids.sequences, skip_special_tokens=False)[0]

    prediction, scores, beam_indices = generated_ids.sequences, generated_ids.scores, generated_ids.beam_indices
    transition_beam_scores = model.compute_transition_scores(
        sequences=prediction,
        scores=scores,
        beam_indices=beam_indices,
    )

    parsed_answer = processor.post_process_generation(sequence=generated_ids.sequences[0], 
        transition_beam_score=transition_beam_scores[0],
        task=task_prompt, image_size=(image.width, image.height)
    )

    print(parsed_answer)

prompt = "<OD>"
run_example_with_score(prompt)

for More detailed examples, please refer to notebook

</details>

Benchmarks

Florence-2 Zero-shot performance

The following table presents the zero-shot performance of generalist vision foundation models on image captioning and object detection evaluation tasks. These models have not been exposed to the training data of the evaluation tasks during their training phase.

The following table continues the comparison with performance on other vision-language evaluation tasks.

Florence-2 finetuned performance

We finetune Florence-2 models with a collection of downstream tasks, resulting two generalist models Florence-2-base-ft and Florence-2-large-ft that can conduct a wide range of downstream tasks.

The table below compares the performance of specialist and generalist models on various captioning and Visual Question Answering (VQA) tasks. Specialist models are fine-tuned specifically for each task, whereas generalist models are fine-tuned in a task-agnostic manner across all tasks. The symbol "▲" indicates the usage of external OCR as input.

BibTex and citation info

@article{xiao2023florence,
  title={Florence-2: Advancing a unified representation for a variety of vision tasks},
  author={Xiao, Bin and Wu, Haiping and Xu, Weijian and Dai, Xiyang and Hu, Houdong and Lu, Yumao and Zeng, Michael and Liu, Ce and Yuan, Lu},
  journal={arXiv preprint arXiv:2311.06242},
  year={2023}
}