Chapter 26

本地推理 Phi-3-Vision

Phi-3-vision-128k-instruct 不仅让 Phi-3 能理解语言，还能“看见”世界。通过 Phi-3-vision-128k-instruct，我们可以解决各种视觉问题，比如 OCR、表格分析、物体识别、图片描述等。许多以前需要大量数据训练才能完成的任务，现在都能轻松实现。以下是 Phi-3-vision-128k-instruct 相关的技术和应用场景。

0. 准备工作

请确保已安装以下 Python 库（推荐使用 Python 3.10+）

bash

pip install transformers -U
pip install datasets -U
pip install torch -U

建议使用 CUDA 11.6+ 并安装 flatten

bash

pip install flash-attn --no-build-isolation

新建一个 Notebook。为了完成示例，建议先创建以下内容。

python

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

model_id = "microsoft/Phi-3-vision-128k-instruct"

kwargs = {}
kwargs['torch_dtype'] = torch.bfloat16

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, torch_dtype="auto").cuda()

user_prompt = '<|user|>\n'
assistant_prompt = '<|assistant|>\n'
prompt_suffix = "<|end|>\n"

1. 使用 Phi-3-Vision 分析图片

我们希望 AI 能分析图片内容并给出相关描述

python

prompt = f"{user_prompt}<|image_1|>\nCould you please introduce this stock to me?{prompt_suffix}{assistant_prompt}"

url = "https://g.foolcdn.com/editorial/images/767633/nvidiadatacenterrevenuefy2017tofy2024.png"

image = Image.open(requests.get(url, stream=True).raw)

inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")

generate_ids = model.generate(**inputs, 
                              max_new_tokens=1000,
                              eos_token_id=processor.tokenizer.eos_token_id,
                              )
generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]

response = processor.batch_decode(generate_ids, 
                                  skip_special_tokens=True, 
                                  clean_up_tokenization_spaces=False)[0]

在 Notebook 中执行以下脚本即可获得相关答案

txt

Certainly! Nvidia Corporation is a global leader in advanced computing and artificial intelligence (AI). The company designs and develops graphics processing units (GPUs), which are specialized hardware accelerators used to process and render images and video. Nvidia's GPUs are widely used in professional visualization, data centers, and gaming. The company also provides software and services to enhance the capabilities of its GPUs. Nvidia's innovative technologies have applications in various industries, including automotive, healthcare, and entertainment. The company's stock is publicly traded and can be found on major stock exchanges.

2. 使用 Phi-3-Vision 进行 OCR

除了分析图片，我们还可以从图片中提取信息。这就是 OCR 过程，以前需要写复杂代码才能完成。

python

prompt = f"{user_prompt}<|image_1|>\nHelp me get the title and author information of this book?{prompt_suffix}{assistant_prompt}"

url = "https://marketplace.canva.com/EAFPHUaBrFc/1/0/1003w/canva-black-and-white-modern-alone-story-book-cover-QHBKwQnsgzs.jpg"

image = Image.open(requests.get(url, stream=True).raw)

inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")

generate_ids = model.generate(**inputs, 
                              max_new_tokens=1000,
                              eos_token_id=processor.tokenizer.eos_token_id,
                              )

generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]

response = processor.batch_decode(generate_ids, 
                                  skip_special_tokens=False, 
                                  clean_up_tokenization_spaces=False)[0]

结果如下

txt

The title of the book is "ALONE" and the author is Morgan Maxwell.

3. 多图比较

Phi-3 Vision 支持多张图片的比较。我们可以用这个模型找出图片之间的差异。

python

prompt = f"{user_prompt}<|image_1|>\n<|image_2|>\n What is difference in this two images?{prompt_suffix}{assistant_prompt}"

print(f">>> Prompt\n{prompt}")

url = "https://hinhnen.ibongda.net/upload/wallpaper/doi-bong/2012/11/22/arsenal-wallpaper-free.jpg"

image_1 = Image.open(requests.get(url, stream=True).raw)

url = "https://assets-webp.khelnow.com/d7293de2fa93b29528da214253f1d8d0/news/uploads/2021/07/Arsenal-1024x576.jpg.webp"

image_2 = Image.open(requests.get(url, stream=True).raw)

images = [image_1, image_2]

inputs = processor(prompt, images, return_tensors="pt").to("cuda:0")

generate_ids = model.generate(**inputs, 
                              max_new_tokens=1000,
                              eos_token_id=processor.tokenizer.eos_token_id,
                              )

generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]

response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]

结果如下

txt

The first image shows a group of soccer players from the Arsenal Football Club posing for a team photo with their trophies, while the second image shows a group of soccer players from the Arsenal Football Club celebrating a victory with a large crowd of fans in the background. The difference between the two images is the context in which the photos were taken, with the first image focusing on the team and their trophies, and the second image capturing a moment of celebration and victory.

免责声明：
本文件使用 AI 翻译服务 Co-op Translator 进行翻译。虽然我们力求准确，但请注意，自动翻译可能包含错误或不准确之处。原始语言的原文应被视为权威来源。对于重要信息，建议采用专业人工翻译。对于因使用本翻译而产生的任何误解或误释，我们不承担任何责任。

Chapter 26

本地推理 Phi-3-Vision

0. 准备工作

请确保已安装以下 Python 库（推荐使用 Python 3.10+）

bash

pip install transformers -U
pip install datasets -U
pip install torch -U

建议使用 CUDA 11.6+ 并安装 flatten

bash

pip install flash-attn --no-build-isolation

新建一个 Notebook。为了完成示例，建议先创建以下内容。

python

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

model_id = "microsoft/Phi-3-vision-128k-instruct"

kwargs = {}
kwargs['torch_dtype'] = torch.bfloat16

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, torch_dtype="auto").cuda()

user_prompt = '<|user|>\n'
assistant_prompt = '<|assistant|>\n'
prompt_suffix = "<|end|>\n"

1. 使用 Phi-3-Vision 分析图片

我们希望 AI 能分析图片内容并给出相关描述

python

prompt = f"{user_prompt}<|image_1|>\nCould you please introduce this stock to me?{prompt_suffix}{assistant_prompt}"

url = "https://g.foolcdn.com/editorial/images/767633/nvidiadatacenterrevenuefy2017tofy2024.png"

image = Image.open(requests.get(url, stream=True).raw)

inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")

generate_ids = model.generate(**inputs, 
                              max_new_tokens=1000,
                              eos_token_id=processor.tokenizer.eos_token_id,
                              )
generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]

response = processor.batch_decode(generate_ids, 
                                  skip_special_tokens=True, 
                                  clean_up_tokenization_spaces=False)[0]

在 Notebook 中执行以下脚本即可获得相关答案

txt

Certainly! Nvidia Corporation is a global leader in advanced computing and artificial intelligence (AI). The company designs and develops graphics processing units (GPUs), which are specialized hardware accelerators used to process and render images and video. Nvidia's GPUs are widely used in professional visualization, data centers, and gaming. The company also provides software and services to enhance the capabilities of its GPUs. Nvidia's innovative technologies have applications in various industries, including automotive, healthcare, and entertainment. The company's stock is publicly traded and can be found on major stock exchanges.

2. 使用 Phi-3-Vision 进行 OCR

除了分析图片，我们还可以从图片中提取信息。这就是 OCR 过程，以前需要写复杂代码才能完成。

python

prompt = f"{user_prompt}<|image_1|>\nHelp me get the title and author information of this book?{prompt_suffix}{assistant_prompt}"

url = "https://marketplace.canva.com/EAFPHUaBrFc/1/0/1003w/canva-black-and-white-modern-alone-story-book-cover-QHBKwQnsgzs.jpg"

image = Image.open(requests.get(url, stream=True).raw)

inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")

generate_ids = model.generate(**inputs, 
                              max_new_tokens=1000,
                              eos_token_id=processor.tokenizer.eos_token_id,
                              )

generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]

response = processor.batch_decode(generate_ids, 
                                  skip_special_tokens=False, 
                                  clean_up_tokenization_spaces=False)[0]

结果如下

txt

The title of the book is "ALONE" and the author is Morgan Maxwell.

3. 多图比较

Phi-3 Vision 支持多张图片的比较。我们可以用这个模型找出图片之间的差异。

python

prompt = f"{user_prompt}<|image_1|>\n<|image_2|>\n What is difference in this two images?{prompt_suffix}{assistant_prompt}"

print(f">>> Prompt\n{prompt}")

url = "https://hinhnen.ibongda.net/upload/wallpaper/doi-bong/2012/11/22/arsenal-wallpaper-free.jpg"

image_1 = Image.open(requests.get(url, stream=True).raw)

url = "https://assets-webp.khelnow.com/d7293de2fa93b29528da214253f1d8d0/news/uploads/2021/07/Arsenal-1024x576.jpg.webp"

image_2 = Image.open(requests.get(url, stream=True).raw)

images = [image_1, image_2]

inputs = processor(prompt, images, return_tensors="pt").to("cuda:0")

generate_ids = model.generate(**inputs, 
                              max_new_tokens=1000,
                              eos_token_id=processor.tokenizer.eos_token_id,
                              )

generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]

response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]

结果如下

txt

The first image shows a group of soccer players from the Arsenal Football Club posing for a team photo with their trophies, while the second image shows a group of soccer players from the Arsenal Football Club celebrating a victory with a large crowd of fans in the background. The difference between the two images is the context in which the photos were taken, with the first image focusing on the team and their trophies, and the second image capturing a moment of celebration and victory.