NaughtyDog97/FormalEnhencedGPS-34B

基于FormalGeo7K的推理模型

快速开始

在运行脚本之前，首先安装如下必要的依赖。

pip install --upgrade pip
pip install torch transformers==4.40.0
pip install sentencepiece protobuf
pip install accelerate pillow
pip install ninja
pip install packaging
pip install flash-attn --no-build-isolation

import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image
import warnings
import numpy as np

# set device
device = 'cuda'  # or cpu
torch.set_default_device(device)

# create model
model = AutoModelForCausalLM.from_pretrained(
    'NaughtyDog97/FormalEnhencedGPS-34B',
    torch_dtype=torch.float16, # float32 for cpu
    device_map='auto',
    trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(
    'NaughtyDog97/FormalEnhencedGPS-34B',
    trust_remote_code=True)

# text prompt
img_path = 'sample/4927.png'
qs = 'As shown in the diagram, AE/AB=1/4, M is the midpoint of segment AC, BE is parallel to CP, EA is parallel to CP. Find the ratio of the length of line BC to the length of line CD.'
prompt = f'Using the provided geometric image and question, first predict the construction_cdl and image_cdl. Then, give a detailed step-by-step solution.\nThe question is:\n{qs}'
text = f'<|im_start|>user\n<image>\n{prompt}<|im_end|>\n<|im_start|>assistant\n'
text_chunks = [tokenizer(chunk).input_ids for chunk in text.split('<image>')]
input_ids = torch.tensor(text_chunks[0] + [-200] + text_chunks[1][1:], dtype=torch.long).unsqueeze(0).to(device)

# image, sample images can be found in images folder
image = Image.open(img_path).convert('RGB')

image_tensor = model.process_images([image], model.config).to(dtype=model.dtype, device=device)

# generate
with torch.inference_mode():
    output_ids = model.generate(
        input_ids,
        images=image_tensor,
        do_sample=False,
        temperature=None,
        top_p=None,
        top_k=None,
        num_beams=1,
        max_new_tokens=3500,
        eos_token_id=tokenizer.eos_token_id,
        repetition_penalty=None,
        use_cache=True
    )[0]


respones = tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip()
print(respones)

我们的模型支持的求解方式有如下三种:

# Q => Predicted CDL + CoT Answer
prompt = f'Using the provided geometric image and question, first predict the construction_cdl and image_cdl. Then, give a detailed step-by-step solution.\nThe question is:\n{qs}'


# Q + Predicted CDL => CoT Answer
prompt = f'Using the provided geometric image, construction_cdl, image_cdl, and question, give a detailed step-by-step solution. Note that there may be minor errors in the construction_cdl and image_cdl.\nThe construction_cdl is:\n{predict_consCDL}\nThe image_cdl is:\n{predict_imgCDL}\nThe question is:\n{qs}'


# Q + Predicted CDL => Calibrated CDL + CoT Answer
prompt = f'Using the provided geometric image and the possibly erroneous construction_cdl and image_cdl, first calibrate the construction_cdl and image_cdl, then give a detailed step-by-step solution to the question.\nThe initial construction_cdl is:\n{predict_consCDL}\nThe initial image_cdl is:\n{predict_imgCDL}\nThe question is:\n{qs}'

结合Formalization模型的推理

import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image
import warnings
import numpy as np
import re


def parse_cdl(input_string):
    # 使用正则表达式查找各个部分
    patterns = {
        'construction_cdl': r'(?:The )?(?:calibrate )?construction_cdl(?: is)?:\n(.*?)(?=\n(?:The )?(?:calibrate )?\w+_cdl is:|\n(?:The )?(?:calibrate )?\w+_cdl:|\nSolution is:|\Z)',
        'image_cdl': r'(?:The )?(?:calibrate )?image_cdl(?: is)?:\n(.*?)(?=\n(?:The )?(?:calibrate )?\w+_cdl is:|\n(?:The )?(?:calibrate )?\w+_cdl:|\nSolution is:|\Z)',
        'text_cdl': r'(?:The )?text_cdl(?: is)?:\n(.*?)(?=\n(?:The )?\w+_cdl is:|\n(?:The )?\w+_cdl:|\nSolution is:|\Z)',
        'goal_cdl': r'(?:The )?goal_cdl(?: is)?:\n(.*?)(?=\n(?:The )?\w+_cdl is:|\n(?:The )?\w+_cdl:|\nSolution is:|\Z)'
    }
    
    results = {}
    
    # 优先匹配包含"calibrate"的版本
    for key, pattern in patterns.items():
        pattern = pattern.replace("(?:calibrate )?", "(?:calibrate )")
        match = re.search(pattern, input_string, re.DOTALL)
        if match:
            results[key] = match.group(1).strip()
        else:
            # 如果未找到包含"calibrate"的版本，尝试匹配不含"calibrate"的版本
            pattern = pattern.replace("(?:calibrate )", "(?:calibrate )?")
            match = re.search(pattern, input_string, re.DOTALL)
            if match:
                results[key] = match.group(1).strip()
    
    return results


# set device
device = 'cuda'  # or cpu
torch.set_default_device(device)

# create model
formalization_model = AutoModelForCausalLM.from_pretrained(
    'NaughtyDog97/GeoFormalizer',
    torch_dtype=torch.float16, # float32 for cpu
    device_map='auto',
    trust_remote_code=True)

formalization_tokenizer = AutoTokenizer.from_pretrained(
    'NaughtyDog97/GeoFormalizer',
    trust_remote_code=True)


reason_model = AutoModelForCausalLM.from_pretrained(
    'NaughtyDog97/FormalEnhencedGPS-34B',
    torch_dtype=torch.float16, # float32 for cpu
    device_map='auto',
    trust_remote_code=True)
reason_tokenizer = AutoTokenizer.from_pretrained(
    'NaughtyDog97/FormalEnhencedGPS-34B',
    trust_remote_code=True)



img_path = 'sample/4927.png'
image = Image.open(img_path).convert('RGB')


# formalization
prompt = 'Based on the image, first describe what you see in the figure, then predict the construction_cdl and image_cdl and calibrate it.'
text = f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<image>\n{prompt}<|im_end|>\n<|im_start|>assistant\n"
text_chunks = [formalization_tokenizer(chunk).input_ids for chunk in text.split('<image>')]
input_ids = torch.tensor(text_chunks[0] + [-200] + text_chunks[1][1:], dtype=torch.long).unsqueeze(0).to(device)

# generate
image_tensor = formalization_model.process_images([image], formalization_model.config).to(dtype=formalization_model.dtype, device=device)
with torch.inference_mode():
    output_ids = formalization_model.generate(
        input_ids,
        images=image_tensor,
        do_sample=False,
        temperature=None,
        top_p=None,
        top_k=None,
        num_beams=1,
        max_new_tokens=3500,
        eos_token_id=formalization_tokenizer.eos_token_id,
        repetition_penalty=None,
        use_cache=True
    )[0]


respones = formalization_tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip()
print(f'Formalization result is\n{respones}')
cdl_info = parse_cdl(respones)
predict_consCDL = cdl_info['construction_cdl']
predict_imgCDL = cdl_info['image_cdl']



# reasoning

qs = 'As shown in the diagram, AE/AB=1/4, M is the midpoint of segment AC, BE is parallel to CP, EA is parallel to CP. Find the ratio of the length of line BC to the length of line CD.'
prompt = f'Using the provided geometric image and the possibly erroneous construction_cdl and image_cdl, first calibrate the construction_cdl and image_cdl, then give a detailed step-by-step solution to the question.\nThe initial construction_cdl is:\n{predict_consCDL}\nThe initial image_cdl is:\n{predict_imgCDL}\nThe question is:\n{qs}'
text = f'<|im_start|>user\n<image>\n{prompt}<|im_end|>\n<|im_start|>assistant\n'
text_chunks = [reason_tokenizer(chunk).input_ids for chunk in text.split('<image>')]
input_ids = torch.tensor(text_chunks[0] + [-200] + text_chunks[1][1:], dtype=torch.long).unsqueeze(0).to(device)



# generate
image_tensor = reason_model.process_images([image], reason_model.config).to(dtype=reason_model.dtype, device=device)
with torch.inference_mode():
    output_ids = reason_model.generate(
        input_ids,
        images=image_tensor,
        do_sample=False,
        temperature=None,
        top_p=None,
        top_k=None,
        num_beams=1,
        max_new_tokens=3500,
        eos_token_id=reason_tokenizer.eos_token_id,
        repetition_penalty=None,
        use_cache=True
    )[0]

respones = reason_tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip()
print(f'Reasoning steps is\n{respones}')

Performance

	Q => Predicted CDL + CoT Answer	Q + Predicted CDL => CoT Answer	Q + Predicted CDL => Calibrated CDL + CoT Answer
siglip-0.4B-yi1.5-34B	71.84/80.58	72.17/81.72	72.33/81.72