sample_llama3-8B.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
import os
import fire
import json
import re
from collections import defaultdict
from datasets import load_dataset
from typing import Optional, List
from llama import Llama
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch



def extract_svg_from_text(text: str) -> Optional[str]:
    """
    从包含SVG的文本中提取出完整的<svg>...</svg>结构。
    如果未匹配到，则返回一个默认的空SVG。
    """
    pattern = r"<svg\b[^>]*>.*?</svg>"
    matches = re.findall(pattern, text, re.DOTALL | re.IGNORECASE)
    if matches:
        return matches[0]
    else:
        return """<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 36 36"></svg>"""


def code_style_prompt(desc: str) -> str:
    return f"""\
// SVG CODE GENERATION TASK FOR CODELLAMA
// OBJECTIVE: Create simple yet accurate SVG contour drawing
// DESCRIPTION: {desc}

// SVG Example(DESCRIPTION=wheelchair)(you do not need to generate an example as well):
<svg xmlns="http://www.w3.org/2000/svg" width="100" height="100" viewBox="0 0 100 100">
  <!-- Wheelchair seat -->
  <path d="M30,40 L50,40 L50,60 L30,60 Z" fill="#555"/>

  <!-- Wheelchair back -->
  <path d="M30,40 L20,30 L20,20 L30,20 L30,40" fill="#555"/>

  <!-- Large wheel -->
  <circle cx="65" cy="65" r="25" stroke="#333" stroke-width="3" fill="none"/>
  <circle cx="65" cy="65" r="5" fill="#333"/>

  <!-- Small wheel -->
  <circle cx="30" cy="70" r="10" stroke="#333" stroke-width="3" fill="none"/>
  <circle cx="30" cy="70" r="3" fill="#333"/>

  <!-- Wheel spokes (large wheel) -->
  <line x1="65" y1="65" x2="80" y2="65" stroke="#333" stroke-width="2"/>
  <line x1="65" y1="65" x2="65" y2="80" stroke="#333" stroke-width="2"/>
  <line x1="65" y1="65" x2="55" y2="75" stroke="#333" stroke-width="2"/>
  <line x1="65" y1="65" x2="55" y2="55" stroke="#333" stroke-width="2"/>

  <!-- Wheel spokes (small wheel) -->
  <line x1="30" y1="70" x2="38" y2="70" stroke="#333" stroke-width="2"/>
  <line x1="30" y1="70" x2="30" y2="78" stroke="#333" stroke-width="2"/>
</svg>


// CODE GENERATION INSTRUCTIONS:
1. Figure out the main parts of the object(animal) according to the DESCRIPTION
1. Fill path data for main-outline using basic commands
2. Position eye element at logical position
3. Keep all coordinates within viewBox
4. Use 2 decimal precision for coordinates
5. Close all path elements properly

// {desc} GENERATION START FROM HERE:
"""


def post_process(code: str) -> str:
    """针对代码模型的输出优化后处理"""
    # 提取闭合的SVG代码块
    svg_match = re.search(r'<svg.*?</svg>', code, re.DOTALL)
    if svg_match:
        code = svg_match.group(0)

    # 确保XML声明
    if '<?xml' not in code:
        code = '<?xml version="1.0" encoding="UTF-8"?>\n' + code

    # 验证必要元素
    required_elements = {
        '<svg': 1,
        '</svg>': 1,
        '<path': 1,
        '<circle': 1
    }
    for elem, count in required_elements.items():
        if code.count(elem) < count:
            code = code.replace('</svg>',
                                f'<!-- Auto-added {elem} -->\n<{elem} />\n</svg>')

    return code.strip()


def strict_svg_postprocess(raw_code: str) -> str:
    """
    严格按照需求设计的SVG后处理器

    处理逻辑：
    1. 按行处理，找到第一个不以<svg开头的行作为内容起点
    2. 逐行检查：去重（最多3次）、完整性、排除</svg>
    3. 自动添加标准头尾
    """
    # 预处理：清理前后空白，分割为行
    lines = [line.strip() for line in raw_code.strip().split('\n')]

    # 阶段1：找到有效内容起始行
    start_index = 0
    for i, line in enumerate(lines):
        if not line.lower().startswith("<svg"):
            start_index = i
            break

    # 阶段2：逐行处理有效内容
    valid_lines = []
    line_counter = {}

    for line in lines[start_index:]:
        # 排除</svg>标签
        if re.match(r'</\s*svg\s*>', line, re.IGNORECASE):
            continue
        if re.match(r'<\s*svg', line, re.IGNORECASE):
            continue
        # 检查完整性（匹配XML标签语法）
        is_valid_tag = re.fullmatch(
            r'\s*<[^>]+/?>\s*',
            line,
            re.IGNORECASE
        )

        # 检查是否已存在3次
        count = line_counter.get(line, 0)

        if is_valid_tag and count < 1:
            valid_lines.append(line)
            line_counter[line] = count + 1

    # 阶段3：组装最终结果
    core_content = '\n'.join(valid_lines)

    return f'''<svg xmlns="http://www.w3.org/2000/svg" width="100" height="100" viewBox="0 0 100 100">
{core_content}
</svg>'''


def load_label_names(json_path: str) -> dict:
    """加载标签映射表"""
    with open(json_path, 'r', encoding='utf-8') as f:
        data = json.load(f)
    return data['dataset_info']['features'][0]['dtype']['class_label']['names']


def main_infer( ):
    # 初始化代码模型
    # generator = Llama.build(
    #     ckpt_dir=ckpt_dir,
    #     tokenizer_path=tokenizer_path,
    #     max_seq_len=max_seq_len,
    #     max_batch_size=max_batch_size,
    # )
    # 加载基础模型（请根据具体模型名称或路径调整）
    base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct",token="")
    # 加载 LoRA 模型，加载 LoRA 权重（此处使用“steve329/llama3-8B-edit-lora-12k”）
    model = PeftModel.from_pretrained(base_model, "steve329/llama3-8B-edit-lora-12k")
    # 设置评估模式
    model.eval()
    # 加载对应的分词器（确保与基础模型匹配）
    tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(device)


    root = "/export/home2/zhanjun001/codellama/codellama/SVGEditBench_clean_llama8b"
    for dir in os.listdir(root):
        print(dir)
        if (dir == "LICENSE-CODE") or (dir == "LICENSE-IMAGES") or (dir == "README.md") or (dir == "CaseGenerator.py") : continue
        output_dir = os.path.jon(root+dir+'generated_svg') # 替换为实际的目标文件夹路径

        # 确保目标文件夹存在，如果不存在则创建
        if not os.path.exists(output_dir):
            os.makedirs(output_dir)

        file_dir = os.path.jon(root+dir+'query')
        i=0
        for file in os.listdir(file_dir):
            
            file_name = os.path.splitext(file)[0]
            file_path = os.path.join(file_dir, file)
            with open(file_path, "r", encoding="utf-8") as file:
                content = file.read()
            print(content)
            if len(content) > 4383:
                file_path = os.path.join(output_dir, file_name + '.svg')

                # 将final_code写入到description.svg文件中
                with open(file_path, 'w', encoding='utf-8') as svg_file:
                    svg_file.write("""<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 36 36"></svg>""")

                print(f"SVG文件已保存至: {file_path}")
                with open('/export/home2/zhanjun001/codellama/codellama/SVGEditBench_clean/'+dir+'/skipped_file.txt', 'w', encoding='utf-8') as f:
                    f.write(f"{file_name}"+"\n")

                continue

            # test_input = (
            #     '{"instruction": "You are an expert SVG graphics generator. You generate clean, valid SVG code according to user instructions.", '
            #     f'"input": {content}'
            # )

            inputs = tokenizer(content, return_tensors="pt")
            input_ids = inputs.input_ids.to(device)
            attention_mask = inputs.attention_mask.to(device)
            # 使用模型生成文本（可以根据需要调整生成参数）
            with torch.no_grad():
                generated_ids = model.generate(
                    input_ids,
                    attention_mask=attention_mask,
                    max_length=4096,  # 指定生成文本的最大长度
                    do_sample=True,  # 是否使用采样，True 可生成更多样化结果
                    top_k=50,  # Top-K 采样参数
                    top_p=0.95  # Top-p (nucleus) 采样参数
                )

            # 解码生成的 token 成为文本
            generated_text = tokenizer.decode(generated_ids[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
            # print("Prompt:", prompt)
            # print("Generated text:", generated_text)
            # print("-" * 80)
            # results = generator.text_completion(
            #     prompts=[content],
            #     max_gen_len=max_gen_len,
            #     temperature=temperature,
            #     top_p=top_p,
            # )





            # 后处理
            # raw_code = results[0]['generation']aa
            print("raw_code:")
            print(generated_text)
            final_code = extract_svg_from_text(generated_text)

            # 输出结果
            print(f"\n=== Input: {file_name} ===")
            print(f"// Generated SVG Code:")
            print(final_code)
            print("\n" + "=" * 40 + "\n")

            # 定义SVG文件的完整路径
            file_path = os.path.join(output_dir, file_name + '.svg')

            # 将final_code写入到description.svg文件中
            with open(file_path, 'w', encoding='utf-8') as svg_file:
                svg_file.write(final_code)

            print(f"SVG文件已保存至: {file_path}")


if __name__ == "__main__":
    main_infer()