# AI的NLP学习/embedding/tensorboard

龙文浩_2026-04-09 9:53

import torch

import torch.nn as nn

from torch.utils.tensorboard import SummaryWriter

import jieba

import os

import shutil

===================== 加载自定义词表，确保人名完整分词 =====================

jieba_tokenizer = jieba.Tokenizer()

custom_words = ["龙文浩", "张三", "李四", "王五", "工程师", "学生", "教师", "医生"]

for word in custom_words:

jieba_tokenizer.add_word(word)

构建语料 + 词表

def build_vocab():

texts = [

"龙文浩是工程师",

"张三是学生",

"李四是教师",

"王五是医生"

]

words = []

for text in texts:

words += jieba_tokenizer.lcut(text)

vocab = sorted(list(set(words)))

word2idx = {w: i for i, w in enumerate(vocab)}

print("词表：", word2idx)

return vocab, word2idx

===================== 构建简单 Embedding 模型 =====================

class EmbeddingModel(nn.Module):

def init (self, vocab_size, embed_dim=10):

super().init ()

self.embedding = nn.Embedding(vocab_size, embed_dim)

def forward(self, x):

return self.embedding(x)

===================== 主函数：训练 + TensorBoard 可视化 =====================

if name == 'main ':

vocab, word2idx = build_vocab()

vocab_size = len(vocab)

复制代码

model = EmbeddingModel(vocab_size, embed_dim=10)
word_indices = torch.tensor(list(word2idx.values()), dtype=torch.long)

with torch.no_grad():
    embeddings = model.embedding(word_indices)

# ===================== 设置根目录 =====================
# 统一使用绝对路径，避免相对路径歧义
log_dir = os.path.abspath(r"runs\embedding_visual")
# 强制清空旧日志，避免缓存/编码问题
if os.path.exists(log_dir):
    shutil.rmtree(log_dir)
os.makedirs(log_dir, exist_ok=True)

# ===================== 写入TensorBoard（简化目录，消除嵌套） =====================
writer = SummaryWriter(log_dir)
# 移除tag参数，避免生成嵌套目录；不指定global_step，简化目录结构
writer.add_embedding(
    mat=embeddings,
    metadata=vocab
)
writer.close()

print("\n✅ TensorBoard 文件已保存到", log_dir)
print("🚀 启动命令（复制直接执行，PowerShell/CMD通用）：")
print(f"tensorboard --logdir={log_dir} --host=127.0.0.1 --port=6006")
print("🌐 浏览器访问：http://127.0.0.1:6006，点击 PROJECTOR 标签查看词向量可视化")