PyTorch使用字母字符串作为Tensor输入报too many dimensions 'str'如何解决

PyTorch 中的torch.Tensor是数值张量类型，原生不支持直接存储字符串，你直接将字符串列表传入torch.Tensor构造函数，就会触发类型不匹配的报错。
神经网络只能接收数值型输入，所以你需要先把字母字符串转换为数值编码，再输入模型，下面是两种可直接落地的实现方案：

方案1：固定字符映射（轻量场景适用） #

先建立字符到数字的映射字典，将每个字母转换为对应的整数，再构造张量即可运行。
修改后的完整代码如下：

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable as var
import torch.optim as optim

# 新增：字符到数值的映射表
char_to_idx = {char: idx for idx, char in enumerate(['a','b','c','d','e','f','g','h','i','j'])}
# 反向映射可选，需要还原字符的时候用
idx_to_char = {idx: char for char, idx in char_to_idx.items()}

class Nnetz(nn.Module):
    def __init__(self):
        super(Nnetz, self).__init__()
        self.lin1 = nn.Linear(10, 10)
        self.lin2 = nn.Linear(10, 10)
        self.lin3 = nn.Linear(10, 10)
        self.lin4 = nn.Linear(10, 10)
        self.lin5 = nn.Linear(10, 10)
        self.lin6 = nn.Linear(10, 10)
        
    def forward(self, x):
        x = F.relu(self.lin1(x))
        x = self.lin2(x)
        x = self.lin3(x)
        x = self.lin4(x)
        x = self.lin5(x)
        x = self.lin6(x)
        return x
    
    def num_flat_features(self, x):
        size = x.size()[1:]
        num = 1
        for i in size:
            num *= i
        return num
        
netz = Nnetz()

for i in range(100):
    # 替换原来的数值输入为字母输入，转成数值列表
    x_char = ['j','i','h','g','f','e','d','c','b','a']
    x = [char_to_idx[c] for c in x_char]
    input = var(torch.Tensor([x for _ in range(10)]))
    out = netz(input)

    x_target_char = ['a','b','c','d','e','f','g','h','i','j']
    target = var(torch.Tensor([[char_to_idx[c] for c in x_target_char] for _ in range(10)]))
    criterion = nn.MSELoss()
    loss = criterion(out, target)
    print(loss)

    netz.zero_grad()
    loss.backward()
    optimizer = optim.SGD(netz.parameters(), lr=0.01)
    optimizer.step()

该方案的优势是改动量极小，无需调整网络结构；劣势是映射的数值是序数，神经网络会默认认为数值有大小关系（比如a对应0 < b对应1），如果你的字符本身没有顺序语义，会引入不必要的偏差。

方案2：Embedding层编码（工业界常用方案） #

如果要消除数值序数的偏差，获得更好的训练效果，可以在网络最前面增加nn.Embedding层，将字符索引映射为可学习的向量，是文本输入场景的标准处理方式。
修改后的完整代码如下：

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable as var
import torch.optim as optim

# 字符到数值的映射表
char_to_idx = {char: idx for idx, char in enumerate(['a','b','c','d','e','f','g','h','i','j'])}
vocab_size = len(char_to_idx) # 字典大小10
embedding_dim = 1 # 这里可以调整嵌入向量维度，比如设为2、4都可以

class Nnetz(nn.Module):
    def __init__(self):
        super(Nnetz, self).__init__()
        # 新增Embedding层
        self.embedding = nn.Embedding(vocab_size, embedding_dim)
        # 输入维度调整为 10 * embedding_dim，因为10个字符每个输出embedding_dim维，展开后输入全连接层
        self.lin1 = nn.Linear(10 * embedding_dim, 10)
        self.lin2 = nn.Linear(10, 10)
        self.lin3 = nn.Linear(10, 10)
        self.lin4 = nn.Linear(10, 10)
        self.lin5 = nn.Linear(10, 10)
        self.lin6 = nn.Linear(10, 10)
        
    def forward(self, x):
        # x的形状是(batch_size, 10)，每个元素是字符索引
        x = self.embedding(x) # 输出形状(batch_size, 10, embedding_dim)
        x = x.flatten(1) # 展平为(batch_size, 10 * embedding_dim)
        x = F.relu(self.lin1(x))
        x = self.lin2(x)
        x = self.lin3(x)
        x = self.lin4(x)
        x = self.lin5(x)
        x = self.lin6(x)
        return x
    
    def num_flat_features(self, x):
        size = x.size()[1:]
        num = 1
        for i in size:
            num *= i
        return num
        
netz = Nnetz()

for i in range(100):
    x_char = ['j','i','h','g','f','e','d','c','b','a']
    x = [char_to_idx[c] for c in x_char]
    # 注意这里转LongTensor，因为Embedding层要求输入是整数索引
    input = var(torch.LongTensor([x for _ in range(10)]))
    out = netz(input)

    x_target_char = ['a','b','c','d','e','f','g','h','i','j']
    target = var(torch.Tensor([[char_to_idx[c] for c in x_target_char] for _ in range(10)]))
    criterion = nn.MSELoss()
    loss = criterion(out, target)
    print(loss)

    netz.zero_grad()
    loss.backward()
    optimizer = optim.SGD(netz.parameters(), lr=0.01)
    optimizer.step()

内容的提问来源于stack exchange，提问作者FruchtigeBanane