PyTorch实现Softmax回归时前向传播异常问题求助

我尝试用Softmax回归将2D数据点分类为0或1，但代码运行存在问题，不确定哪里出错。我初始化了接收2个输入、输出2个类别的Softmax模型，训练时把包含x、y列的训练数据输入模型，以下是我的代码、数据输出和完整CSV文件：

原代码 #

import csv
import torch
import numpy

data_path = "C:\\Users\\bchen\\Downloads\\2d-2class-dataset.csv"

data_numpy = numpy.loadtxt(data_path, dtype=int, delimiter=',', skiprows=1)

data_torch = torch.from_numpy(data_numpy)

row_count = data_torch.size()[0]
training_count = 16

training_xy = data_torch[:training_count, [0, 1]]
training_labels = data_torch[:training_count, 2]

test_xy = data_torch[:training_count + 1, [0, 1]]
test_labels = data_torch[:training_count + 1, 2]

class Softmax(torch.nn.Module):
    def __init__(self, n_inputs, n_outputs):
        super().__init__()
        self.linear = torch.nn.Linear(n_inputs, n_outputs)

    def forward(self, x):
        pred = self.linear(x)
        return pred

model_softmax = Softmax(2, 2)

optimizer = torch.optim.SGD(model_softmax.parameters(), lr=0.01)

criterion = torch.nn.CrossEntropyLoss

Loss = []
epochs = 100
for epoch in range(epochs):
    optimizer.zero_grad()
    model_softmax.train()
    xy_pred = model_softmax(training_xy) #ISSUE
    loss = criterion(xy_pred, training_labels)
    Loss.append(loss)
    loss.backward()
    optimizer.step()

print(Loss)

with torch.inference_mode:
    y_pred = model_softmax(test_xy)

print(test_labels)
print(y_pred)

training_xy输出 #

tensor([[ -85,   40],
        [ -18,   82],
        [-154,  150],
        [ 140,  162],
        [ 102,   64],
        [  22,   28],
        [ 126, -108],
        [ 171,   52],
        [ 276,   43],
        [  77,  -63],
        [ 251, -196],
        [  52, -201],
        [ 181,  -30],
        [  32,  -22],
        [ -34, -129],
        [ -77,  -60]], dtype=torch.int32)

完整CSV文件 #

# x,y,label
-85,40,1
-18,82,1
-154,150,1
140,162,0
102,64,0
22,28,0
126,-108,0
171,52,0
276,43,0
77,-63,0
251,-196,0
52,-201,0
181,-30,0
32,-22,1
-34,-129,1
-77,-60,1
-127,-207,1
-244,-63,1
-189,60,1
-48,198,1
73,214,0
217,221,0
-72,-248,0

问题排查与修正 #

你的代码存在以下几个关键问题：

1. 数据类型不匹配
   training_xy是int32类型，而PyTorch线性层默认使用float32进行计算，类型不兼容会导致报错。需要将输入数据转换为浮点型：

training_xy = data_torch[:training_count, [0, 1]].float()
test_xy = data_torch[training_count:, [0, 1]].float()

2. 损失函数未实例化
   criterion = torch.nn.CrossEntropyLoss只是引用了类，没有创建实例，应该加上括号：

criterion = torch.nn.CrossEntropyLoss()

3. 测试集划分错误
   原代码里测试集取了前17行，和训练集（前16行）大量重叠，应该取训练集之后的数据：

test_xy = data_torch[training_count:, [0, 1]]
test_labels = data_torch[training_count:, 2]

4. inference_mode使用错误
   torch.inference_mode是上下文管理器，需要加括号调用：

with torch.inference_mode():
    y_pred = model_softmax(test_xy)

5. 损失值存储问题
   直接存储张量会占用不必要的内存，应该存储数值：

Loss.append(loss.item())

修正后完整代码 #

import torch
import numpy

data_path = "C:\\Users\\bchen\\Downloads\\2d-2class-dataset.csv"

# 加载数据并转换为PyTorch张量
data_numpy = numpy.loadtxt(data_path, dtype=int, delimiter=',', skiprows=1)
data_torch = torch.from_numpy(data_numpy)

row_count = data_torch.size()[0]
training_count = 16

# 划分训练集和测试集，转换为浮点型
training_xy = data_torch[:training_count, [0, 1]].float()
training_labels = data_torch[:training_count, 2]
test_xy = data_torch[training_count:, [0, 1]].float()
test_labels = data_torch[training_count:, 2]

# 定义Softmax模型（CrossEntropyLoss已包含Softmax，无需额外实现）
class Softmax(torch.nn.Module):
    def __init__(self, n_inputs, n_outputs):
        super().__init__()
        self.linear = torch.nn.Linear(n_inputs, n_outputs)

    def forward(self, x):
        return self.linear(x)

model_softmax = Softmax(2, 2)
optimizer = torch.optim.SGD(model_softmax.parameters(), lr=0.01)
criterion = torch.nn.CrossEntropyLoss()  # 实例化损失函数

Loss = []
epochs = 100
for epoch in range(epochs):
    model_softmax.train()
    optimizer.zero_grad()
    
    xy_pred = model_softmax(training_xy)
    loss = criterion(xy_pred, training_labels)
    Loss.append(loss.item())  # 存储损失数值
    
    loss.backward()
    optimizer.step()

print("训练损失变化：", Loss)

# 测试模型
with torch.inference_mode():
    y_pred = model_softmax(test_xy)
    # 转换为预测类别
    pred_labels = torch.argmax(y_pred, dim=1)

print("测试集真实标签：", test_labels)
print("测试集预测类别：", pred_labels)

内容的提问来源于stack exchange，提问作者codingcultivator445