我正在尝试使用 Logistic 回归使用 pytorch 执行 ML 任务，以解决具有 17 个数字属性作为输入的二元分类问题。
这是我的模型：
class LogisticRegression(nn.Module):
def __init__(self):
super(LogisticRegression, self).__init__()
self.linear = nn.Linear(17, 1)
def forward(self, x):
output = torch.sigmoid(self.linear(x))
return output
纪元代码：
self.loss_func = torch.nn.BCELoss()
self.net = LogisticRegression()
self.opti = optim.Adam(self.net.parameters(), lr=0.01)
for epoch in range(local_epochs):
for data, label in self.train_dl:
data, label = data.to(self.dev), label.to(self.dev)
self.opti.zero_grad()
preds = self.net(data)
preds = torch.round(preds)
loss = self.loss_func(preds, label.to(torch.float32))
loss.backward()
self.opti.step()
计算精度的代码：
sum_accu = 0
num = 0
for data, label in self.test_dl:
data, label = data.to(self.dev), label.to(self.dev)
preds = self.net(data)
print(preds)
rounded_preds = torch.round(preds)
sum_accu += (rounded_preds == label).float().mean()
num += 1
accuracy = sum_accu/num
每个时代的准确度总是相同的，所以当通过在每个时代结束时打印“preds”进行调查时，我观察到这些值完全相同，一遍又一遍地打印。这是从我的终端复制的两个这样的实例。
tensor([[0.3958],
[0.3945],
[0.7899],
[0.7307],
[0.4597],
[0.6161],
[0.4561],
[0.5829],
[0.5005],
[0.5272],
[0.5041],
[0.5438],
[0.5004],
[0.3623],
[0.5250],
[0.5050],
[0.7511],
[0.5194],
[0.4359],
[0.4961],
[0.4741],
[0.2358],
[0.5108],
[0.6026],
[0.5581],
[0.7312],
[0.5047],
[0.5456],
[0.3654],
[0.3890],
[0.5150],
[0.4835],
[0.5635],
[0.6935],
[0.4463],
[0.5423],
[0.6504],
[0.5503],
[0.5756],
[0.6759],
[0.5986],
[0.4919],
[0.4506],
[0.5991],
[0.6428],
[0.5743],
[0.4644],
[0.5459],
[0.4973],
[0.5363],
[0.4168],
[0.4297],
[0.4899],
[0.4357],
[0.4491],
[0.7026],
[0.5691],
[0.3235],
[0.5053],
[0.6034],
[0.5739],
[0.3975],
[0.5293],
[0.4870],
[0.4976],
[0.4651],
[0.7204],
[0.5242],
[0.3649],
[0.5264],
[0.4867],
[0.3726],
[0.5398],
[0.5339],
[0.5334],
[0.3884],
[0.4933],
[0.4489],
[0.3119],
[0.3826],
[0.4244],
[0.5612],
[0.5480],
[0.5575],
[0.5411],
[0.6343],
[0.5174],
[0.4965],
[0.5172],
[0.4460],
[0.5153],
[0.3980],
[0.5763],
[0.4840],
[0.4682],
[0.5400],
[0.7034],
[0.4811],
[0.4481]])
tensor([[0.3958],
[0.3945],
[0.7899],
[0.7307],
[0.4597],
[0.6161],
[0.4561],
[0.5829],
[0.5005],
[0.5272],
[0.5041],
[0.5438],
[0.5004],
[0.3623],
[0.5250],
[0.5050],
[0.7511],
[0.5194],
[0.4359],
[0.4961],
[0.4741],
[0.2358],
[0.5108],
[0.6026],
[0.5581],
[0.7312],
[0.5047],
[0.5456],
[0.3654],
[0.3890],
[0.5150],
[0.4835],
[0.5635],
[0.6935],
[0.4463],
[0.5423],
[0.6504],
[0.5503],
[0.5756],
[0.6759],
[0.5986],
[0.4919],
[0.4506],
[0.5991],
[0.6428],
[0.5743],
[0.4644],
[0.5459],
[0.4973],
[0.5363],
[0.4168],
[0.4297],
[0.4899],
[0.4357],
[0.4491],
[0.7026],
[0.5691],
[0.3235],
[0.5053],
[0.6034],
[0.5739],
[0.3975],
[0.5293],
[0.4870],
[0.4976],
[0.4651],
[0.7204],
[0.5242],
[0.3649],
[0.5264],
[0.4867],
[0.3726],
[0.5398],
[0.5339],
[0.5334],
[0.3884],
[0.4933],
[0.4489],
[0.3119],
[0.3826],
[0.4244],
[0.5612],
[0.5480],
[0.5575],
[0.5411],
[0.6343],
[0.5174],
[0.4965],
[0.5172],
[0.4460],
[0.5153],
[0.3980],
[0.5763],
[0.4840],
[0.4682],
[0.5400],
[0.7034],
[0.4811],
[0.4481]])
我原以为随着模型学习和更新参数，每个时期的输出都会发生变化，但是每个时期相同的输出导致相同的准确度，这清楚地表明模型没有学习。尝试使用从 0.1 到 0.00001 的学习率。