pytorch FANISHION-MNIST训练了1280次loss降不下去了，一直在1.73左右，模型最后一步用的softmax_卷积神经网络训练mnist数据集损失值一直在2.3左右怎么回事-优快云博客

本文链接：https://blog.youkuaiyun.com/studyvcmfc/article/details/106612631

本文详细介绍了使用Python和PyTorch库训练神经网络的过程，包括数据加载、网络前向传播、损失计算、反向传播及权重更新等关键步骤。通过实际代码展示了如何调整网络参数以最小化损失函数。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

import collections
# TODO: Train the network here
print("initial fc1.weights=",model.fc1.weight)
epoches = 1280
for i in range(epoches):
images ,labels = next(iter(trainloader))
#print("images.size()=",images.size()[0])
#print("labels=",labels)
t= images.size()[0]
#print("t=",t)
images.resize_(t,784)
optimizer.zero_grad()
output = model.forward(images)
criterion = nn.CrossEntropyLoss()
loss = criterion(output,labels)
loss.backward()
# print("gradient=",model.fc1.weight.grad)
print("loss=",loss)
optimizer.step()

#各步打印的loss

loss= tensor(1.8425, grad_fn=<NllLossBackward>)
loss= tensor(1.7099, grad_fn=<NllLossBackward>)
loss= tensor(1.6928, grad_fn=<NllLossBackward>)
loss= tensor(1.8132, grad_fn=<NllLossBackward>)
loss= tensor(1.7459, grad_fn=<NllLossBackward>)
loss= tensor(1.7712, grad_fn=<NllLossBackward>)
loss= tensor(1.7174, grad_fn=<NllLossBackward>)
loss= tensor(1.8287, grad_fn=<NllLossBackward>)
loss= tensor(1.8221, grad_fn=<NllLossBackward>)
loss= tensor(1.7079, grad_fn=<NllLossBackward>)
loss= tensor(1.8615, grad_fn=<NllLossBackward>)
loss= tensor(1.8535, grad_fn=<NllLossBackward>)
loss= tensor(1.7789, grad_fn=<NllLossBackward>)
loss= tensor(1.8826, grad_fn=<NllLossBackward>)
loss= tensor(1.7923, grad_fn=<NllLossBackward>)
loss= tensor(1.7446, grad_fn=<NllLossBackward>)
loss= tensor(1.7387, grad_fn=<NllLossBackward>)
loss= tensor(1.8063, grad_fn=<NllLossBackward>)
loss= tensor(1.7541, grad_fn=<NllLossBackward>)
loss= tensor(1.7205, grad_fn=<NllLossBackward>)
loss= tensor(1.7760, grad_fn=<NllLossBackward>)
loss= tensor(1.7013, grad_fn=<NllLossBackward>)
loss= tensor(1.8267, grad_fn=<NllLossBackward>)
loss= tensor(1.6684, grad_fn=<NllLossBackward>)
loss= tensor(1.7829, grad_fn=<NllLossBackward>)
loss= tensor(1.7570, grad_fn=<NllLossBackward>)
loss= tensor(1.7603, grad_fn=<NllLossBackward>)
loss= tensor(1.6776, grad_fn=<NllLossBackward>)
loss= tensor(1.7989, grad_fn=<NllLossBackward>)
loss= tensor(1.7488, grad_fn=<NllLossBackward>)
loss= tensor(1.7861, grad_fn=<NllLossBackward>)
loss= tensor(1.7122, grad_fn=<NllLossBackward>)
loss= tensor(1.7824, grad_fn=<NllLossBackward>)
loss= tensor(1.7683, grad_fn=<NllLossBackward>)
loss= tensor(1.7521, grad_fn=<NllLossBackward>)
loss= tensor(1.7753, grad_fn=<NllLossBackward>)
loss= tensor(1.7420, grad_fn=<NllLossBackward>)
loss= tensor(1.7792, grad_fn=<NllLossBackward>)
loss= tensor(1.8161, grad_fn=<NllLossBackward>)
loss= tensor(1.7306, grad_fn=<NllLossBackward>)
loss= tensor(1.7359, grad_fn=<NllLossBackward>)
loss= tensor(1.6829, grad_fn=<NllLossBackward>)
loss= tensor(1.7521, grad_fn=<NllLossBackward>)
loss= tensor(1.8347, grad_fn=<NllLossBackward>)
loss= tensor(1.7310, grad_fn=<NllLossBackward>)
loss= tensor(1.7629, grad_fn=<NllLossBackward>)
loss= tensor(1.7304, grad_fn=<NllLossBackward>)