神经网络多损失函数联合方式

RockLiu@805

于 2024-09-24 23:05:01 发布

阅读量413

点赞数 7

文章标签：神经网络深度学习 python

本文链接：https://blog.csdn.net/qq_18943707/article/details/142502481

版权

多损失函数联合

loss_fn_1用于优化网络权重，loss_fn_2用于优化中心矢量

optim_1 = torch.optim.AdamW(weights.parameters(), lr=lr, weight_decay=1e-4)
optim_2 = torch.optim.RMSprop(weights.parameters(), lr=5e-3)
loss_fn_1 = nn.CrossEntropyLoss().to(device)
loss_fn_2 = CenterLoss(6, 6, size_average=True).to(device)
loss_rate = 5e-3
epochs = 15
index_train = 1
index_test = 1
acc = 0.0
start_time = time.time()
metrics_list = []
end_loss = 0

for epoch in range(epochs):
    print(f"开始第{epoch}轮训练")
    weights.train()
    for data in train_transforms_loader:
        img, label = data
        img = img.to(device)
        label = label.to(device)
        output = weights(img).to(device)
        optim_1.zero_grad()
        optim_2.zero_grad()
        loss_1 = loss_fn_1(output, label)
        loss_2 = loss_fn_2(output, label)
        # writer.add_scalar(' train_loss', loss.item(), index_train)
        result_loss = loss_1 + loss_rate * loss_2
        result_loss.backward(retain_graph=True)
        loss_2 *= (1. / loss_rate)
        loss_2.backward()
        optim_1.step()

        optim_2.step()
        if index_train % 10 == 0:
            print(f'index_train{index_train}')
        index_train += 1

如果多个损失项均用于优化网络权重，那么只采用一个优化器即可

optim_1 = torch.optim.AdamW(weights.parameters(), lr=lr, weight_decay=1e-4)
loss_fn_1 = nn.CrossEntropyLoss().to(device)
loss_fn_2 = CenterLoss(6, 6, size_average=True).to(device)
loss_rate = 5e-3
epochs = 15
index_train = 1
index_test = 1
acc = 0.0
start_time = time.time()
metrics_list = []
end_loss = 0

for epoch in range(epochs):
    print(f"开始第{epoch}轮训练")
    weights.train()
    for data in train_transforms_loader:
        img, label = data
        img = img.to(device)
        label = label.to(device)
        output = weights(img).to(device)
        optim_1.zero_grad()
        loss_1 = loss_fn_1(output, label)
        loss_2 = loss_fn_2(output, label)
        # writer.add_scalar(' train_loss', loss.item(), index_train)
        result_loss = loss_1 + loss_rate * loss_2
        result_loss.backward()
        optim_1.step()
        if index_train % 10 == 0:
            print(f'index_train{index_train}')
        index_train += 1