PyTorch ReduceLROnPlateau调度器异常降LR问题排查

我用PyTorch训练MobileNetV3Large模型，配置了ReduceLROnPlateau学习率调度器，但发现验证精度持续上升的情况下，调度器还是降低了学习率。以下是训练代码片段和前7轮训练日志，帮忙排查原因：

训练代码 #

bench_val_loss = 1000
bench_acc = 0.0
epochs = 15
optimizer = optim.Adam(embeddingNet.parameters(), lr=1e-3) 
loss_optimizer = torch.optim.Adam(loss_fn.parameters(), lr=1e-3)

scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=3, threshold=0.02)

for epoch in range(1, epochs + 1):

    print(f'current lr: {scheduler.get_last_lr()}')
    loss=train(embeddingNet, loss_fn, device, train_dataloader, optimizer, loss_optimizer, epoch)
    val_loss, accuracy =test(train_dataset, val_dataset, embeddingNet, accuracy_calculator, loss_fn, epoch, val_dataloader)
    #val_loss = simpleTest(train_dataset, val_dataset, embeddingNet, accuracy_calculator, loss_fn, epoch, val_dataloader)

    
    torch.save(embeddingNet.state_dict(), 'my/path/mobileNetV3L_ArcFaceLAST.pth')

    if accuracy >= bench_acc:
      bench_val_loss = val_loss
      torch.save(embeddingNet.state_dict(), 'my/path/mobileNetV3L_ArcFaceBEST.pth')

    scheduler.step(accuracy)

    writer.add_scalars('Training vs. Validation Loss',
                       {'Training': loss, 'Validation': val_loss},
                       global_step=epoch+1)

前7轮训练日志 #

Test set accuracy (Precision@1) = 0.17834772304046048
current lr: [0.001]
Epoch 3: Loss = 39.68284225463867
Epoch 3: valLoss = 39.9765007019043
100%|██████████| 962/962 [01:43<00:00,  9.28it/s]
100%|██████████| 370/370 [00:41<00:00,  8.92it/s]
Computing accuracy
Test set accuracy (Precision@1) = 0.31242593533096324
current lr: [0.001]
Epoch 4: Loss = 39.4412841796875
Epoch 4: valLoss = 39.67761562450512
100%|██████████| 962/962 [01:45<00:00,  9.11it/s]
100%|██████████| 370/370 [00:41<00:00,  8.86it/s]
Computing accuracy
Test set accuracy (Precision@1) = 0.3633824276282377
current lr: [0.001]
Epoch 5: Loss = 39.09823989868164
Epoch 5: valLoss = 39.54649614901156
100%|██████████| 962/962 [01:42<00:00,  9.37it/s]
100%|██████████| 370/370 [00:41<00:00,  8.87it/s]
Computing accuracy
Test set accuracy (Precision@1) = 0.44244117149145085
current lr: [0.001]
Epoch 6: Loss = 38.70449447631836
Epoch 6: valLoss = 39.1865906792718
100%|██████████| 962/962 [01:45<00:00,  9.15it/s]
100%|██████████| 370/370 [00:39<00:00,  9.25it/s]
Computing accuracy
Test set accuracy (Precision@1) = 0.5167597765363129
current lr: [0.0001]

问题分析与修复 #

核心错误 #

调度器的配置和传入的监控指标不匹配：

• 你设置了scheduler = ReduceLROnPlateau(..., mode='min')，这个模式表示调度器会监控指标的下降，当指标连续patience轮没有下降（即没有改善）时降低学习率。
• 但你调用scheduler.step(accuracy)时传入的是精度，这是一个需要最大化的指标——精度持续上升对mode='min'的调度器来说，相当于指标一直在"没有改善"（因为它在等指标变小），所以触发了学习率下降逻辑。

两种修复方式 #

1. 调整调度器的mode参数：因为监控的是精度（需要最大化），将mode改为'max'：

scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', factor=0.1, patience=3, threshold=0.02)

2. 切换监控指标为验证损失：如果你想保留mode='min'，就把scheduler.step(accuracy)改成scheduler.step(val_loss)——验证损失是需要最小化的指标，和mode='min'逻辑匹配。

额外注意 #

你的代码中使用了两个独立的Adam优化器（optimizer用于模型参数，loss_optimizer用于损失函数参数），但当前调度器只绑定了optimizer，如果loss_fn包含可训练参数，loss_optimizer的学习率不会被调度，若需要统一调度，需为其也配置对应的学习率调度器。

内容的提问来源于stack exchange，提问作者elbarto