Why doesn't the loss function go down？

[LYJvincent]

Hi！Excuse me！ I use this network to detect defects.Use MixSoftmaxCrossEntropyLoss loss function, training to 400 eopch, but loss function has been drop down between 0.5 to 0.7.May I ask what the problem is?

[LYJvincent]

MixSoftmaxCrossEntropyOHEMLoss and MixSoftmaxCrossEntropyLoss have the same effect, The learning rate is not going down

[feitiandemiaomi]

@LYJvincent I meet the same problem with you, I use my dataset to train it , the loss always shakes.

[Tramac]

How much is the learning rate?

[feitiandemiaomi]

@Tramac I'm wrong ,the lr=0.001 and I use the adam optimizer, I will go on training

[LYJvincent]

@Tramac the lr = 1e-4, now I get better results。thanks！

[feitiandemiaomi]

@LYJvincent What the values of miou can achieve? and how long have you trained?

[LYJvincent]

@feitiandemiaomi current mean_iou is 0.97602%， about 3.5 hour

[LYJvincent]

@feitiandemiaomi is just for my database not cityspaces

[Tramac]

@feitiandemiaomi current mean_iou is 0.97602%， about 3.5 hour

0.97602% or 97.602%?

[LYJvincent]

@Tramac Oh,sorry, is my fault。not 0.97602%, but 97.602%

[feitiandemiaomi]

@LYJvincent Can you share hyperparameters ?

[feitiandemiaomi]

@LYJvincent The paper's miou is 69%, your result is amazing. Hope to learn it ,Thanks

[Tramac]

It depends on the datasets.

[CuttlefishXuan]

@feitiandemiaomi Hi, I meet the same problem as you, the training loss fluctuates between 0.9 and 1.1 Have you solved this problem? Could you get the better result

[Alohomora0130]

@feitiandemiaomi Hi, I meet the same problem as you. Have you solve this problem? If you have solve this problem, can you share me with your learning rate and batch_size? Thanks!

[Alohomora0130]

@CuttlefishXuan Hi! Have you solve this problem? Can you tell me your learning rate and batch_size?T hank you very much!