During the training process, the loss has been rising

[amnesiac555]

i use the mobilenet and fpn_inception, training data is Gopro, config is as follows:
anyone has trained successfully？

[amnesiac555]

i didn't change training code, feel confused about it

[NoviceCrom]

Maybe you can change the "disc_loss" in the config.yml to ragan_ls which is stated in the paper. By the way, could you share your Gopro dataset download from the link with me? I have no access to it for some reason.

[amnesiac555]

Maybe you can change the "disc_loss" in the config.yml to ragan_ls which is stated in the paper.

OK THANKS ！ i will try it immediately

[ldlshizhu]

请问你训练成功了吗

[ldlshizhu]

你好，我看到你train_batches_per_epoch: 1000
val_batches_per_epoch: 100这两行注释掉了，但是我下载的代码里没有注释这两行。请问是什么原因呢

[zmwang0574]

你好，我看到你train_batches_per_epoch: 1000
val_batches_per_epoch: 100这两行注释掉了，但是我下载的代码里没有注释这两行。请问是什么原因呢

Hi, I'm also wondering why you # these two lines. Are they unnecessary for the training? BTW, I'm also wondering what does "bounds" do, which is also #-ed by you.

[cianhwang]

你好，我看到你train_batches_per_epoch: 1000
val_batches_per_epoch: 100这两行注释掉了，但是我下载的代码里没有注释这两行。请问是什么原因呢

Hi, I'm also wondering why you # these two lines. Are they unnecessary for the training? BTW, I'm also wondering what does "bounds" do, which is also #-ed by you.

Commenting train/val_batches_per_epoch ensures each epoch the network sees all the training/validation samples. "bounds" s are intended to split the train/val set, which should all be set to [0.0, 1.0] if you specify separate training/validation sets.

[KMCzajkowski]

Have you solved this issue? I also have a similar problem. The loss is not improving during training regardless of what I do.