PSNR: Nan

[YuhangLiu98]

I've run it on multiple other datasets and this is happening and I'm very confused.

[Mytttttttt]

这是来自QQ邮箱的假期自动回复邮件。你好，我最近正在休假中，无法亲自回复你的邮件。我将在假期结束后，尽快给你回复。

[HuaqingHe]

I'm encountering a similar issue in my experiments and would like to respectfully seek your guidance. My current configuration is as follows:

Dataset: DIV2K + FFHQ
train:
ema_decay: 0
optim_g:
type: Adam
lr: 4e-4
weight_decay: 0
betas: [0.9, 0.99]

scheduler:
type: CosineAnnealingRestartLR
periods: [200000]
restart_weights: [1]
eta_min: 1e-7
pretrain_network_g: experiments/pretrained_models/Restormer_real_denoising.pth
Notably, I'm observing the [specific issue you're referring to] as early as Step 2000. Could you kindly advise on potential reasons for this early occurrence and any recommended adjustments to the configuration?
Also, I have changed the scheduler to
type: MultiStepLR
milestones: [100000, 750000]
#milestones: [5000, 10000, 15000, 20000，25000，30000，35000]
gamma: 0.5
The problem is still.

[HuaqingHe]

I've run it on multiple other datasets and this is happening and I'm very confused.

I find that use_grad_clip is helpful. Did you turn it on?

[YuhangLiu98]

I've run it on multiple other datasets and this is happening and I'm very confused.

I find that use_grad_clip is helpful. Did you turn it on?

在yml文件中，use_grad_clip: true。需要将它设置为False吗？

[HuaqingHe]

可以看看 model 的optimize_parameters里面这个参数有没有被用到。是要打开的。 `if self.opt['train'].get('use_grad_clip', False): torch.nn.utils.clip_grad_norm_(self.net_g.parameters(), 0.01)` 打开之后会限制每次梯度更新不超过0.01。如果还是有问题，可以试试把learning rate调小。 我试着可以把 0.01 调大点。不然每次更细的梯度太小，模型参数变化不大，收敛很慢。

…

---Original--- From: "Yuhang ***@***.***> Date: Tue, Apr 22, 2025 17:07 PM To: ***@***.***>; Cc: "Huaqing ***@***.******@***.***>; Subject: Re: [swz30/Restormer] PSNR: Nan (Issue #121) image.png (view on web) I've run it on multiple other datasets and this is happening and I'm very confused. I find that use_grad_clip is helpful. Did you turn it on? 在yml文件中，use_grad_clip: true。需要将它设置为False吗？ — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***> YuhangLiu98 left a comment (swz30/Restormer#121) image.png (view on web) I've run it on multiple other datasets and this is happening and I'm very confused. I find that use_grad_clip is helpful. Did you turn it on? 在yml文件中，use_grad_clip: true。需要将它设置为False吗？ — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>