Gangwei XU

Results 27 comments of Gangwei XU

Have you tested it on stereo matching?

> Hello! > > When I went through the RawNeRF paper, it says that it was trained using multi-exposure dataset. But when I downloaded the RawNeRF dataset, I could only...

can you give me your gray images?

You can use KITTI pretrained model, that will perform well.

请问你是用哪一个.py文件测试的呢?是evaluate_stereo.py还是demo_imgs.py?在这两个文件中有一行disp = model(,),这个disp直接出的就是视差数值

不太清楚你的具体问题,按照现有的sceneflow处理方式就可以了啊

这是不正常的,我们训练的时候没有出现NAN 2023-11-02 14:40:03>1790426751 写道: 你好,我在复现论文时发现训练过程中会出现nan,所以我在代码中添加以下代码(图中画圈的)想要查看跑到第几个epoch时会出现nan。然后,我发现在训练一开始就有参数的梯度是nan,请问这是正常的吗? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: ***@***.***>

我觉得看你的需求,如果你想快一点,可以设置一下2023-11-03 09:54:08>1790426751 写道: 感谢您之前的回复,我还有一个小问题想要请教:就是关于训练参数--slow_fast_gru的设置,我看这个参数在前身RAFT-Stereo中说可以提高计算速度,但是在您的论文并没有提到这点,在代码中这参数也是默认为false的。请问这个参数我要设置为true还是保持默认就好? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: ***@***.***>

batchsize最好设置大一点,太小容易震荡容易出现nan,如果batchsize太小的话,试试梯度累计,将多个batch的loss求平均后再反向传播。或者把 core中的igev_stereo.py with autocast(enabled=self.args.mixed_precision): 改成enabled=False试试,我能想到的了,如果还不行,可能还是需要调整batchsize。2023-11-30 19:50:29>Wade_Lucky 写道: 作者您好,我在sceneflow数据集上训练您的模型时,发现如果将lr设置为0.001,batchsize=2,然后跑到5k步左右的时候,3像素偏差会出现nan,我想咨询一下,如果batchsize设置只能是2的话,如果让网络可以快速下降,lr应该设置为多少呢 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: ***@***.***>