yuxiangsun / rtfnet Goto Github PK

Dear Dr. Yuxiang Sun,
I would like to implement this very meaningful project into a machine that runs only with CPU.
Could you please share your take on this issue, regarding feasibility of implementing it using CUDA toolkit without gpu or converting gpu arrays into cpu
arrays andd its effect on computation time?
Thank you in advance!

您好，尝试训练您的模型的时候，发现代码中没有提供完整的参数。尝试从github上您提供的链接下载非常慢。能否提供一下google 云盘的链接或者百度网盘？万分感谢！

大佬您好，我最近在做多模态语义分割方向，我们有一个实验需要用到您FuseSeg网络的代码进行复现，请问这个网络的代码可以开源一下吗？

大佬您好，我们最近在做多模态语义分割方面的工作，需要您的帮助，我们要对FuseSeg论文进行官方的结果复现，您的团队可以考虑开源一下FuseSeg的代码吗？

Hello, Mr. Sun.

Thanks for your job. I would like to know some details about your two work of RGB-T Semantic Segmentaiton.

About the params and FPS. I have a question about your RTFNet and FuseSeg. According to TABLE IX of your FuseSeg, FuseSeg-161 has 1.5 times params fewer than RTFNet-101, but why FuseSeg-161 slower than RTFNet-101? Is there more FLOPs in FuseSeg-161 than RTFNet-101?

Thanks for your work again.

another question...
i downloaded the ir_seg_dataset but the pics show rgb type in a software and ir in another software. and the label folder pics r all almost black. is it normal? if it isnt. would u mind share the dataset with me. thanks QAQ

Excuse me，I have an issue，Why is the dataset four-channel？ How to train RGB image and infrared image separately input？I would be most grateful if you could answer me.

Hi,

Thank you very much for your nice work!

I have a question about the result of NTE and NRE shown in Table IV of the paper.

When you test NTE and NRE, did you retrain the model? Or simple remove RGB encoder or thermal encoder from RTFNet model?

I have tried to remove the RGB encoder to have NTE, and remove the thermal encoder to obtain NRE. But in my case, the macc of NTE at daytime is only 27.3, and the macc of NTE at nighttime is only 16.1. The macc of NRE at daytime is only 11.1, and the macc of NRE at nighttime is only 11.1.

Therefore, I am wondering how you guys produced the results in Table IV. Could you please give some explanations?

Bests,
Xingchen

I was not able to download the dataset.
The site seems problem.
Could you validate the server of the dataset?

first of all, thank you for sharing you code, i'm beginner in python and tensorflow and i don't know how to show the qualitative results ( predicted images) of testing, i just got quantitative values (probability of each class).

请问您训练这个网络花了多久时间？最终训练了多少个epoch？每100个需要多久？
我这里每训练一个epoch需要花费1个小时左右（batch size = 2 num_workers= 0），耗时太长让我很困惑，想知道这个时间是否是正常范围？

First of all, thank you for your code. However, I downloaded your software and tried to train again, but i got wrong result, and i don’t why it happened.
My test result is that: Miou: 51.4899%(resnet152) 48.20%(resnet101) 40.5% (resnet34) 45%(resnet18).
I have trained those model for about 140 to 180 epoch. I think that it is enough. So do you have any idea about this problem? Is it caused by dropout? Thank you very much.

您好 最近研读了您的论文 我注意到单独用白天跟夜晚的数据得到的结果（macc miou）比总体的都低， 理论上 总体的值不应该介于两者之间吗？ 这一点让我非常困惑，希望您能给一些指导。。。

Thanks for your work, but why the results in Table 3 of your paper is so different from the results in the paper of MFNet?

When training RTFNet not use Docker, I run: python train.py in my conda,
but GPU(RTX2080) is out of memory even set batch_size = 1 and num_workers = 1?

#############################################################################################
parser = argparse.ArgumentParser(description='Train with pytorch')
#############################################################################################
parser.add_argument('--model_name', '-m', type=str, default='RTFNet')
#batch_size: RTFNet-152: 2; RTFNet-101: 2; RTFNet-50: 3; RTFNet-34: 10; RTFNet-18: 15;
parser.add_argument('--batch_size', '-b', type=int, default=2)
parser.add_argument('--lr_start', '-ls', type=float, default=0.01)
parser.add_argument('--gpu', '-g', type=int, default=0)
#############################################################################################
parser.add_argument('--lr_decay', '-ld', type=float, default=0.95)
parser.add_argument('--epoch_max', '-em', type=int, default=1000) # please stop training mannully
parser.add_argument('--epoch_from', '-ef', type=int, default=0)
parser.add_argument('--num_workers', '-j', type=int, default=1)
parser.add_argument('--n_class', '-nc', type=int, default=9)
parser.add_argument('--data_dir', '-dr', type=str, default='./dataset/')
args = parser.parse_args()
#############################################################################################

Thanks for your great work! I am wondering whether the algorithm can perform in real-time, and what's the approximate processing frequency during test?

我运行了您的train.py，使用ResNet-152作为编码器，batch_size设置为2，其他参数均维持您代码中的原值，目前位置已经训练了88个epoch，但是loss一直在不停的震荡不下降，请问您知道怎么解决这个问题吗？

I've been able to reproduce the results in your paper using the pretrained weights (RTFNet 152).
Since you didn't provide a training script, I've used my own. However I wasn't able to get anywhere close to the same mAcc and mIoU.

Do you plan to release your training script? Could you perhaps detail your procedure?

I have tried the same parameters detailed in your paper, and freezing the ResNET weights.