shocked by the initial val result

[myalos]

Before training, the function val evaluate the initial model on the NYUv2 test set, and the result is
abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |
& 0.323 & 0.448 & 1.002 & 0.365 & 0.520 & 0.783 & 0.905
That shocks me, am i wrong ? why the initial model perform pretty well on the NYUv2 test set ?

[niujinshuchong]

Hi, the depth map is aligned with GT depth during evaluation.

[Cresynia]

Hi, I read your blog and noted that you have run the structdepth before,and I met some problems when I run it. It said 'normD_down = D_down + norm_down RuntimeError: The size of tensor a (378) must match the size of tensor b (281) at non-singleton dimension 2'.I don't know where I made errors.And I'm a beginner,the first time to run,really hope you can give some help.Thanks!

[myalos]

Hi,
rgb = torch.permute(rgb, (0, 2, 3, 1))
rgb_down = self.pdist(rgb[:, 1:, :, :], rgb[:, :-1, :, :])
rgb_right = self.pdist(rgb[:, :, 1:, :], rgb[:, :, :-1, :])
...
aligned_norm = torch.permute(rgb, (0, 2, 3, 1))
norm_down = self.pdist(aligned_norm[:, 1:, :, :], aligned_norm[:, :-1, :, :])
norm_right = self.pdist(aligned_norm[:, :, 1:, :], aligned_norm[:, :, :-1, :])

I remember I changed these two places before, hope this help. By the way, did you run the monodepth2 on nyu2 successfully?

[Cresynia]

Wow!It's you!Thanks a lot!Haven't yet.And I want to do the indoor depth estimation for my undergraduate final year paper,but my programming skill is so poor and I discovered the structdepth,and I did this these days.

…

------------------ 原始邮件 ------------------ 发件人: ***@***.***>; 发送时间: 2023年3月17日(星期五) 中午1:20 收件人: ***@***.***>; 抄送: ***@***.***>; ***@***.***>; 主题: Re: [svip-lab/Indoor-SfMLearner] shocked by the initial val result (Issue #18) Hi, rgb = torch.permute(rgb, (0, 2, 3, 1)) rgb_down = self.pdist(rgb[:, 1:, :, :], rgb[:, :-1, :, :]) rgb_right = self.pdist(rgb[:, :, 1:, :], rgb[:, :, :-1, :]) ... aligned_norm = torch.permute(rgb, (0, 2, 3, 1)) norm_down = self.pdist(aligned_norm[:, 1:, :, :], aligned_norm[:, :-1, :, :]) norm_right = self.pdist(aligned_norm[:, :, 1:, :], aligned_norm[:, :, :-1, :]) I remember I changed these two places before, hope this help. By the way, did you run the monodepth2 on nyu2 successfully? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[Cresynia]

Problems solved!Thanks!I can run it!
And could you give me some advice about it,sometimes I felt so confused because I felt I nearly know nothing and can do nothing and still have a lot to learn.About this project,I learned for a while but I always felt not clear about it,can't get the point.Because I'm weak in all aspects.Could you give some advice,and can I ask you questions if I met problems?
Really thanks for your help! Best wishes!

[myalos]

I understand this feeling of confusion. But I am not expert in depth estimation(maybe beginner plus), and your expectations of my abilitiy make me a little nervous. I have run the sfmlearner indoor and structdepth before, if you met questions, feel free to ask, and i will reply if i see it.

[Cresynia]

OK，sincerely thanks for your help!

…

---Original--- From: ***@***.***> Date: Sat, Mar 18, 2023 10:15 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [svip-lab/Indoor-SfMLearner] shocked by the initial val result(Issue #18) I understand this feeling of confusion. But I am not expert in depth estimation(maybe beginner plus), and your expectations of my abilitiy make me a little nervous. I have run the sfmlearner indoor and structdepth before, if you met questions, feel free to ask, and i will reply if i see it. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[Cresynia]

Hi！When I ran this project,I met some problems,I want to ask you if you ever met it.I'm finding the solutions.

And I'm doing my graduation project based on structdepth,but I met some problems.Due to my limited knowledge,I want to do some work, and I only add attention to the network.The result is not good,instead of increasing,it decreased.And I'm finding the reason.I think maybe structdepth trained on the P2net's model,they have the same network,but if I changed the network,the result will decrease,but the structdepth have higher requirements for the pretrained model ,otherwise the result will not good.I don't know if I think right or it have this reason.So I tried to do it on the P2net and then use structdepth. I don't know if I can do like this.I have no direction.Thanks!

[myalos]

A1:
maybe you should change this
color_aug = transforms.ColorJitter.get_params(self.brightness, self.contrast, self.saturation, self.hue).
to this
color_aug = transforms.ColorJitter(self.brightness, self.contrast, self.saturation, self.hue)
A2:
sorry, i cannot give you advice, since i am not familiar with depth + attention.

[Cresynia]

Thanks！I can run it! Sincerely thanks for your help!
And for the second question,I just want to ask if it may have that reason for the structdepth because of changed network.I just guessed that.And I'm going to have a try.I hope I can graduate smoothly.Thanks!

[Cresynia]

Sorry,I have another question.When I'm running ,it occured:


I have no idea why this happen

[myalos]

i suggest you could set the breakpoint and debug, i cannot find the reason from this information. maybe gradient explode or other reason. i have not met this case before.

[Cresynia]

OK,thanks!

[Cresynia]

Do you know why this happen when I debug.

It's so strange.I haven't run to the val_dataset,but it occurred.It occurred at the beginning.And if I run into a function,it will occur again.Will this influence code execution?

[Cresynia]

I make a new one and the problem disappeared about the NameError,I think maybe I changed something but I didn't notice that.And I'm running to see if that problems will occur again.

[myalos]

ctrl + f search where val_dataset is used, comment it.

[Cresynia]

Thanks! Problems disappeared miraculously. I created a new one,and it didn't occur.But the Nan problems still exist sometimes, and sometimes it can run. I don't know why. Hope it won't occur again.And now I'm going to write the eval_res_for_each_epoch.txt just as the structdepth do,because it's not convenient to see.And I also want to use the tensorboard on the server,but I couldn't open the website it provides,do you have any solutions?Thanks!

[myalos]

A1:This needs you set the breakpoint and debug, or search in the website or ask person who is familiar with this bug.There are many reasons that can cause nan problems, and i have limited experience in deep learning.
A2:launch the tensorboard with --ip = 0.0.0.0

[Cresynia]

Q1：OK,Thanks!
Q2:
It still can't work.
I use the command :
tensorboard --logdir=./ --host=0.0.0.0
and go into the ip:6006
but it still can't work.

[myalos]

the command "tensorboard with --ip = 0.0.0.0" works for me

[Cresynia]

Ok,thanks!I'll find other solutions,this is in no hurry.

…

---Original--- From: ***@***.***> Date: Sat, Apr 29, 2023 17:30 PM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [svip-lab/Indoor-SfMLearner] shocked by the initial val result(Issue #18) the command "tensorboard with --ip = 0.0.0.0" works for me — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>