What changes did you make in the skimage code

[myalos]

Thanks for sharing code!
What changes did you make in the skimage code，can you release the changing code and explain why ?

[Cresynia]

Hello!I met a problem and want to ask if you have ever met it before.I made some changes to the network, and at first it can run successfully,but after several epochs,the problem occur:

[myalos]

no, pred_depth is generated by _, depth = disp_to_depth(disp, self.opt.min_depth, self.opt.max_depth)
maybe you could check this and then continue to check according to the function call relationship.

[Cresynia]

OK,thanks!I'll try it.And I want to ask when I haven't make changes to the network,this problem didn't occur, and after I made changes,it occur,can it explain the problem is in the network?

[myalos]

yes

[Cresynia]

OK!

[Cresynia]

I think maybe I found the reasons, it's the GPU memory problem,I can continue trainig on the previous model.And I want to ask is it normal that the accuracy decrease sharply for the first several epochs after I add network to the original network?

[myalos]

i think yes, since it is normal that adding a not compatible network at first.

[Cresynia]

Thanks！It becomes better now!😃

…

---Original--- From: ***@***.***> Date: Wed, May 3, 2023 07:59 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [SJTU-ViSYS/StructDepth] What changes did you make in the skimagecode (Issue #8) i think yes, since it is normal that adding a not compatible network at first. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[Cresynia]

Hello!I want to ask you some questions about structdepth.I want to use it to be my baseline(without the losses),and I don't know if the result of it is the DSO one in this picture?

[Cresynia]

And when I run structdepth with the pretrained model on the Imagenet, the problems occurred:

it didn't change!So strange.At first,I used the P2net's model as the structdepth do as pretrained model, the result is normal,because I want to use the one without structdepth's losses as my baseline,I want to change the pretrained model(use the Imagenet one).Have you ever meet this problems?

[myalos]

no, i haven't. i think that maybe training structdepth needs a good depth network at first. That means, if you change the pretrained model, the depth network will be far worse than p2net.

[Cresynia]

Yes,it needs the good network when using the norm loss and the planar loss,but if I don't use the two losses at first,and change the network,training a good network and use losses later,is that ok?I do this,and the problem occurred.Thanks for your reply! Original Email Sender:"goon"< ***@***.*** &gt;; Sent Time:2023/5/8 21:33 To:"SJTU-ViSYS/StructDepth"< ***@***.*** &gt;; Cc recipient:"Cresynia"< ***@***.*** &gt;;"Comment"< ***@***.*** &gt;; Subject:Re: [SJTU-ViSYS/StructDepth] What changes did you make in the skimagecode (Issue #8) no, i haven't. i think that training structdepth needs a good depth network at first. If you change the pretrained model, the depth network is far worse than p2net. &mdash; Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***&gt;

[myalos]

I didn't try it. I think it could work.

[Cresynia]

Ok,thanks!I want to ask the p2net without its planar loss is the same as the struct depth without its Manhattan norm loss and planar loss?

…

---Original--- From: ***@***.***&gt; Date: Tue, May 9, 2023 18:11 PM To: ***@***.***&gt;; Cc: ***@***.******@***.***&gt;; Subject: Re: [SJTU-ViSYS/StructDepth] What changes did you make in the skimagecode (Issue #8) I didn't try it. I think it could work. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***&gt;

[myalos]

Sorry for late, if "same" refer to accuracy, maybe not the same, since batch size, training set, pretrained model are different. If we don't take these into account, i think they are the same, both use same network structure and training method, both sparse rgb points.