somethng about the dataset

[haha1958]

hello, i want to run your code ,but i got a wrong result . i guess it was due to the dataset .Is subfolder meas "profile " in your code ,and the target meas "frontal"?
And ,how could you make the dataset like 001 001.jpg , i can only make it 001 002 .... 001.jpg
will it influence the result?

[opetrova]

Hi,
Yes, each subfolder contains one or more profile images of a person, and the target for them is a frontal image of the same person.

And ,how could you make the dataset like 001 001.jpg , i can only make it 001 002 .... 001.jpg
will it influence the result?

Hmmm I am not sure what you mean.

[haha1958]

i guess i should read the code more carefully . i just want to know whether the sequence of files could influence the result or not？

…

------------------ 原始邮件 ------------------ 发件人: "OlgaP"<[email protected]>; 发送时间: 2019年5月22日(星期三) 晚上7:35 收件人: "scaleway/frontalization"<[email protected]>; 抄送: "杜慧敏"<[email protected]>;"Author"<[email protected]>; 主题: Re: [scaleway/frontalization] somethng about the dataset (#2) Hi, Yes, each subfolder contains one or more profile images of a person, and the target for them is a frontal image of the same person. And ,how could you make the dataset like 001 001.jpg , i can only make it 001 002 .... 001.jpg will it influence the result? Hmmm I am not sure what you mean. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[opetrova]

i guess i should read the code more carefully . i just want to know whether the sequence of files could influence the result or not？
…

Still not sure what you mean by sequence of files. Do you mean the alphanumerical order of file names and subfolder names?

[haha1958]

actually, i can not understand what does this mean: but if we are to separately sort all the subfolders and all the targets alphanumerically, the ones corresponding to the same subject must appear at the same position on the two lists of names. so i do not know what should i do to get a correct dataset ,this is my dataset and this is my result I change nothing in your code ,but the result is wrong . so i just guess it result from dataset. i will try read code once more ,it should not be a big problem .

…

------------------ 原始邮件 ------------------ 发件人: "OlgaP"<[email protected]>; 发送时间: 2019年5月22日(星期三) 晚上7:53 收件人: "scaleway/frontalization"<[email protected]>; 抄送: "杜慧敏"<[email protected]>; "Author"<[email protected]>; 主题: Re: [scaleway/frontalization] somethng about the dataset (#2) i guess i should read the code more carefully . i just want to know whether the sequence of files could influence the result or not？ … Still not sure what you mean by sequence of files. Do you mean the alphanumerical order of file names and subfolder names? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[opetrova]

@haha1958 I think you may have tried to attach some images (?) to your response, but they did not go through.

What I mean is the following: in your dataset folder, you have images and subfolders (which also contain images). Lets say your images are named 001.jpg, 002.jpg and 137.jpg. Your subfolders are named 001, 137 and 49. (I am only using digits inside these names for simplicity, could be letters as well.) Then the correspondence between frontal images and subfolders with profile images will be the following:
001.jpg - 001
002.jpg - 137
137.jpg - 49

Does that clear it up?

[haha1958]

Thank you very much , I get it !! first image and first subfolder are same people , second image and second subfolder are same people....Is it right? and i show you my result :

[opetrova]

Thank you very much , I get it !! first image and first subfolder are same people , second image and second subfolder are same people....Is it right?

Exactly :)

and i show you my result

Hmm, this is very strange. Looks like either the images and/or subfolders don't get sorted for some reason, or their ordering ends up being different than what was used on my machine. If you are using the same convention as mine for your training dataset ("first image and first subfolder are same people , second image and second subfolder are same people" etc), I would recommend going inside the ExternalInputIterator class in data.py to make sure that self.frontals, self.frontal_indices and self.profiles contain the filenames and the indices that they should (the meanings of these should be clear enough from the code, if not, don't hesitate to ask questions!)

A quick check: what happens if you turn off the shuffling for the training set? Do inputs and "reals" correspond to the same people?

[LaviLiu]

@haha1958 Could you tell me the pytorch version you using in the project?

[haha1958]

@haha1958 Could you tell me the pytorch version you using in the project?

0.4.1

[mohitgupta07]

Just a comment, seems like your model doesn't work well for unseen data. Or maybe because of distribution of faces(w.r.t. countries and regions) it can severly impact the performance.

[hangon666]

@haha1958 @Laviyy
Hello, did you re-train the network or test the given pre-trained model?
My pytorch version is 0.4.1, but I can't load the given model generator_v0.pt

[haha1958]

@hangon666 I find a picture named test.jpg in my computer , which generated by load re-trained model 'netG_49.pt' . I am not sure which pytorch version i was use, maybe still 0.4.1. but I try load it again use pytorch 1.1.0 today , it caused runtimeerror.

[hangon666]

@haha1958 ，thanks for your reply, if convinient, could you share the re-trained model?

[haha1958]

@hangon666 It hadnot get an excellent result, but you can use it just for test https://pan.baidu.com/s/1OKnZKbL8M5uGhj1qe77CyQ

[evelynyhc]

hello, thanks you for your work. I am working on this project recently, and i want to know if there any other dataset can instand of CMU dataset? Another question is the re-trained model you shared before have been repired, can you please share it again?

[redaghanem]

I get this error

--> 108 self.decode = ops.ImageDecoder(device = 'mixed', output_type = types.RGB)

AttributeError: module 'nvidia.dali.ops' has no attribute 'nvJPEGDecoder'

any help? please

[redaghanem]

please i need to run this code @haha1958