Open phoenixkillerli opened 4 years ago
The released model was trained using bash scripts/base.sh EXP_NAME
.
What differences are you finding in performance?
I used MIRFLICKR-25000 image collection to train model. and then test encode and decode some images. All are failed. But use the released model, all are succeed. I will try again and find problem. Thanks.
I used MIRFLICKR-25000 image collection to train model. and then test encode and decode some images. All are failed. But use the released model, all are succeed. I will try again and find problem. Thanks.
Hello! Have you solved this problem? I have the same problem
Hello Guys, have anybody solver this problem? training with original parameters does not work at all in my case.
Hello Guys, have anybody solver this problem? training with original parameters does not work at all in my case.
No, I use the code and parameters published by the author, but the accuracy of the training can not meet the requirements of the author. I don't know why.
fujiaixin09
Hello Guys, have anybody solver this problem? training with original parameters does not work at all in my case.
No, I use the code and parameters published by the author, but the accuracy of the training can not meet the requirements of the author. I don't know why.
Hello fujiaixin09, Thank you for your answer! But when you said that accuracy couldn't meet. Is it mean that sometimes it works (model can decode)? In my case it is just nothing, it looks like CNN was not trained at all.
No, it can't decode. The bit accuracy of my training model can only reach more than 80.
------------------ 原始邮件 ------------------ 发件人: "tancik/StegaStamp" @.>; 发送时间: 2021年5月23日(星期天) 晚上6:51 @.>; @.**@.>; 主题: Re: [tancik/StegaStamp] could you share training parameters? (#17)
fujiaixin09
Hello Guys, have anybody solver this problem? training with original parameters does not work at all in my case.
No, I use the code and parameters published by the author, but the accuracy of the training can not meet the requirements of the author. I don't know why.
Hello fujiaixin09, Thank you for your answer! But when you said that accuracy couldn't meet. Is it mean that sometimes it works (model can decode)? In my case it is just nothing, it looks like CNN was not trained at all.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.
The released model was trained using
bash scripts/base.sh EXP_NAME
. What differences are you finding in performance?
Hello Tancik, Thank you for your repository it is really cool! Can you suggest us anything about training, it looks nobody can train model that will be able to decode watermarked image. But at the same time I was able to run your model for decoding and it works fine.
Can you confirm the following,
No modification of the code were makes
The parameters in scripts/base.sh
are being used
The MIRFLICKR data is being used for training
tf1.3 is being used
Additionally, how many iterations do you train for and how long does this training take (in hours). Can you post a screenshot of the tensorboard after training.
Can you confirm the following,
No modification of the code were makes The parameters in
scripts/base.sh
are being used The MIRFLICKR data is being used for training tf1.3 is being usedAdditionally, how many iterations do you train for and how long does this training take (in hours). Can you post a screenshot of the tensorboard after training.
Hello @tancik, Thank you for fast answer,
_ No modification of the code were makes - no modification just git clone
_ The parameters in scripts/base.sh are being used - yes, I have run as described in https://github.com/tancik/StegaStamp/blob/master/README.md
_ The MIRFLICKR data is being used for training - yes MIRFLICKR it has 25000 images
_ tf1.3 is being used - if I use tf1.3 - it gives error: cannot import 'keras'. And after fixing this, new errors appeared. I think original code has been written on the other tf - if I use tf1.13 as it described in README it works but after training it is not recognised anything.
how many iterations do you train - 140 000 as should be
how long does this training take (in hours). - I think +- 12 hours on gtx 1080 ti
screenshot in attachment.
Kind regards
Hmm, interesting. Can you also provide a screenshot of the remaining scalar logs and screenshot of the image logs.
Yes sure.
It is odd that the input images are black. It doesn't seem to be using the dataset images.
hmm. yes, I see.
hello, I ran into the same problem。 could you tell me how to solve it,pls? @EvgenGunya
hmm. yes, I see.
hello, I ran into the same problem。 could you tell me how to solve it,pls?
My problem was path to the image folder.
My problem was path to the image folder.
thx I modified the variable TRAIN_PATH of train.py, but useless Do I need to modify other variables? @EvgenGunya
I try to train model, bug could not get the same performance like the released model. Could you share training parameters or told me how to adjust parameters.
Thanks.