Bias Image Inpainting to paint water

[gcosne]

Exploring inpainting methods such as : Generative Image Inpainting to paint water.

[gcosne]

To do :

1. Images must be constant size
2. Adapt code to use custom masks

It seems that masks are randomly generated and some modifications of the data loader are required to use custom masks : 275.

Author suggests having a look at neuralgym.data_from_fnames. cf neuralgym

[gcosne]

Actual version of the code I'm modifying is the V2. The first version of the paper & code was designed for mask of rectangular shape while v2 works well with any mask.

Current state: - Training with random mask and monitoring tensorboard on slurm is working.

But I have several issues while loading custom mask with the neuralgym dataloader
error : 'data_from_fnames.py:153: image is None' .

Suggestion:

• Building a custom dataloader for image-mask pairs.
• The difficulty lies in not breaking the data-augmentation procedure.

[sashavor]

Did you see this issue? JiahuiYu/generative_inpainting#48 and #15 too

…

On Thu, Jun 27, 2019 at 11:29 AM Gautier Cosne ***@***.***> wrote: Actual version of the code I'm modifying is the V2 <https://github.com/Yukariin/generative_inpainting>. The first version of the paper & code was designed for mask of rectangular shape while v2 works well with any mask. [image: image] <https://user-images.githubusercontent.com/25148268/60278775-066cf680-98ce-11e9-9e3e-e45c65bff560.png> *Current state:* - Training with random mask and monitoring tensorboard on slurm is working. But I have several issues while loading custom mask with the neuralgym dataloader error : *'data_from_fnames.py:153: image is None'* . *Suggestion:* - Building a custom dataloader for image-mask pairs. - The difficulty lies in not breaking the data-augmentation procedure. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#77?email_source=notifications&email_token=ADMMIIVVVCXJL4MLTQ4L3DDP4TMGJA5CNFSM4HZAH7I2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODYXPSSQ#issuecomment-506394954>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADMMIIVIBS5DFMST2FUOU73P4TMGJANCNFSM4HZAH7IQ> .

-- Sasha Luccioni Postdoctoral Researcher Climate Change AI https://mila.quebec/en/person/sasha-luccioni/

[sashavor]

Also this: JiahuiYu/generative_inpainting#9 On Thu, Jun 27, 2019 at 1:01 PM Sasha Luccioni <luccionis@mila.quebec> wrote:

…

Did you see this issue? JiahuiYu/generative_inpainting#48 and #15 too On Thu, Jun 27, 2019 at 11:29 AM Gautier Cosne ***@***.***> wrote: > Actual version of the code I'm modifying is the V2 > <https://github.com/Yukariin/generative_inpainting>. The first version > of the paper & code was designed for mask of rectangular shape while v2 > works well with any mask. > [image: image] > <https://user-images.githubusercontent.com/25148268/60278775-066cf680-98ce-11e9-9e3e-e45c65bff560.png> > > *Current state:* - Training with random mask and monitoring tensorboard > on slurm is working. > > But I have several issues while loading custom mask with the neuralgym > dataloader > error : *'data_from_fnames.py:153: image is None'* . > > *Suggestion:* > > - Building a custom dataloader for image-mask pairs. > - The difficulty lies in not breaking the data-augmentation procedure. > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#77?email_source=notifications&email_token=ADMMIIVVVCXJL4MLTQ4L3DDP4TMGJA5CNFSM4HZAH7I2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODYXPSSQ#issuecomment-506394954>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/ADMMIIVIBS5DFMST2FUOU73P4TMGJANCNFSM4HZAH7IQ> > . > -- Sasha Luccioni Postdoctoral Researcher Climate Change AI https://mila.quebec/en/person/sasha-luccioni/

-- Sasha Luccioni Postdoctoral Researcher Climate Change AI https://mila.quebec/en/person/sasha-luccioni/

[gcosne]

I shamefully had an issue with the path & number of channels, I still have some issues with data augmentation here but it works if I disable it.

First epochs..

[sashavor]

It's not shameful! It's part of research!

…

On Thu, Jun 27, 2019, 16:10 Gautier Cosne ***@***.***> wrote: I shamefully had an issue with the path & number of channels, I still have some issues with data augmentation here but it works if I disable it. First epochs.. [image: image] <https://user-images.githubusercontent.com/25148268/60297284-d76a7b00-98f5-11e9-8ef3-0cb7209d7425.png> [image: image] <https://user-images.githubusercontent.com/25148268/60297339-f49f4980-98f5-11e9-92b1-4e3b7f01d762.png> — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#77?email_source=notifications&email_token=ADMMIIQPF2KZC467MF7OXP3P4UNCDA5CNFSM4HZAH7I2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODYYHLJI#issuecomment-506492325>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADMMIIUCMG4HY2II7KVFKATP4UNCDANCNFSM4HZAH7IQ> .

[gcosne]

Link to the code: https://github.com/cc-ai/generative_inpainting

GitHub

cc-ai/generative_inpainting

Generative Image Inpainting with Contextual Attention https://arxiv.org/abs/1801.07892, demo http://jiahuiyu.com/deepfill - cc-ai/generative_inpainting