Error downloading `jpeg` & `pillow` with conda about download-celeba-hq HOT 14 OPEN

Thanks for reporting this.

@vgunjal you are not doing something wrong. I was able to reproduce the error... The script needs to be updated. The problem is that I believe that these versions are necessary for the script to work correctly.

I am trying now the script with new versions, but I think it will not work.

from download-celeba-hq.

So I've been doing a lot of experimenting to get the script to work over the past 2 days. A lot of it are hacky workarounds, but if they end up giving me the output I need, I can post an update / make a PR with the changes if that's okay with you?

from download-celeba-hq.

That would be awsome... Thanks a lot.

from download-celeba-hq.

Did you check that the images are correctly reconstructed? You should have a look at a 100 of them to see if anything weird is happening...

from download-celeba-hq.

The process is yet to complete one successful run xD I've been babysitting it all day. Will post an update by tomorrow.

from download-celeba-hq.

@nperraud The script completed successfully! I'm now trying to verify if the images are correct. I used numpy and matplotlib to look at the npy files as images and I get examples that look like these -

Is the output supposed to look like this? What do you use to visualize the images?

from download-celeba-hq.

They are not supposed to look like this. But I suppose that it might just be a plotting issue. Have a look at this notebook.
https://github.com/nperraud/download-celebA-HQ/blob/master/Checking%20the%20dataset.ipynb

from download-celeba-hq.

It was indeed a plotting issue. The output looks similar to the output from the notebook.

Note that some of the images have random green / blue pixels scattered around in them. I'm not sure why those occur.

from download-celeba-hq.

I think it should look like that...

I believe the wrong pixels comes from the fact that you are using the wrong versions of the packages...

from download-celeba-hq.

do you think the output I get is usable for the training of the Progressive GANs?

from download-celeba-hq.

No, unfortunatly I do not think so...

But did you try the original repo? https://github.com/tkarras/progressive_growing_of_gans
I have the feeling that their code improved a lot since I created these scripts. Today, it might be easier to just use their repo... or to do a mix of both repo...

I see now that their requirements has been updated:
https://github.com/tkarras/progressive_growing_of_gans/blob/master/requirements-pip.txt

Maybe simply updating the requirements should do the trick

from download-celeba-hq.

I am sorry, I do not have the time yet to help more...

from download-celeba-hq.

I already tried with that repo - however as I understand it, their dataset_tool.py file only helps generate the deltas. I didn't understand how they recreate the original images from the deltas and, hence, I ended up using your script.

from download-celeba-hq.

Could you still make a pull request? I will try to have a look at it...

from download-celeba-hq.