Great Work!!!!!

[apple]

Great Work!!!!!

[Xiaoming-Zhao]

Thanks a lot for your interests in this work. We may not be able to provide the processed data trying to avoid any potential LICENSE issue.

exceed off quota limit is a common issue as GDrive does not support the command line downloading well (I encountered the same thing). You can use the browser to download the data. I know it is a little bit cumbersome but this may be the only way (I also use this tedious way to download the data).

Hope this helps.

[fangchangma]

@VIROBO-15 We will check with legal and see if we can release the pre-processed datasets. Please stay tuned.

[3dsf]

@VIROBO-15

Hi, Can you Please provide the FFHQ data as, I am unable to download the ffhq dataset from the given google drive (https://drive.google.com/drive/folders/1LTBpJ0W_WLjqza3zdayligS8Dh1V1gA6) due to exceed off quota limit.

In the past for issues like this, I've cloned the drive to my own google drive and then you avoid the data caps.
(Also I just downloaded it outside of my google account with no issues)

[VIROBO-15]

@3dsf
Thank You for helping but I have also made a copy of the drive in my own google drive but It still shows the same issue.

Can you Please help me out in this...

[cnnlstm]

Thanks a lot for your interests in this work. We may not be able to provide the processed data trying to avoid any potential LICENSE issue.

exceed off quota limit is a common issue as GDrive does not support the command line downloading well (I encountered the same thing). You can use the browser to download the data. I know it is a little bit cumbersome but this may be the only way (I also use this tedious way to download the data).

Hope this helps.

Thanks for your amazing work, I also encounter the same problem when I download the tf-records, but I have the ffhq-image 1024*1024 files (https://drive.google.com/drive/folders/1tZUcXDBeOibC6jcMCtgRRz67pzrAHeHL). So did the tf-recoards data contains the same 1024 resolution faces as ffhq-image 1024*1024 ?

[Xiaoming-Zhao]

@cnnlstm My gut is that should be fine but I have not tried that. Just two things:

1. The officially-provided 1024*1024 images (the link you shared) use folder name as 69000. However, if you follow StyleGAN2-ADA's processing script to extract data from the TFRecords, the folder name is 00069. Our processed pose file use the same folder name as 00069. You need to slightly modify the dataset class:

   a) We directly read image from zipfile, you need to change it to read from disk:
   

   ml-gmpi/gmpi/datasets.py

   Line 49 in 566e9c5

   all_f_list = zipfile.ZipFile(self.zipf).namelist()

   
   b) Modify the folder name to point to the right path when you read data:
   

   ml-gmpi/gmpi/datasets.py

   Lines 62 to 63 in 566e9c5

   	lm_path = [i.replace("png", "mat") for i in im_path]
   	pose_data = [os.path.join(pose_data_path, i) for i in lm_path]

2. The zipfile of 1024x1024 occupies ~200GB. However, the officially provided 1024x1024 images (the link you shared) only occupies ~90GB. I guess it is because of no compression in the zipfile but I am not 100% sure: see this line.

Overall, I think you can definitely train GMPI with the data you already have.

[fangchangma]

@VIROBO-15 We will check with legal and see if we can release the pre-processed datasets. Please stay tuned.

Unfortunately we were informed by Apple's legal team that we could not release these pre-processed datasets because of privacy concerns.

[cnnlstm]

@cnnlstm My gut is that should be fine but I have not tried that. Just two things:

1. The officially-provided 1024*1024 images (the link you shared) use folder name as 69000. However, if you follow StyleGAN2-ADA's processing script to extract data from the TFRecords, the folder name is 00069. Our processed pose file use the same folder name as 00069. You need to slightly modify the dataset class:
   a) We directly read image from zipfile, you need to change it to read from disk:
   

   ml-gmpi/gmpi/datasets.py

   Line 49 in 566e9c5

   all_f_list = zipfile.ZipFile(self.zipf).namelist()

   b) Modify the folder name to point to the right path when you read data:
   

   ml-gmpi/gmpi/datasets.py

   Lines 62 to 63 in 566e9c5

   	lm_path = [i.replace("png", "mat") for i in im_path]
   	pose_data = [os.path.join(pose_data_path, i) for i in lm_path]

2. The zipfile of 1024x1024 occupies ~200GB. However, the officially provided 1024x1024 images (the link you shared) only occupies ~90GB. I guess it is because of no compression in the zipfile but I am not 100% sure: see this line.

Overall, I think you can definitely train GMPI with the data you already have.

Thanks for your detailed reply.

I convert the ffhq-image 1024*1024 to tf_recoard files, and obtain different resolution files, just like files in the Nvidia released tf_recoard (https://drive.google.com/drive/folders/1M24jfI-Ylb-k2EGhELSnxssWi9wGUokg).

Then I download small tf_recoard file (ffhq-r07.tfrecords) from the link, and uncompress them to images, meanwhile uncompress my own ffhq-r07.tfrecords files to images.

I check the corresponding images and found that they are exactly the same. So I believe the ffhq-r10.recoards data contains the same 1024 resolution faces as ffhq-image 1024*1024.

[Xiaoming-Zhao]

@cnnlstm Thanks a lot for your time in checking this. I think this will surely be useful for others. Hope you enjoy this repo.