Source: The ALAE and Gan part of this repo is orginiated from https://github.com/ariel415el/SimplePytorch-ALAE
Difference:
-Added New Dataset for Anime Character.
-Added Saving Function for the Image Generated or Reconstructed by ALAE
-Added Style Mixing and Image Truncation Trick by Modifying the forward function in ALAE
-Added FID and PPL Metrics' Calculation
Source: The FID part of this repo is originated from https://github.com/hukkelas/pytorch-frechet-inception-distance/tree/c4bef90e502e7e1aec2a1a8f45b259630e093f8b
Difference:
-Original Repo will crash if loading a large database since it read all images at once and they are stored in the RAM
-The new implementation will read the path to get the images during calculation of the fid
Source: The PPL is originated from https://github.com/NVlabs/stylegan/blob/master/metrics/perceptual_path_length.py
Difference:
-Slerp Function is defined in a different way
-The Handling of Latent Space W and Z has changed
Please checkout their github and thank you for their contribution
The anime dataset can be downloaded in here: https://drive.google.com/drive/folders/13H-8T0fbRxTT8SPE_5PoEG9hVUCe8sWO?usp=sharing Put the folder named Twdne-128 into data folder , them put another .pt dataset file inside the Twdne-128 Folder. Please checkout the following file structure.
The result after training for 100000 iterations is as below,
The first row : Original Image The second row : Reconstructed Image The third row : Random Generated Image by Noise
Lassests
Ldnn
...
Ldata
LTwdne-128
Ltwdne.pt
Remember to add argument when you run training:
python train_StyleALAE--dataset_name twdne
If you do not have two dataset for comparison, you need to generate a reconstructed one or a random generated one to compare with the original dataset. Anyways, you can just use the test dataset in the repo.
Just a reminder, the two datasets do not necessary need to have the same size
LHS is Original , RHS is Reconstructed Image
Command:
python ./score_calculator/fid.py --path1 ../data/Test/original --path2 ../data/Test/reconstruct --batch-size 8
Result:
Looking for images in ../data/Test/original/*.png
Looking for images in ../data/Test/original/*.jpg
Looking for images in ../data/Test/reconstruct/*.png
Looking for images in ../data/Test/reconstruct/*.jpg
Calculating Activation
Progress: [------------------->] 100 %--Done--
Calculating Activation
Progress: [------------------->] 100 %--Done--
The FID of these datasets : 133.45649989750808
Finished
Please download the vgg.pt file from Nvidia Website : 'https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metrics/vgg16.pt' or download it from Google Drive: https://drive.google.com/file/d/1EC1m_4Ae-54zYSfW8swKSyrmou-MSQtj/view?usp=sharing
Then , Put the vgg16.pt into the "score_calculator" folder
Command:
python score_calculator/ppl.py
Result:
Checkpoint ckpt_gs-120000_res-5=64x64_alpha-0.40.pt loaded. Starting from resolution number 5 and global_step 120000
Calculating on w latent space , path length is full
Progress: [------------------> ] 99 %Finished
The PPL is 152.44447
Calculating on w latent space , path length is end
Progress: [------------------> ] 99 %Finished
The PPL is 96.42851
Calculating on z latent space , path length is full
Progress: [------------------> ] 99 %Finished
The PPL is 384.7624
Calculating on z latent space , path length is end
Progress: [------------------> ] 99 %Finished
The PPL is 424.26343
Process finished with exit code 0
If you want to reconstruct / generate image or perform StyleMixing / Truncation , you can run the test_StyleALAE file, but remember that you need to set up the pretrained model path:
-Select your model in the path
-Select the location of saving in the saved_path
-Remember to set the alpha value to a correct value , this info should be available on the name of your pt file
-Remember only to Uncomment one of the function: generate_img , generate_img_with_truncation , reconstruct_img , generate_style_mixing_img
if __name__ == '__main__':
path = "./data/FFHQ-thumbnails/thumbnails128x128"
reconstruction_saved_path = "./data/FFHQ-thumbnails/reconstructed_thumbnails128x128"
generation_saved_path = "./data/FFHQ-thumbnails/generated_thumbnails128x128"
model = StyleALAE(model_config=config, device=device)
model.load_train_state('./archived/FFHQ/StyleALAE-z-256_w-256_prog-(4,256)-(8,256)-(16,128)-(32,128)-(64,64)-(64,32)/checkpoints/ckpt_gs-120000_res-5=64x64_alpha-0.40.pt')
batch_size = 32
generate_img(model, config, generation_saved_path, True, 3, alpha=0.4, batch_size=32)
# reconstruct_img(model, config, path, reconstruction_saved_path, True, alpha=0.4)
# generate_style_mixing_img(model, config, alpha=0.4, copystyleto=[[4,8],[8,16],[16,32],[32,64]])
# generate_img_with_truncation(model, config, alpha=0.4)
Command:
python test_StyleALAE
Function:
generate_img(model, config, saved_path, show, generation_size, alpha=0.4, batch_size=32)
show - option for displaying the generated image
generation size - represent how many images are going to be generated,
Result:
Function:
reconstruct_img(model, config, path, saved_path, show, alpha=0.4)
show - option for displaying the generated image
Result:
Function:
generate_style_mixing_img(model, config, alpha=0.4, copystyleto=[4,8])
Two latent code are used in style mixing , -The main latent code cotributes to most input of generator -The latent code for copying contributes to part of the input, this can be control using the "copystyleto" option
copystyleto - option for using the latent code for copying in different resolutions
Result:
Function:
generate_img_with_truncation(model, config, alpha=0.4)
Result
