Skip to content

Latest commit

 

History

History
128 lines (104 loc) · 3.88 KB

UNUSED_INFO.md

File metadata and controls

128 lines (104 loc) · 3.88 KB
Raw

SSuperGAN Module

  • An example code piece to run a forward pass with SStyleGAN model:
import torch
import torch.nn as nn
import torchvision
import torchvision.models as models
import numpy as np

from networks.ssupergan import SSuperGAN

seq_args = {
    "lstm_hidden": 256,    # Hidden size of LSTM module
    "embed": 256,          # Last size for mean and std outputs
    "cnn_embed": 2048,     # The output dim retrieved from CNN embedding module
    "fc_hiddens": [],      # Sizes of FC layers after LSTM output, if there are any
    "lstm_dropout": 0,     # Set to 0 if num_lstm_layers is 1, otherwise set to [0, 0.5]
    "fc_dropout": 0,       # Dropout ratio of FC layers if there are any
    "num_lstm_layers": 1   # Number of stacked LSTM layers
}

gen_img_dim = 64           # Dimension of the generated image

model = SSuperGAN(seq_args, image_dim=gen_img_dim).cuda()

C, W, H = 3, 360, 360
x = torch.randn(16, 3, C, H, W).cuda()                    # Sample sequential images
y = torch.randn(16, 3, gen_img_dim, gen_img_dim).cuda()   # Sample target images to generate

with torch.no_grad():
    kl, recon, discr = model(x, y)

print("KL Loss:", kl)
print("Reconstruction Loss:", recon)
print("Discriminator Loss:", discr)
  • To train the SSuperGAN network, you have to specify the following parameters in the ssupergan_config.yaml file under the configs folder.
# Sequential Encoder Configs
lstm_hidden: 256       # hidden size of LSTM module
cnn_embed: 2048        # the output dim retrieved from CNN embedding module
fc_hiddens: []         # sizes of FC layers after LSTM output, if there are any
lstm_dropout: 0        # set to 0 if num_lstm_layers is 1, otherwise set to [0, 0.5]
fc_dropout: 0          # dropout ratio of FC layers if there are any
num_lstm_layers: 1     # lstm num_layers parameter

# GAN Configs
image_dim: 64          # Size of the face image to processs
latent_dim: 256        # Latent z dimension
g_hidden_size: 1024    # Generator hidden size dimension
d_hidden_size: 1024    # Discriminator hidden size dimension
e_hidden_size: 1024    # Encoder hidden size dimension

# GAN Trainig Confings
batch_size: 8
train_epochs: 10

discriminator_lr: 0.0002
discriminator_weight_decay: 0.000025
discriminator_beta_1: 0.5
discriminator_beta_2: 0.999

generator_lr: 0.0002
generator_weight_decay: 0.000025
generator_beta_1: 0.5
generator_beta_2: 0.999

seq_encoder_lr: 0.0002
seq_encoder_weight_decay: 0.000025
seq_encoder_beta_1: 0.5
seq_encoder_beta_2: 0.999

BiGAN Module

  • In order to run the module 'bigan_config.yaml' file should be created under configs.
  • Example Config:
face_image_folder_train_path: /home/gsoykan20/Desktop/ffhq_thumbnails/thumbnails128x128/
face_image_folder_test_path: /home/gsoykan20/Desktop/ffhq_thumbnails/thumbnails128x128/
num_training_samples: 10240
test_samples_range:
    - 10240
    - 10640
image_dim: 32
batch_size: 32
train_epochs: 10
discriminator_lr: 0.0002
discriminator_weight_decay: 0.000025
discriminator_beta_1: 0.5
discriminator_beta_2: 0.999
generator_lr: 0.0002
generator_weight_decay: 0.000025
generator_beta_1: 0.5
generator_beta_2: 0.999

Face Recognition Module

  • In order to run the module 'face_recognition_config.yaml' file should be created under configs.
  • Example Config:
face_image_folder_train_path: /datasets/iCartoonFace2020/personai_icartoonface_rectrain/icartoonface_rectrain
face_image_folder_test_path: /datasets/iCartoonFace2020/personai_icartoonface_rectrain/icartoonface_rectrain
num_training_samples: 1024
test_samples_range: 
    - 388000
    - 389678
image_dim: 64
batch_size: 128
train_epochs: 5

Comment: > 
    face_image_folder_path: /home/gsoykan20/Desktop/amazing-mysteries-gutter-comics/comics/data/raw_panel_images
    TODO: Add VALIDATION PATH
    total number of images in icf train set : 389678
    num_training_samples: 20480
    test_samples_range: 
        - 20480
        - 22528