Conditional-DCGAN using Pytorch

We use the c-Deep Convolutional Generative Adverserial Network model to generate animated faces 👫 and emojis 😃 from a random noise fed as input.

Refrence:Pytorch REF paper

Requirements

 1. Pytorch

Usage

Prepare Training Dataset

Create folder named data . Then create folders according to your class-label names(n folders for n classes).Then place all images into the corresponding folder(class-label).

Folder structure

Root
│
├── data
│    └── train
│          ├── ONE
│          │    ├── ONE_0.png 
│    	   │	├── ONE_1.png
│    	   │	└── ...
│    	   ├── TWO
│          │    │ 
│    	   │	├── TWO_0.png
│    	   │	└── TWO_1.png
│    	   ├──	.
│    	   └── 	.
│
│ 
│
├── c-dcgan.ipynb
├── G_chk.pt
└── D_chk.pt 

Larger the number and variation in images, larger the accuracy of classification.

c-DCGAN Architecture

Generator Neural Network

Generator (
  (DCV1_data): ConvTranspose2d(100, 256, kernel_size=(4, 4), stride=(1, 1))
  (BN1_data): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
  (DCV1_label): ConvTranspose2d(2, 256, kernel_size=(4, 4), stride=(1, 1))
  (BN1_label): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
  (DCV2): ConvTranspose2d(512, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
  (BN2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
  (DCV3): ConvTranspose2d(256, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
  (BN3): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
  (DCV4): ConvTranspose2d(128, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
  (BN4): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
  (DCV5): ConvTranspose2d(64, 3, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
)

Discriminator Neural Network

Discriminator (
  (CV1_data): Conv2d(3, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
  (CV1_label): Conv2d(2, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
  (CV2): Conv2d(128, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
  (BN2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
  (CV3): Conv2d(256, 512, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
  (BN3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
  (CV4): Conv2d(512, 1024, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))
  (BN4): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True)
  (CV5): Conv2d(1024, 1, kernel_size=(4, 4), stride=(1, 1))
)

Training

Run the c-dcgan.ipynb jupyter notebook to start the Training.

Once the training is completed the result is stored the main folder as D_chk_100.pt and G_chk_100.pt

Input-data

Sampled mini-batch images

Training through Epochs

Final Generated Images

Loss

ToDo

• Train.
• Save/Load checkpoint.
• Generate.
• UI for input, output and to download checkpoint.

Credits

@inproceedings{aneja2016modeling,
  title={Modeling Stylized Character Expressions via Deep Learning},
  author={Aneja, Deepali and Colburn, Alex and Faigin, Gary and Shapiro, Linda and Mones, Barbara},
  booktitle={Asian Conference on Computer Vision},
  pages={136--153},
  year={2016},
  organization={Springer}
}