README.md

T2F

Text-to-Face generation using Deep Learning. This project combines two of the recent architectures StackGAN and ProGAN for synthesizing faces from textual descriptions.
The project uses Face2Text dataset containing of 400 images and textual captions for each of them. The data is included in the repository under data/LFW/Face2Text/face2text_v0.1 directory.

Some Examples:

Architecture:

The textual description is encoded into a summary vector using an LSTM network. The summary vector, i.e. Embedding (psy_t) as shown in the diagram is passed through the Conditioning Augmentation block (a single linear layer) to obtain the textual part of the latent vector (uses VAE like reparameterization technique) for the GAN as input. The second part of the latent vector is random gaussian noise. The latent vector so produced is fed to the generator part of the GAN, while the embedding is fed to the final layer of the discriminator for conditional distribution matching. The training of the GAN progresses exactly as mentioned in the ProGAN paper; i.e. layer by layer at increasing spatial resolutions. The new layer is introduced using the fade-in technique to avoid destroying previous learning.

Running the code:

The code is present in the implementation/ subdirectory. The implementation is done using the PyTorch framework. So, for running this code, please install PyTorch version 0.4.0 before continuing.

Code organization:
configs: contains the configuration files for training the network. (You can use any one, or create your own)
data_processing: package containing data processing and loading modules
networks: package contains network implementation
processed_annotations: directory stores output of running process_text_annotations.py script
process_text_annotations.py: processes the captions and stores output in processed_annotations/ directory. (no need to run this script; the pickle file is included in the repo.)
train_network.py: script for running the training the network

Sample configuration:

# All paths to different required data objects
images_dir: "../data/LFW/lfw"
processed_text_file: "processed_annotations/processed_text.pkl"
log_dir: "training_runs/11/losses/"
sample_dir: "training_runs/11/generated_samples/"
save_dir: "training_runs/11/saved_models/"

# Hyperparameters for the Model
captions_length: 100
img_dims:
  - 64
  - 64

# LSTM hyperparameters
embedding_size: 128
hidden_size: 256
num_layers: 3  # number of LSTM cells in the encoder network

# Conditioning Augmentation hyperparameters
ca_out_size: 178

# Pro GAN hyperparameters
depth: 5
latent_size: 256
learning_rate: 0.001
beta_1: 0
beta_2: 0
eps: 0.00000001
drift: 0.001
n_critic: 1

# Training hyperparameters:
epochs:
  - 160
  - 80
  - 40
  - 20
  - 10

# % of epochs for fading in the new layer
fade_in_percentage:
  - 85
  - 85
  - 85
  - 85
  - 85

batch_sizes:
  - 16
  - 16
  - 16
  - 16
  - 16

num_workers: 3
feedback_factor: 7  # number of logs generated per epoch
checkpoint_factor: 2  # save the models after these many epochs
use_matching_aware_discriminator: True  # use the matching aware discriminator

Use the requirements.txt to install all the dependencies for the project.

$ workon [your virtual environment]
$ pip install -r requirements.txt

Sample run:

$ mkdir training_runs
$ mkdir training_runs/generated_samples training_runs/losses training_runs/saved_models
$ train_network.py --config=configs/11.comf

Other links:

blog: https://medium.com/@animeshsk3/t2f-text-to-face-generation-using-deep-learning-b3b6ba5a5a93
training_time_lapse video: https://www.youtube.com/watch?v=NO_l87rPDb8

#TODO:

1.) Create a simple demo.py for running inference on the trained models
2.) Separate the implementation/networks/C_Pro_GAN.py as a standalone library