Commit 78a92c7e authored by Guillaume HEUSCH's avatar Guillaume HEUSCH
Browse files

[doc] added doc for Conditional GAN

parent e6a05829
Pipeline #22458 passed with stage
in 10 minutes and 22 seconds
.. py:currentmodule:: bob.learn.pytorch
Conditional GAN
The conditional GAN is an extension of the original GAN, by adding a conditioning variable
in the process.
As an illustration, consider MNIST digits: instead of generating a digit between 0 and 9, the
condition variable would allow to generate a particular digit.
More particularly, the input to the generator (i.e. noise) will be
concatenated with a variable specifying the particular condition to generate the fake data.
Also, there are different strategies to embed the conditioning variable in the discriminator.
Here we chose to concatenate the one-hot encoded conditional variable as feature maps to the
input image.
.. figure:: img/cond_map.png
:align: center
Conditional input to the discriminator
The article describing this algorithm is the following [cgan]_::
Author = {Mirza, M. and Osindero, S.},
Title = {Conditional {G}enerative {A}dversarial {N}ets},
eprint = {arXiv:1411.1784},
seq-number = {49},
year = 2014
There are other articles dealing specifically with conditional GANs for face processing, and this work is
loosely based on `Conditional Generative Adversarial Nets for Convolutional Face Generation <>`_
and inspired from the following `code <>`_.
Here we consider the conditioning variable to be the pose of the face (in terms of yaw). We will
use the Multi-PIE database, since it contains face images with 13 different poses.
First of all, we need to write a configuration file, containing the dataset and the
network definition:
.. code-block:: python
### DATA ###
from bob.learn.pytorch.datasets.multipie import MultiPIEDataset
from bob.learn.pytorch.datasets import RollChannels
from bob.learn.pytorch.datasets import ToTensor
from bob.learn.pytorch.datasets import Normalize
import torchvision.transforms as transforms
dataset = MultiPIEDataset(root_dir='/idiap/temp/heusch/data/multipie-cropped-64x64',
RollChannels(), # bob to skimage:
Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
### NETWORK ###
from bob.learn.pytorch.architectures import ConditionalGAN_generator
from bob.learn.pytorch.architectures import ConditionalGAN_discriminator
from bob.learn.pytorch.architectures import weights_init
noise_dim = 100
conditional_dim = 13
generator = ConditionalGAN_generator(noise_dim, conditional_dim)
discriminator = ConditionalGAN_discriminator(conditional_dim)
.. note::
You should have the ``bob.db.multipie`` package installed on your environment
To train the conditional GAN to generate faces with different poses, you should execute::
$ ./bin/ -vv
If you have access to a GPU, you should run the script using this option::
$ ./bin/ -vv --use-gpu
Generated samples will be saved after each training epoch. You can specify where to write
images using the ``--output-dir`` option (default to ``./conditionalgan``).
Here are some examples of generated face images with different poses:
.. figure:: img/cgan-multipie-epoch-1.png
:align: center
After 1 epoch
.. figure:: img/cgan-multipie-epoch-20.png
:align: center
After 20 epoch
.. figure:: img/cgan-multipie-epoch-50.png
:align: center
After 50 epoch
.. [cgan] *M. Mirza, S. Osindero*. **Conditional Generative Adversarial Nets** `arXiv:1411.1784. <>`__
......@@ -14,6 +14,7 @@ Users Guide
Reference Manual
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment