Skip to content
Snippets Groups Projects
Commit 714007eb authored by Guillaume HEUSCH's avatar Guillaume HEUSCH
Browse files

[doc] added some images in the DR-GAN section

parent 9e2cb575
No related branches found
No related tags found
No related merge requests found
......@@ -66,6 +66,15 @@ and you can generate faces based on this dataset by doing::
$ ./bin/train_conditionalgan_casia.py
Also, the "Improved Wasserstein with gradient penalty" training procedure has been implemented for
this model. But the first attempt using it does not reach good generation abilities (and you should
carefully tune the discriminator learning rate), and seems to be less stable.
.. figure:: img/iw-gp-conditionalgan-iteration-33000.png
:align: center
Iteration 33000 of the Improved Wasserstein training.
.. todo:: script that samples from the models
......@@ -37,4 +37,39 @@ The generator model consists in three parts:
Architecture of the DR-GAN
To train the DR-GAN, simply execute the following script::
$ ./bin/train_drgan_multipie.py
To monitor the training, we show a randomly selected real example (left)
and the corresponding generated images of the supposedly same identity
*at a random pose* (right)
.. figure:: img/dr-gan-epoch-1-174.png
:align: center
Epoch 1: Target pose (left profile) not met
.. figure:: img/dr-gan-epoch-1-80.png
:align: center
Epoch 1: Target pose is better, but ID is clearly not
.. figure:: img/dr-gan-epoch-10-244.png
:align: center
Epoch 10: Target pose met, good quality (frontal), but ID problem
.. figure:: img/dr-gan-epoch-50-176.png
:align: center
Epoch 50: Able to generate view from not frontal image
.. figure:: img/dr-gan-epoch-51-109.png
:align: center
Epoch 50: Going from one profile to another
doc/img/dr-gan-epoch-1-174.png

44.4 KiB

doc/img/dr-gan-epoch-1-80.png

43.7 KiB

doc/img/dr-gan-epoch-10-244.png

44 KiB

doc/img/dr-gan-epoch-50-176.png

40.9 KiB

doc/img/dr-gan-epoch-51-109.png

39.5 KiB

doc/img/iw-gp-conditionalgan-iteration-33000.png

121 KiB

0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment