Generalizing Across Domains via Cross-Gradient Training
Tensorflow implementation of the paper Generalizing Across Domains via Cross-Gradient Training.
Usage
The file crossgrad.py contains the base tensorflow implementation with no dependencies (except for tensorflow itself and the util.py supplied in this same repository).
An example of usage can be found in the notebook CrossGrad.ipynb, which requires to run:
- SenseTheFlow branch v3.0 (https://github.com/gpascualg/SenseTheFlow/tree/v3.0)
- MNIST python wrappers (https://github.com/datapythonista/mnist)
Understanding crossgrad function
In crossgrad.py you can find the base function for the model. Its parameters map to the paper ones as (see Figure 2, Algorithm 1):
-
labels_fn:
$C_{\theta_l}$ -
latent_fn:
$G_{\theta^1_d}$ (defaults to a simple CNN + dense network with final dimension controlled by latent_space_dimensions) -
domain_fn:
$S_{\theta^2_d}$ (defaults to a simple 2-dense network) - $x$: As the paper, features
-
domain:
$d$ -
labels:
$y$ -
epsilon_d:
$\epsilon_d$ -
epsilon_l:
$\epsilon_l$ -
alpha_d:
$\alpha_d$ -
alpha_l:
$\alpha_l$ - params: Not in the paper, some configuration (such as 'summaries' and 'data_format')