This Python package implements the activation function transformations, weight initializations, and dataset preprocessing used in Deep Kernel Shaping (DKS) and Tailored Activation Transformations (TAT). DKS and TAT, which were introduced in the DKS paper and TAT paper, are methods for constructing/transforming neural networks to make them much easier to train. For example, these methods can be used in conjunction with K-FAC to train deep vanilla deep convnets (without skip connections or normalization layers) as fast as standard ResNets of the same depth.
The package supports the JAX, PyTorch, and TensorFlow tensor programming frameworks.
Questions/comments about the code can be sent to [email protected].
NOTE: we are not taking code contributions from Github at this time. All PRs from Github will be rejected. Instead, please email us if you find a bug.
For each of the supported tensor programming frameworks, there is a
corresponding subpackage which handles the activation function transformations,
weight initializations, and (optional) data preprocessing. (These are dks.jax
,
dks.pytorch
, and dks.tensorflow
.) It's up to the user to import these and
use them appropriately within their model code. Activation functions are
transformed by the function get_transformed_activations()
in the module
activation_transform
of the appropriate subpackage. Sampling initial
parameters is done using functions inside of the module
parameter_sampling_functions
of said subpackage. And data preprocessing is
done using the function per_location_normalization
inside of the module
data_preprocessing
of said subpackage. Note that in order to avoid having to
import all of the tensor programming frameworks, the user is required to
individually import whatever framework subpackage they want. e.g. import dks.jax
. Meanwhile, import dks
won't actually do anything.
get_transformed_activations()
requires the user to pass either the "maximal
slope function" for DKS, the "subnet maximizing function" for TAT with Leaky
ReLUs, or the "maximal curvature function" for TAT with smooth activation
functions. (The subnet maximizing function also handles DKS and TAT with smooth
activations.) These are special functions that encode information about the
particular model architecture. See the section titled "Summary of our method" of
the DKS paper for a procedure to construct the maximal slope function for a
given model, or the appendix section titled "Additional details and pseudocode
for activation function transformations" of the TAT paper for procedures to
construct the other two functions.
In addition to these things, the user is responsible for ensuring that their model meets the architectural requirements of DKS/TAT, and for converting any weighted sums into "normalized sums" (which are weighted sums whose non-trainable weights have a sum of squares equal to 1). See the section titled "Summary of our method" of the DKS paper for more details.
Note that the data preprocessing method implemented, called Per-Location Normalization (PLN), may not always be needed in practice, but we have observed certain situations where not using can lead to problems. (For example, training on datasets that contain all-zero pixels, such as CIFAR-10.) Also note that ReLUs are only partially supported by DKS, and unsupported by TAT, and so their use is highly discouraged. Instead, one should use Leaky ReLUs, which are fully supported by DKS, and work especially well with TAT.
dks.examples.haiku.modified_resnet
is a Haiku ResNet model which has been
modified as described in the DKS/TAT papers, and includes support for both DKS
and TAT. When constructed with its default arguments, it removes the
normalization layers and skip connections found in standard ResNets, making it a
"vanilla network". It can be used as an instructive example for how to build
DKS/TAT models using this package. See the section titled "Application to
various modified ResNets" from the DKS paper for more details.
This package can be installed directly from GitHub using pip
with
pip install git+https://github.com/deepmind/dks.git
or
pip install -e git+https://github.com/deepmind/dks.git#egg=dks[<extras>]
Or from PyPI with
pip install dks
or
pip install dks[<extras>]
Here <extras>
is a common-separated list of strings (with no spaces) that can
be passed to install extra dependencies for different tensor programming
frameworks. Valid strings are jax
, tf
, and pytorch
. So for example, to
install dks
with the extra requirements for JAX and PyTorch, one does
pip install dks[jax,pytorch]
To run tests in a Python virtual environment with specific pinned versions of all the dependencies one can do:
git clone https://github.com/deepmind/dks.git
cd dks
./test.sh
However, it is strongly recommended that you run the tests in the same Python
environment (with the same package versions) as you plan to actually use dks
.
This can be accomplished by installing dks
for all three tensors programming
frameworks (e.g. with pip install dks[jax,pytorch,tf]
or some other
installation method), and then doing
pip install pytest-xdist
git clone https://github.com/deepmind/dks.git
cd dks
python -m pytest -n 16 tests
This is not an official Google product.