ASDL: Automatic Second-order Differentiation (for Fisher, Gradient covariance, Hessian, Jacobian, and Kernel) Library
Project description
ASDL: Automatic Second-order Differentiation Library
ASDL is an extension library of PyTorch to easily perform gradient preconditioning using second-order information (e.g., Hessian, Fisher information) for deep neural networks.
ASDL provides various implementations and a unified interface (GradientMaker) for gradient preconditioning for deep neural networks. For example, to train your model with gradient preconditioning by K-FAC algorithm, you can replace a <Standard>
gradient calculation procedure (i.e., a forward pass followed by a backward pass) with one by <ASDL>
with KfacGradientMaker like the following:
from asdl.precondition import PreconditioningConfig, KfacGradientMaker
# Initialize model
model = Net()
# Initialize optimizer (SGD is recommended)
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
# Initialize KfacGradientMaker
config = PreconditioningConfig(data_size=batch_size, damping=0.01)
gm = KfacGradientMaker(model, config)
# Training loop
for x, t in data_loader:
optimizer.zero_grad()
# <Standard> (gradient calculation)
# y = model(x)
# loss = loss_fn(y, t)
# loss.backward()
# <ASDL> ('preconditioned' gradient calculation)
dummy_y = gm.setup_model_call(model, x)
gm.setup_loss_call(loss_fn, dummy_y, t)
y, loss = gm.forward_and_backward()
optimizer.step()
You can apply a different gradient preconditioning algorithm by replacing gm
with another XXXGradientMaker(model, config)
(XXX: algorithm name, e.g., ShampooGradientMaker for Shampoo algorithm) with the same interface.
This enables a flexible switching/comparison of a range of gradient preconditioning algorithms.
Installation
You can install the latest version of ASDL by running:
$ pip install asdfghjkl
ASDL is tested with Python 3.7 and is compatible with PyTorch 1.13.
Example
The training script for training MLPs, CNNs, or ResNets using varous types of gradient preconditionig methods (which reproduces the results in the ASDL paper).
Resource
- ASDL paper
- ASDL poster @ HOOML2022 workshop at NeurIPS 2022
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file asdfghjkl-0.1a4.tar.gz
.
File metadata
- Download URL: asdfghjkl-0.1a4.tar.gz
- Upload date:
- Size: 75.7 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/5.1.1 CPython/3.9.4
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | a934411a0ffdee6fcdccb19672196498ea6a8e55e3e67abbe67200c84b46ddee |
|
MD5 | b936ad5784c919929072f5cb45f7e56a |
|
BLAKE2b-256 | c573c81489fe1ed8b12099c8e034e8e623374016e5e6db2010c38d8f386acc2d |
File details
Details for the file asdfghjkl-0.1a4-py3-none-any.whl
.
File metadata
- Download URL: asdfghjkl-0.1a4-py3-none-any.whl
- Upload date:
- Size: 89.7 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/5.1.1 CPython/3.9.4
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 9cedec38494d7907075941d03d018d8c2b64047e6c41713a49f30e3790816c51 |
|
MD5 | 5cdd6f457d28249dea0c576c76e3ccf8 |
|
BLAKE2b-256 | 329a7548968acfacb162c88dd9c2247a54f36136221ebb13d457016b28d47523 |