asdfghjkl

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