(NeurIPS 2021 Spotlight) [arXiv]
Second-order Neural ODE Optimizer
✔️
faster convergence in wall-clock time |
✔️
O(1) memory cost |
✔️
better test-time performance |
✔️
architecture co-optimization
This repo provides PyTorch code of Second-order Neural ODE Optimizer (SNOpt), a second-order optimizer for training Neural ODEs that retains O(1) memory cost with superior convergence and test-time performance.
Installation
This code is developed with Python3. PyTorch >=1.7 (we recommend 1.8.1) and torchdiffeq >= 0.2.0 are required.
- Install the dependencies with Anaconda and activate the environment
snopt
withconda env create --file requirements.yaml python=3 conda activate snopt
- [Optional] This repo provides a modification (with 15 lines!) of
torchdiffeq
that allows SNOpt to collect 2nd-order information during adjoint-based training. If you wish to runtorchdiffeq
on other commit, simply copy-and-paste the folder to this directory then apply the providedsnopt_integration.patch
.cp -r <path_to_your_torchdiffeq_folder> . git apply snopt_integration.patch
Run the code
We provide example code for 8 datasets across image classification (main_img_clf.py
), time-series prediction (main_time_series.py
), and continuous normalizing flow (main_cnf.py
). The command lines to generate similar results shown in our paper are detailed in scripts
folder. Datasets will be automatically downloaded to data
folder at the first call, and all results will be saved to result
folder.
bash scripts/run_img_clf.sh <dataset> # dataset can be {mnist, svhn, cifar10}
bash scripts/run_time_series.sh <dataset> # dataset can be {char-traj, art-wr, spo-ad}
bash scripts/run_cnf.sh <dataset> # dataset can be {miniboone, gas}
For architecture (specifically integration time) co-optimization, run
bash scripts/run_img_clf.sh cifar10-t1-optimize
Integration with your workflow
snopt
can be integrated flawlessly with existing training work flow. Below we provide a handy checklist and pseudo-code to help your integration. For more complex examples, please refer to main_*.py
in this repo.
- Import
torchdiffeq
that is patched with snopt integration; otherwise simply usetorchdiffeq
in this repo. - Inherit
snopt.ODEFuncBase
as your vector field; implement the forward pass inF
rather thanforward
. - Create Neural ODE with ode layer(s) using
snopt.ODEBlock
; implement propertiesodes
andode_mods
. - Initialize
snopt.SNOpt
as preconditioner; calltrain_itr_setup()
andstep()
before standardoptim.zero_grad()
andoptim.step()
(see the code below). - That's it
🤓 ! Enjoy your second-order training🚂 →🚅 !
import torch
from torchdiffeq import odeint_adjoint as odesolve
from snopt import SNOpt, ODEFuncBase, ODEBlock
from easydict import EasyDict as dict
class ODEFunc(ODEFuncBase):
def __init__(self, opt):
super(ODEFunc, self).__init__(opt)
self.linear = torch.nn.Linear(input_dim, input_dim)
def F(self, t, z):
return self.linear(z)
class NeuralODE(torch.nn.Module):
def __init__(self, ode):
super(NeuralODE, self).__init__()
self.ode = ode
def forward(self, z):
return self.ode(z)
@property
def odes(self): # in case we have multiple odes, collect them in a list
return [self.ode]
@property
def ode_mods(self): # modules of all ode(s)
return [mod for mod in self.ode.odefunc.modules()]
# Create Neural ODE
opt = dict(
optimizer='SNOpt',tol=1e-3,ode_solver='dopri5',use_adaptive_t1=False,snopt_step_size=0.01)
odefunc = ODEFunc(opt)
integration_time = torch.tensor([0.0, 1.0]).float()
ode = ODEBlock(opt, odefunc, odesolve, integration_time)
net = NeuralODE(ode)
# Create SNOpt optimizer
precond = SNOpt(net, eps=0.05, update_freq=100)
optim = torch.optim.SGD(net.parameters(), lr=0.001)
# Training loop
for (x,y) in training_loader:
precond.train_itr_setup() # <--- additional step for precond
optim.zero_grad()
loss = loss_function(net(x), y)
loss.backward()
# Run SNOpt optimizer
precond.step() # <--- additional step for precond
optim.step()
What the library actually contains
This snopt
library implements the following objects for efficient 2nd-order adjoint-based training of Neural ODEs.
ODEFuncBase
: Defines the vector field (inheritstorch.nn.Module
) of Neural ODE.CNFFuncBase
: Serves the same purposes asODEFuncBase
except for CNF applications.ODEBlock
: A Neural-ODE module (torch.nn.Module
) that solves the initial value problem (given the vector field, integration time, and a ODE solver) and handles integration time co-optimization with feedback policy.SNOpt
: Our primary 2nd-order optimizer (torch.optim.Optimizer
), implemented as a "preconditioner" (see example code above). It takes the following arguments.net
is the Neural ODE. Note that the entire network (rather thannet.parameters()
) is required.eps
is the the regularization that stabilizes preconditioning. We recommend the value in [0.05, 0.1].update_freq
is the frequency to refresh the 2nd-order information. We recommend the value 100~200.alpha
decides the running averages of eigenvalues. We recommend fixing the value to 0.75.full_precond
decides whether we wish to precondition layers aside from those in Neural ODEs.
SNOptAdjointCollector
: A helper to collect information fromtorchdiffeq
to construct 2nd-order matrices.IntegrationTimeOptimizer
: Our 2nd-order method that co-optimizes the integration time (i.e.,t1
). This is done by callingt1_train_itr_setup(train_it)
andupdate_t1()
together withoptim.zero_grad()
andoptim.step()
(seetrainer.py
).
The options are passed in as opt
and contains the following fields (see options.py
for full descriptions.)
optimizer
is the training method. Use "SNOpt" to enable our method.ode_solver
specifies the ODE solver (default is "dopri5") with the absolute/relative tolerancetol
.- For CNF applications, use
divergence_type
to specify how divergence should be computed. snopt_step_size
determines the step sizes SNOpt will sample along the integration to compute 2nd-order matrices. We recommend the value 0.01 for integration time [0,1], which yield around 100 sampled points.- For integration time (t1) co-optimization, enable the flag
use_adaptive_t1
and setup the following options.adaptive_t1
specifies t1 optimization method. Choices are "baseline" and "feedback"(ours).t1_lr
is the learning rate. We recommend the value in [0.05, 0.1].t1_reg
is the coefficient of the quadratic penalty imposed on t1. The performance is quite sensitive to this value. We recommend the value in [1e-4, 1e-3].t1_update_freq
is the frequency to update t1. We recommend the value 50~100.
Remarks & Citation
The current library only supports adjoint-based training, yet it can be extended to normal odeint
method (stay tuned!). The pre-processing of tabular and uea datasets are adopted from ffjord and NeuralCDE, and the eigenvalue-regularized preconditioning is adopted from EKFAC-pytorch.
If you find this library useful, please cite
@inproceedings{liu2021second,
title={Second-order Neural ODE Optimizer},
author={Liu, Guan-Horng and Chen, Tianrong and Theodorou, Evangelos A},
booktitle={Advances in Neural Information Processing Systems},
year={2021},
}