PyTorch implementation of Normalizer-Free Networks and SGD - Adaptive Gradient Clipping

Paper: https://arxiv.org/abs/2102.06171.pdf

Original code: https://github.com/deepmind/deepmind-research/tree/master/nfnets

Do star this repository if it helps your work!

Note: See this comment for a generic implementation for any optimizer as a temporary reference for anyone who needs it.

Installation

Install from PyPi:

pip3 install nfnets-pytorch

or install the latest code using:

pip3 install git+https://github.com/vballoli/nfnets-pytorch

Usage

WSConv2d

Use WSConv2d and WSConvTranspose2d like any other torch.nn.Conv2d or torch.nn.ConvTranspose2d modules.

import torch
from torch import nn
from nfnets import WSConv2d

conv = nn.Conv2d(3,6,3)
w_conv = WSConv2d(3,6,3)

conv_t = nn.ConvTranspose2d(3,6,3)
w_conv_t = WSConvTranspose2d(3,6,3)

Generic AGC (recommended)

import torch
from torch import nn, optim
from torchvision.models import resnet18

from nfnets import WSConv2d
from nfnets.agc import AGC # Needs testing

conv = nn.Conv2d(3,6,3)
w_conv = WSConv2d(3,6,3)

optim = optim.SGD(conv.parameters(), 1e-3)
optim_agc = AGC(conv.parameters(), optim) # Needs testing

# Ignore fc of a model while applying AGC.
model = resnet18()
optim = torch.optim.SGD(model.parameters(), 1e-3)
optim = AGC(model.parameters(), optim, model=model, ignore_agc=['fc'])

SGD - Adaptive Gradient Clipping

Similarly, use SGD_AGC like torch.optim.SGD

# The generic AGC is preferable since the paper recommends not applying AGC to the last fc layer.
import torch
from torch import nn, optim
from nfnets import WSConv2d, SGD_AGC

conv = nn.Conv2d(3,6,3)
w_conv = WSConv2d(3,6,3)

optim = optim.SGD(conv.parameters(), 1e-3)
optim_agc = SGD_AGC(conv.parameters(), 1e-3)

Using it within any PyTorch model

import torch
from torch import nn
from torchvision.models import resnet18

from nfnets import replace_conv

model = resnet18()
replace_conv(model)

Docs

Find the docs at readthedocs

TODO

• WSConv2d
• SGD - Adaptive Gradient Clipping
• Function to automatically replace Convolutions in any module with WSConv2d
• Documentation
• Generic AGC wrapper.(See this comment for a reference implementation) (Needs testing for now)
• WSConvTranspose2d
• NFNets
• NF-ResNets

Cite Original Work

To cite the original paper, use:

@article{brock2021high,
  author={Andrew Brock and Soham De and Samuel L. Smith and Karen Simonyan},
  title={High-Performance Large-Scale Image Recognition Without Normalization},
  journal={arXiv preprint arXiv:},
  year={2021}
}