NFNet Pytorch Implementation
This repo contains pretrained NFNet models F0-F6 with high ImageNet accuracy from the paper High-Performance Large-Scale Image Recognition Without Normalization. The small models are as accurate as an EfficientNet-B7, but train 8.7 times faster. The large models set a new SOTA top-1 accuracy on ImageNet.
| NFNet | F0 | F1 | F2 | F3 | F4 | F5 | F6+SAM |
|---|---|---|---|---|---|---|---|
| Top-1 accuracy Brock et al. | 83.6 | 84.7 | 85.1 | 85.7 | 85.9 | 86.0 | 86.5 |
| Top-1 accuracy this implementation | 82.82 | 84.63 | 84.90 | 85.46 | 85.66 | 85.62 | TBD |
All credits go to the authors of the original paper. This repo is heavily inspired by their nice JAX implementation in the official repository. Visit their repo for citing.
Get started
git clone https://github.com/benjs/nfnets_pytorch.git
pip3 install -r requirements.txt
Download pretrained weights from the official repository and place them in the pretrained folder.
from pretrained import pretrained_nfnet
model_F0 = pretrained_nfnet('pretrained/F0_haiku.npz')
model_F1 = pretrained_nfnet('pretrained/F1_haiku.npz')
# ...
The model variant is automatically derived from the parameter count in the pretrained weights file.
Validate yourself
python3 eval.py --pretrained pretrained/F0_haiku.npz --dataset path/to/imagenet/valset/
You can download the ImageNet validation set from the ILSVRC2012 challenge site after asking for access with, for instance, your .edu mail address.
Scaled weight standardization convolutions in your own model
Simply replace all your nn.Conv2d with WSConv2D and all your nn.ReLU with VPReLU or VPGELU (variance preserving ReLU/GELU).
import torch.nn as nn
from model import WSConv2D, VPReLU, VPGELU
# Simply replace your nn.Conv2d layers
class MyNet(nn.Module):
def __init__(self):
super(MyNet, self).__init__()
self.activation = VPReLU(inplace=True) # or VPGELU
self.conv0 = WSConv2D(in_channels=128, out_channels=256, kernel_size=1, ...)
# ...
def forward(self, x):
out = self.activation(self.conv0(x))
# ...
SGD with adaptive gradient clipping in your own model
Simply replace your SGD optimizer with SGD_AGC.
from optim import SGD_AGC
optimizer = SGD_AGC(
named_params=model.named_parameters(), # Pass named parameters
lr=1e-3,
momentum=0.9,
clipping=0.1, # New clipping parameter
weight_decay=2e-5,
nesterov=True)
It is important to exclude certain layers from clipping or momentum. The authors recommends to exclude the last fully convolutional from clipping and the bias/gain parameters from weight decay:
import re
for group in optimizer.param_groups:
name = group['name']
# Exclude from weight decay
if len(re.findall('stem.*(bias|gain)|conv.*(bias|gain)|skip_gain', name)) > 0:
group['weight_decay'] = 0
# Exclude from clipping
if name.startswith('linear'):
group['clipping'] = None
Train your own NFNet
Adjust your desired parameters in default_config.yaml and start training.
python3 train.py --dataset /path/to/imagenet/
There is still some parts missing for complete training from scratch:
- Multi-GPU training
- Data augmentations
- FP16 activations and gradients
Contribute
The implementation is still in an early stage in terms of usability / testing. If you have an idea to improve this repo open an issue, start a discussion or submit a pull request.
Development status
- Pre-trained NFNet Models
- F0-F5
- F6+SAM
- Scaled weight standardization
- Squeeze and excite
- Stochastic depth
- FP16 activations
- SGD with unit adaptive gradient clipping (SGD-AGC)
- Exclude certain layers from weight-decay, clipping
- FP16 gradients
- PyPI package
- PyTorch hub submission
- Label smoothing loss from Szegedy et al.
- Training on ImageNet
- Pre-trained weights
- Tensorboard support
- general usability improvements
- Multi-GPU support
- Data augmentation
- Signal propagation plots (from first paper)
0 Oct 18, 2021
63 Oct 26, 2021
4 Jul 04, 2021
134 Dec 06, 2022
78 Dec 27, 2022
492 Jan 06, 2023
157 Dec 26, 2022
2 Mar 02, 2022
137 Dec 20, 2022
495 Dec 28, 2022
15 Dec 15, 2022
2.1k Jan 09, 2023
20 Jun 29, 2022
724 Jan 04, 2023
55 Dec 21, 2022
40 Nov 03, 2022
25 Dec 03, 2022
17 Nov 05, 2022
13k Dec 29, 2022
15 Dec 24, 2022