ID	seresnext26d_32x4d
LR	0.6
Epochs	100
Layers	26
Dropout	0.2
Crop Pct	0.875
Momentum	0.9
Batch Size	1024
Image Size	224
Interpolation	bicubic

ID	seresnext26t_32x4d
LR	0.6
Epochs	100
Layers	26
Dropout	0.2
Crop Pct	0.875
Momentum	0.9
Batch Size	1024
Image Size	224
Interpolation	bicubic

ID	seresnext50_32x4d
LR	0.6
Epochs	100
Layers	50
Dropout	0.2
Crop Pct	0.875
Momentum	0.9
Batch Size	1024
Image Size	224
Interpolation	bicubic

SEResNeXt

rwightman / pytorch-image-models

Last updated on Feb 14, 2021

Parameters 17 Million

FLOPs 4 Billion

File Size 64.30 MB

Training Data ImageNet

Training Resources 8x NVIDIA Titan X GPUs

Training Time

Training Techniques	SGD with Momentum, Weight Decay, Label Smoothing
Architecture	1x1 Convolution, Squeeze-and-Excitation Block, Batch Normalization, Convolution, Grouped Convolution, Global Average Pooling, ResNeXt Block, Residual Connection, ReLU, Max Pooling, Softmax
ID	seresnext26d_32x4d
LR	0.6
Epochs	100
Layers	26
Dropout	0.2
Crop Pct	0.875
Momentum	0.9
Batch Size	1024
Image Size	224
Interpolation	bicubic
SHOW MORE
SHOW LESS

Parameters 17 Million

FLOPs 3 Billion

File Size 64.29 MB

Training Data ImageNet

Training Resources 8x NVIDIA Titan X GPUs

Training Time

Training Techniques	SGD with Momentum, Weight Decay, Label Smoothing
Architecture	1x1 Convolution, Squeeze-and-Excitation Block, Batch Normalization, Convolution, Grouped Convolution, Global Average Pooling, ResNeXt Block, Residual Connection, ReLU, Max Pooling, Softmax
ID	seresnext26t_32x4d
LR	0.6
Epochs	100
Layers	26
Dropout	0.2
Crop Pct	0.875
Momentum	0.9
Batch Size	1024
Image Size	224
Interpolation	bicubic
SHOW MORE
SHOW LESS

Parameters 28 Million

FLOPs 5 Billion

File Size 105.45 MB

Training Data ImageNet

Training Resources 8x NVIDIA Titan X GPUs

Training Time

Training Techniques	SGD with Momentum, Weight Decay, Label Smoothing
Architecture	1x1 Convolution, Squeeze-and-Excitation Block, Batch Normalization, Convolution, Grouped Convolution, Global Average Pooling, ResNeXt Block, Residual Connection, ReLU, Max Pooling, Softmax
ID	seresnext50_32x4d
LR	0.6
Epochs	100
Layers	50
Dropout	0.2
Crop Pct	0.875
Momentum	0.9
Batch Size	1024
Image Size	224
Interpolation	bicubic
SHOW MORE
SHOW LESS

README.md

Summary

SE ResNeXt is a variant of a ResNext that employs squeeze-and-excitation blocks to enable the network to perform dynamic channel-wise feature recalibration.

How do I load this model?

To load a pretrained model:

import timm
m = timm.create_model('seresnext26d_32x4d', pretrained=True)
m.eval()

Replace the model name with the variant you want to use, e.g. seresnext26d_32x4d. You can find the IDs in the model summaries at the top of this page.

How do I train this model?

You can follow the timm recipe scripts for training a new model afresh.

Citation

@misc{hu2019squeezeandexcitation,
      title={Squeeze-and-Excitation Networks}, 
      author={Jie Hu and Li Shen and Samuel Albanie and Gang Sun and Enhua Wu},
      year={2019},
      eprint={1709.01507},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Results

Image Classification on ImageNet

Image Classification

BENCHMARK	MODEL	METRIC NAME	METRIC VALUE	GLOBAL RANK
ImageNet	seresnext50_32x4d	Top 1 Accuracy	81.27%	# 63
		Top 5 Accuracy	95.62%	# 63
ImageNet	seresnext26t_32x4d	Top 1 Accuracy	77.99%	# 168
		Top 5 Accuracy	93.73%	# 168
ImageNet	seresnext26d_32x4d	Top 1 Accuracy	77.59%	# 178
		Top 5 Accuracy	93.61%	# 178