| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_002 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_004 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_006 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_008 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_016 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_032 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_040 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_064 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_080 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_120 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_160 |
| SHOW MORE |
| Training Techniques | SGD with Momentum, Weight Decay |
|---|---|
| Architecture | 1x1 Convolution, Batch Normalization, Convolution, Dense Connections, Global Average Pooling, Grouped Convolution, ReLU, Squeeze-and-Excitation Block |
| ID | regnety_320 |
| SHOW MORE |
RegNetY is a convolutional network design space with simple, regular models with parameters: depth $d$, initial width $w_{0} > 0$, and slope $w_{a} > 0$, and generates a different block width $u_{j}$ for each block $j < d$. The key restriction for the RegNet types of model is that there is a linear parameterisation of block widths (the design space only contains models with this linear structure):
$$ u_{j} = w_{0} + w_{a}\cdot{j} $$
For RegNetX authors have additional restrictions: we set $b = 1$ (the bottleneck ratio), $12 \leq d \leq 28$, and $w_{m} \geq 2$ (the width multiplier).
For RegNetY authors make one change, which is to include Squeeze-and-Excitation blocks.
To load a pretrained model:
import timm
m = timm.create_model('regnety_008', pretrained=True)
m.eval()
Replace the model name with the variant you want to use, e.g. regnety_008. You can find the IDs in the model summaries at the top of this page.
You can follow the timm recipe scripts for training a new model afresh.
@misc{radosavovic2020designing,
title={Designing Network Design Spaces},
author={Ilija Radosavovic and Raj Prateek Kosaraju and Ross Girshick and Kaiming He and Piotr Dollár},
year={2020},
eprint={2003.13678},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
Image Classification on ImageNet
| MODEL | TOP 1 ACCURACY | TOP 5 ACCURACY |
|---|---|---|
| regnety_032 | 82.01% | 95.91% |
| regnety_320 | 80.8% | 95.25% |
| regnety_120 | 80.38% | 95.12% |
| regnety_160 | 80.28% | 94.97% |
| regnety_080 | 79.87% | 94.83% |
| regnety_064 | 79.73% | 94.76% |
| regnety_040 | 79.23% | 94.64% |
| regnety_016 | 77.87% | 93.73% |
| regnety_008 | 76.32% | 93.07% |
| regnety_006 | 75.27% | 92.53% |
| regnety_004 | 74.02% | 91.76% |
| regnety_002 | 70.28% | 89.55% |
