An open source framework for slimmable training on tasks of ImageNet classification and COCO detection, which has enabled numerous projects. 1, 2, 3
1. Slimmable Neural Networks ICLR 2019 Paper | OpenReview | Detection | Model Zoo
Illustration of slimmable neural networks. The same model can run at different widths (number of active channels), permitting instant and adaptive accuracy-efficiency trade-offs.
2. Universally Slimmable Networks and Improved Training Techniques ICCV 2019 Paper | Model Zoo
Illustration of universally slimmable networks. The same model can run at arbitrary widths.
3. AutoSlim: Towards One-Shot Architecture Search for Channel Numbers NeurIPS 2019 Workshop Paper | Model Zoo
AutoSlimming MobileNet v1, MobileNet v2, MNasNet and ResNet-50: the optimized number of channels under each computational budget (FLOPs).
apps
dir, based on PyTorch.logs
dir and directly run command.test_only
and pretrained
in config file. You will need to manage visible gpus by yourself.python train.py app:{apps/***.yml}
. {apps/***.yml}
is config file. Do not miss app:
prefix.Model | Switches (Widths) | Top-1 Err. | FLOPs | Model ID |
---|---|---|---|---|
S-MobileNet v1 | 1.00 0.75 0.50 0.25 |
28.5 30.5 35.2 46.9 |
569M 325M 150M 41M |
a6285db |
S-MobileNet v2 | 1.00 0.75 0.50 0.35 |
29.5 31.1 35.6 40.3 |
301M 209M 97M 59M |
0593ffd |
S-ShuffleNet | 2.00 1.00 0.50 |
28.6 34.5 42.8 |
524M 138M 38M |
1427f66 |
S-ResNet-50 | 1.00 0.75 0.50 0.25 |
24.0 25.1 27.9 35.0 |
4.1G 2.3G 1.1G 278M |
3fca9cc |
Universally Slimmable Networks and Improved Training Techniques
Model | Model ID | Spectrum | |||||||||||||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
US‑MobileNet v1 | 13d5af2 | Width MFLOPs Top-1 Err. |
1.0 568 28.2 |
0.975 543 28.3 |
0.95 517 28.4 |
0.925 490 28.7 |
0.9 466 28.7 |
0.875 443 29.1 |
0.85 421 29.4 |
0.825 389 29.7 |
0.8 366 30.2 |
0.775 345 30.3 |
0.75 325 30.5 |
0.725 306 30.9 |
0.7 287 31.2 |
0.675 267 31.7 |
0.65 249 32.2 |
0.625 232 32.5 |
0.6 217 33.2 |
0.575 201 33.7 |
0.55 177 34.4 |
0.525 162 35.0 |
0.5 149 35.8 |
0.475 136 36.5 |
0.45 124 37.3 |
0.425 114 38.1 |
0.4 100 39.0 |
0.375 89 40.0 |
0.35 80 41.0 |
0.325 71 41.9 |
0.3 64 42.7 |
0.275 48 44.2 |
0.25 41 44.3 |
US‑MobileNet v2 | 3880cad | Width MFLOPs Top-1 Err. |
1.0 300 28.5 |
0.975 299 28.5 |
0.95 284 28.8 |
0.925 274 28.9 |
0.9 269 29.1 |
0.875 268 29.1 |
0.85 254 29.4 |
0.825 235 29.9 |
0.8 222 30.0 |
0.775 213 30.2 |
0.75 209 30.4 |
0.725 185 30.7 |
0.7 173 31.1 |
0.675 165 31.4 |
0.65 161 31.7 |
0.625 161 31.7 |
0.6 151 32.4 |
0.575 150 32.4 |
0.55 106 34.4 |
0.525 100 34.6 |
0.5 97 34.9 |
0.475 96 35.1 |
0.45 88 35.8 |
0.425 88 35.8 |
0.4 80 36.6 |
0.375 80 36.7 |
0.35 59 37.7 |
AutoSlim: Towards One-Shot Architecture Search for Channel Numbers
Model | Top-1 Err. | FLOPs | Model ID |
---|---|---|---|
AutoSlim-MobileNet v1 | 27.0 28.5 32.1 |
572M 325M 150M |
9b0b1ab |
AutoSlim-MobileNet v2 | 24.6 25.8 27.0 |
505M 305M 207M |
a24f1f2 |
AutoSlim-MNasNet | 24.6 25.4 26.8 |
532M 315M 217M |
31477c9 |
AutoSlim-ResNet-50 | 24.0 24.4 26.0 27.8 |
3.0G 2.0G 1.0G 570M |
f95f419 |
Implementing slimmable networks and slimmable training is straightforward:
models/slimmable_ops
.train.py
.CC 4.0 Attribution-NonCommercial International
The software is for educaitonal and academic research purpose only.