- Searching for MobileNetV3 paper
- Author: Andrew Howard(Google Research), Mark Sandler(Google Research, Grace Chu(Google Research), Liang-Chieh Chen(Google Research), Bo Chen(Google Research), Mingxing Tan(Google Brain), Weijun Wang(Google Research), Yukun Zhu(Google Research), Ruoming Pang(Google Brain), Vijay Vasudevan(Google Brain), Quoc V. Le(Google Brain), Hartwig Adam(Google Research)
-
The model.py comes from https://github.com/xiaolai-sqlai/mobilenetv3
-
We test both small and large mobileNetv3 on CIFAR-100 using our trained model, train code and test code. Our trained mode can be found from ./checkpoint
-
For the experiment on imageNet, we didn't train our own model, and we just use the contained pretrained model from the above link. We found that there is a accuracy difference between our result and their result. We got an acc1 of 65.07%, while they obatained 69.037%. This difference was also issued by someone else - xiaolai-sqlai/mobilenetv3#18
- For CIFAR-100 and imageNet data, we experimented with resize (224, 224).
| Datasets | Model | Accuracy | Epoch | Training Time | Parameters |
|---|---|---|---|---|---|
| CIFAR-100 | MobileNetV3-LARGE | 69.92% | 34 | 3h 58min | 2.8M |
| CIFAR-100 | MobileNetV3-SMALL | 68.67% | 35 | 1h 25min | 1.35M |
- For imageNet, we didn't train our own model. We just report the results from original paper, the pretrained model provider and ourselves(use the same pretrained model).
| Datasets | Model | Accuracy | Parameters |
|---|---|---|---|
| IMAGENET | MobileNetV3-LARGE(paper) | 75.2% | 5.4M |
| IMAGENET | MobileNetV3-SMALL(paper) | 67.4% | 2.9M |
| IMAGENET | MobileNetV3-SMALL(pretrained model provider) | 69.037% | 2.52M |
| IMAGENET | MobileNetV3-SMALL(Ours) | 65.07% | 2.52M |
- torch==1.1.0
- 32G Mem + 1GPU