This is an unofficial PyTorch implementation for MobileNetV3. Multi-GPUs training is supported. We trained it on ImageNet-1K and released the model parameters. This work was implemented by Peng Xu and Jin Feng.
This project is designed with these goals:
- Train MobileNetV3-Small 1.0 on ImageNet-1K dataset.
- Train MobileNetV3-Large 1.0 on ImageNet-1K dataset.
Ubuntu 14.04
Python 2.7
PyTorch 0.4.0
Our server details: 2 Intel(R) Xeon(R) CPUs (E5-2620 v3 @ 2.40GHz), 128 GB RAM, 4 GTX 1080 Ti GPUs.
For fast IO, ImageNet-1K dataset is stored in our SSD.
We report the performance (Top-1 accuracy) on ImageNet-1K validation set.
| Network | Top-1 Accuracy | Pretrained Model |
|---|---|---|
| MobileNetV3-Small 1.0 (Official Implementation) | 67.4% | none |
| MobileNetV3-Small 1.0 (Our Implementation) | soon | 9.8 MB, Baiduyun Disk, [Google Drive] (coming soon) |
Optimizer
Learning rate
Batch Size = 2560
Running costs are summarized in following table.
| GPU RAM | RAM | Running Time |
|---|---|---|
| 30 GB | 100 GB | 72 hours |
Please see more details in our training file train.py.
from MobileNetV2 import MobileNetV2
net = MobileNetV2(model_mode="SMALL", num_classes=1000)
state_dict = torch.load('xx')["network"]
net.load_state_dict(state_dict)
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent