ScaleNet
By Yi Li, Zhanghui Kuang, Yimin Chen, Wayne Zhang
SenseTime.
Table of Contents
Introduction
This is a PyTorch implementation of Data-Driven Neuron Allocation for Scale Aggregation Networks.(CVPR2019) with pretrained models.
Citation
If you use these models in your research, please cite:
@inproceedings{Li2019ScaleNet,
title={Data-Driven Neuron Allocation for Scale Aggregation Networks},
author={Li, Yi and Kuang, Zhanghui and Chen, Yimin and Zhang, Wayne},
booktitle={CVPR},
year={2019}
}
Approach
Figure 1: architecture of ScaleNet-50.
Figure 2: scale aggregation block.
Trained models
| Model | Top-1 err. | Top-5 err. |
|---|---|---|
| ScaleNet-50-light | 22.80 | 6.57 |
| ScaleNet-50 | 22.02 | 6.05 |
| ScaleNet-101 | 20.82 | 5.42 |
| ScaleNet-152 | 20.06 | 5.18 |
Pytorch:
from pytorch.scalenet import *
model = scalenet50(structure_path='structures/scalenet50.json', ckpt=None) # train from stratch
model = scalenet50(structure_path='structures/scalenet50.json', ckpt='weights/scalenet50.pth') # load pretrained model
The weights are available on BaiduYun with extract code: f1c5
Unlike the paper, we used better training settings: increase the epochs to 120 and replace multi-step learning rate by cosine learning rate.
Experiments
Figure 3: experiments on imagenet classification.
Figure 4: experiments on ms-coco detection.
GPU time
| Model | Top-1 err. | FLOPs(10^9) | GPU time(ms) |
|---|---|---|---|
| ResNet-50 | 24.02 | 4.1 | 95 |
| SE-ResNet-50 | 23.29 | 4.1 | 98 |
| ResNeXt-50 | 22.2 | 4.2 | 147 |
| ScaleNet-50 | 22.2 | 3.8 | 93 |
TensorFlow: (empty models of ResNet, SE-ResNet, ResNeXt, ScaleNet for speed test)
python3 tensorflow/test_speed.py scale|res|se|next
All networks were tested using Tensorflow with GTX 1060 GPU and i7 CPU at batch size 16 and image side 224 on 1000 runs.
Some static-graph frameworks like Tensorflow and TensorRT execute multi-branch models in parallel, while Pytorch and Caffe do not. So we suggest to deploy ScaleNets on Tensorflow and TensorRT.