By Yi Li, Zhanghui Kuang, Yimin Chen, Wayne Zhang

SenseTime.

0. Introduction
1. Citation
2. Approach
3. Trained models
4. Experiments
5. GPU time

This is a PyTorch implementation of Data-Driven Neuron Allocation for Scale Aggregation Networks.(CVPR2019) with pretrained models.

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}
}

Figure 1: architecture of ScaleNet-50.

Figure 2: scale aggregation block.

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.

Figure 3: experiments on imagenet classification.

Figure 4: experiments on ms-coco detection.

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.