Update Logs

2019-9-12 Add different up-/down-sampling methods to ScaleNet. 在ScaleNet上增加不同上采样/下采样方法的对比.
2019-9-19 Add a new CNN model named WaveNet, which combines Conv-DeConv couple with DenseNet architecture. ... ... 增加新模型 WaveNet, 将Conv-DeConv结构与DeseNet结构相结合.
2019-9-25 Add val-precision curves of Efficient-b0 & ScaleNet-vo21/-vo69 & ResNet-50 .... 增加EfficientNet-B0训练曲线的对比.
Todo: add a training schedule and data augmentation of EfficientNet.

一、ScaleNet Architecture

二、Requirements

Pytorch ≥ 0.4

TensorboardX

三、 How to Train ScaleNet

model architecture params are in folder: ./arch_params

model training configs are in folder: ./cfg_params

for example, when trainining ImageNet, net1 = 'vo21 || vo69 || vo72 || vo76',
check './arch_params/scalenet_imagenet_params.py'

data root, checkpoints path, tensorboard-logs-path are in xxxmodel_xxxdata_cfg.py

Train by Terminal Commands

Entry function is 'run_main.py'.

It can be used as following:

cd /to/your/project/root

# 
python run_main.py -name 'scalenet' -arch 'net1' -cfg 'cfgnet1' -exp 'exp.net1' -gpu 1 3

# also can run by nohup to save logs into a file
nohup python run_main.py -name 'scalenet' -arch 'net1' -cfg 'cfgnet1' -exp 'exp.net1' -gpu 1 3 1>printlog/net1.out 2>&1 &

Train by Pycharm Client

Find the following lines in 'run_main.py', and remove the comments on these lines. Then run this file in Pycharm.

args.arch_name = 'scalenet'
args.arch_list = ['net1']
args.cfg_dict = 'cfgnet1'
args.exp_version = 'exp.net1'
args.gpu_ids = [0, 1, 2, 3, 5, 6]
print('\n=> Your Args is :', args, '\n')

Validation curves of ScaleNet & EfficientNet & ResNet

ResNet: ./xmodels.tvm_resnet.py ， which is forked from pytorch-official

EfficentNet: ./xmodels/efficientnet.py ， which is forked from https://github.com/lukemelas/EfficientNet-PyTorch

model	Input	Layers	Params	FLOPs	Stages	Top1-Accuracy	GPU-time
EfficientNet-B0	224	82	5.29M	0.39G	5	68.69%	0.01082s
ScaleNet-vo69	224	103	5.08M	4.77G	1	71.34%	0.01567s
ResNet-50	224	54	25.56M	4.11G	4	76.36%	0.00778s
ScaleNet-vo21	224	54	25.02M	4.64G	4	74.62%	0.00799s

四、CAM Camparision of ScaleNet & DenseNet

`top-row: scalenet , bottom-row:densenet`

`top-(WhiteFont): scalenet , bottom-(CyanFont):densenet`

五、Pre-trained Models on ImageNet

BaiduYunPan: https://pan.baidu.com/s/1EbLnt0X-nIndwRlh6zNN8Q key：9qpg

GoogleDrive: uploading ... coming soon!

六、Other Features

Unified Training Framework for Classification.
Unified Data Factory, including CIFAR, IMAGENET, SVHN, LSV etc..
Unified Model Factory, including Pytorch-official models and New Models in 2019.
Including New Models in 2019: ScaleNet, EfficientNet, MobieNet-V3, HighResolutionNet etc..

七、Paper Citation

Paper and Citation can be downloaded here:

https://ieeexplore.ieee.org/document/8863492

GuideWsp / scalenet