microNet

microNet Recurrence and Adaption
mostly based on liyunsheng13/micronet

test the performance of the model

you may code like this

import torch
from micronet import *
from torchsummaryX import summary

model = micronet('m1')
dummy_input = torch.rand(32,3,224,224)
profile = summary(model, dummy_input)

or

python3 micronet.py

train on imagenet

using m1 as demo

python3 -W ignore main.py --arch m1 -d ./imagenet -c ./output/m1

test on imagenet

using m1 as demo

python3 -W ignore main.py --arch m1 -d ./imagenet -c ./output/m1_test -e --weight ./output/m1/model_best.pth

ImageNet format

imagenet/
│   ├── train/
│   │   └──n********/ * n
│   │       └──*.JPEG * n
│   │
│   ├── val/
│   │   └──n********/ * n
│   │       └──*.JPEG * n

If the foldername is not 'train' and 'val', like 'ILSVRC2012_img_train' and 'ILSVRC2012_img_val',
please try to rewrite the str 'train' and 'val' in the .py file utils/dataloaders.py in line 95 and line 115

2021/10/29

Update the micronet.py, using the block_a, block_b, block_c to build the model as it in the paper.

2021/11/02

Our research shows that using dali-cpu may not help increasing the speed,
meanwhile using dali-gpu may run out of memory.
So updata the utils/dataloaders.py, using pytorch_only and cut the 'if' branch in main.py.

2021/11/03

Adapt the structure of part of the code,
it seems to be easier to read and adapt.

2021/11/04

Pull the micronet out and change some default args.
Now it can be tested singly,
The performance_test code was shown above.

2021/11/05

Update the DyShiftMax activation,
add lambdas and alphas on it.
the acc might rises but the mflops rise either.
for m1 as example, the Mult-Adds rises from 5.244M to 5.688M.

Ready to put the model into training :)

2021/11/06

Fix some bugs encountered when we train on distributed parallel with the updated version.

2021/11/07

Results:

training on imagenet_t:

Model	Param	MAdds	Acc@1	Acc@5
microNet_m1	1.700M	5.244M	55.385	78.714
microNet_m1_updated	1.700M	5.688M	57.648	80.330
ghostNet_width=0.1	1.442M	6.911M	51.407	78.538

2021/11/30

the loss of rvm_micronet is not good as expect,
so I try to rewrite the backbone of micronet.
add ghost_module, SE_module and so on...
the mflops even lighter than usual...
put it on the coeus later, hope it performs well.

Citation

@article{li2021micronet,
   author    = {Li, Yunsheng and Chen, Yinpeng and Dai, Xiyang and Chen, Dongdong and Liu, Mengchen and Yuan, Lu and Liu, Zicheng and Zhang, Lei and Vasconcelos, Nuno},
   title     = {MicroNet: Improving Image Recognition with Extremely Low FLOPs},
   journal   = {arXiv preprint arXiv:2108.05894},
   year      = {2021},
}