The implement code of UPANets: Learning from the Universal Pixel Attention Networks
Paper is in here.
Among image classification, skip and densely-connection-based networks have dominated most leaderboards. Recently, from the successful development of multi-head attention in natural language processing, it is sure that now is a time of either using a Transformer-like model or hybrid CNNs with attention. However, the former need a tremendous resource to train, and the latter is in the perfect balance in this direction. In this work, to make CNNs handle global and local information, we proposed UPANets, which equips channel-wise attention with a hybrid skip-densely-connection structure. Also, the extreme-connection structure makes UPANets robust with a smoother loss landscape. In experiments, UPANets surpassed most well-known and widely-used SOTAs with an accuracy of 96.47% in Cifar-10, 80.29% in Cifar-100, and 67.67% in Tiny Imagenet. Most importantly, these performances have high parameters efficiency and only trained in one customer-based GPU. We share implementing code of UPANets in here.
This code was referring another project in here.
UPANets has surpassed a series of SOTAs by reimplementing in the same enviroment and experiment setting. The performance in Cifar-10 and Cifar-100 can be seen in the following plots:
where accuracy is in 1 - top-1 error, and efficiency follows a equation in Acc/Params.
There is a pre-defined package throught pipenv in ./Pipfile. Please install the needed packages in:
$ pipenv install
In this step, the pipenv will automatically get the Pipfile.lock and install the needed parameter or packages as the Pipfile content.
In this project, we implement UPANets in ./models/upanets.py. The simple usage is:
python main.py
If you install the needed packages through pipenv, please use:
pipenv run python main.py
The operation can be simply embedded with CNNs.
Apart from the place in the above picture. A further data flow explanation can be seen in ./models/upanets,.py's CPA function. And, a demonstration of setting sc_x is under blew:
*same=False:
This scenario can be easily embedded after any CNNs, if size is same.
x (OG) ---------------
| |
sc_x (from CNNs) CPA(x)
| |
out + <---------------
*same=True:
This can be embedded after the CNNs where the size are different.
x (OG) ---------------
| |
sc_x (from CNNs) |
| CPA(x)
CPA(sc_x) |
| |
out + <---------------
*sc_x=False
This operation can be seen a channel embedding with CPA
EX: x (3, 32, 32) => (16, 32, 32)
x (OG)
|
CPA(x)
|
out
Please feel free to add it after any CNNs.
Be aware if sc_x=False, it is not matter what is the value of input sc_x as it will be ignored eventually. For example:
def __init___:
...
cpa = CPA(16, 32, sc_x=False)
...
def forward(x):
out = CNN(x)
out = cpa(out, out) # <--- it is acceptable
- Please don't expect that applying CPA as an embedding layer will increase the performance surely but it could offer another option of embedding with global information considering. Further researches about whether placing an embed-CPA in a propoer place to raise performance are needed