This implement is an improved version of real-valued capsule network from our paper《Cv-CapsNet:complex-valued capsule network》 URL:https://ieeexplore.ieee.org/document/8744220， In this implement, a multi-scale feature fusion mechanism based on attention is proposed, named adaptive-diverse-model, which eliminates the manual setting of capsule size in coding stage, we also use the bottleneck from the MobilenetV3 to improve it(adaptive-diverse-model+). results of ablation research can be seen in Results.
Step 1. Clone this repository to local.
git clone https://github.com/Johnnan002/Adaptive-diverse-capsule-network
cd Adaptive-diverse-capsule-network
Step 2. Train the Adaptive-diverse-capsule-network model on CIFAR10
Training with default settings:
$ python Adaptive-diverse-capsule-network.py
More detailed usage run for help:
$ python Adaptive-diverse-capsule-network.py -h
Step 3. Test a pre-trained Adaptive-diverse-capsule-network model
Suppose you have trained a model using the above command, then the trained model will be
saved to result/trained_model.h5. Now just launch the following command to get test results.
$ python Adaptive-diverse-capsule-network.py -t -w result/trained_model.h5
It will output the testing accuracy . The testing data is same as the validation data. It will be easy to test on new data, just change the code as you want
Validation accuracy > 88.5% after 25 epochs on CIFAR10.
About 600 seconds per epoch on a single tesla k80 GPU card
________________________________________________________________________________________
| Models | Parameters | Accuracy(25epoch) | upgrade |
|———————————————————————|—————————————————————|————————————————————|—————————————————————|
| original model | 7.99M | 71.56% | —— —— |
|———————————————————————|—————————————————————|————————————————————|—————————————————————|
| diverse-model | 5.3M | 86.7% | ↑ 15.14% |
|———————————————————————|—————————————————————|————————————————————|—————————————————————|
| adaptive-diverse-model| 5.3M | 87.8% | ↑ 16.24% |
|———————————————————————|—————————————————————|————————————————————|—————————————————————|
|adaptive-diverse-model+| 5.3M | 88.5% | ↑ 16.94% |
|_______________________|_____________________|____________________|_____________________|
If you use the code in your research or wish to refer to the baseline results published in the Model , please use the following BibTeX entry.
@ARTICLE{8744220,
author={X. {Cheng} and J. {He} and J. {He} and H. {Xu}},
journal={IEEE Access},
title={Cv-CapsNet: Complex-Valued Capsule Network},
year={2019},
volume={7},
number={},
pages={85492-85499},
doi={10.1109/ACCESS.2019.2924548},
ISSN={2169-3536},
month={},}