Implementing Zongyi Li's FNO on image classifcation which he used it through Pytorch.
The work doing in here is to transform Pytorch into Tensorflow.
Later I will show not only Image Classification case, but using pure Data-driven method to fit in the Navier Stokes Theorem and Burger's Equation.
The table shows different model's evaluation in MNIST classification. Though FNO has highest accuracy, it takes the longest time to compute. (Your results may be slighty different.)
Model | Train Accuracy | Train Loss | Test Accuracy | Test Loss |
---|---|---|---|---|
Fully connected network - After 5 Epochs | 0.8923 | 0.2935 | 0.8731 | 0.2432 |
Convolutional network - After 5 Epochs | 0.8860 | 0.3094 | 0.9048 | 0.1954 |
Residual network - After 5 Epochs | 0.9064 | 0.2610 | 0.8713 | 0.3398 |
Fourier Neural Operator - After 5 Epochs | 0.9486 | 0.1622 | 0.9455 | 0.1806 |
Citation
@code{Fourier Neural Operator with Tensorflow,
author = "Kozak Hou"
email = "kozak20010716@g.ncu.edu.tw"
Tel : +886-905804898
Affiliation = "Department of Space Science and Engineering, National Central University"
}