In this repository, we apply function combinations in low-dimensional data to design Kolmogorov-Arnold Networks, referred to as FC-KAN (Function Combinations in Kolmogorov-Arnold Networks). The experiments demonstrate that these combinations improve the model performance.
Our paper, "FC-KAN: Function Combinations in Kolmogorov-Arnold Networks," is available at https://arxiv.org/abs/2409.01763. The paper contains several errors (equations), but we will update the content soon.
We can use some element-wise operations to combine the functions' outputs by different methods.
def forward(self, x: torch.Tensor):
#x = self.drop(x)
#device = x.device
if (len(self.func_list) == 1):
raise Exception('The number of functions (func_list) must be larger than 1.')
X = torch.stack([x] * len(self.func_list)) # size (number of functions, batch_size, input_dim)
for layer in self.layers:
X = layer(X)
output = X.detach().clone()
if (self.combined_type == 'sum'): output = torch.sum(X, dim=0)
elif (self.combined_type == 'product'): output = torch.prod(X, dim=0)
elif (self.combined_type == 'sum_product'): output = torch.sum(X, dim=0) + torch.prod(X, dim=0)
elif (self.combined_type == 'quadratic'):
output = torch.sum(X, dim=0) + torch.prod(X, dim=0)
for i in range(X.shape[0]):
output = output + X[i, :, :].squeeze(0)*X[i, :, :].squeeze(0)
elif (self.combined_type == 'quadratic2'): # not better than "quadratic"
output = torch.prod(X, dim=0)
for i in range(X.shape[0]):
output = output + X[i, :, :].squeeze(0)*X[i, :, :].squeeze(0)
elif (self.combined_type == 'cubic'): # not good
outsum = torch.sum(X, dim=0)
output = outsum + torch.prod(X, dim=0)
for i in range(X.shape[0]):
output = output + X[i, :, :].squeeze(0)*X[i, :, :].squeeze(0)
output = output*outsum
elif (self.combined_type == 'concat'):
X_permuted = X.permute(1, 0, 2)
output = X_permuted.reshape(X_permuted.shape[0], -1)
elif (self.combined_type == 'concat_linear'):
X_permuted = X.permute(1, 0, 2)
output = X_permuted.reshape(X_permuted.shape[0], -1)
output = F.linear(output, self.concat_weight)
elif (self.combined_type == 'max'):
output, _ = torch.max(X, dim=0)
elif (self.combined_type == 'min'):
output, _ = torch.min(X, dim=0)
elif (self.combined_type == 'mean'):
output = torch.mean(X, dim=0)
else:
raise Exception('The combined type "' + self.combined_type + '" does not support!')
# Write more combinations here...
#output = self.base_activation(output) # SiLU
#output = output + F.normalize(output, p=2, dim=1)
return output
- numpy==1.26.4
- numpyencoder==0.3.0
- torch==2.3.0+cu118
- torchvision==0.18.0+cu118
- tqdm==4.66.4
- mode: working mode ("train" or "test"). Note that we did not write the test() function. =))
- ds_name: dataset name ("mnist" or "fashion_mnist").
- model_name: type of models (bsrbf_kan, efficient_kan, fast_kan, faster_kan, mlp, and fc_kan).
- epochs: the number of epochs.
- batch_size: the training batch size (default: 64).
- n_input: The number of input neurons (default: 28^2 = 784).
- n_hidden: The number of hidden neurons. We use only 1 hidden layer. You can modify the code (run.py) for more layers.
- n_output: The number of output neurons (classes). For MNIST and Fashion-MNIST, there are 10 classes.
- grid_size: The size of grid (default: 5). Use with bsrbf_kan and efficient_kan.
- spline_order: The order of spline (default: 3). Use with bsrbf_kan and efficient_kan.
- num_grids: The number of grids, equals grid_size + spline_order (default: 8). Use with fast_kan and faster_kan.
- device: use "cuda" or "cpu" (default: "cuda").
- n_examples: the number of examples in the training set used for training (default: 0, mean use all training data)
- note: A note saved in the model name file.
- n_part: the part of data used to train data (default: 0, mean use all training data, 0.1 means 10%).
- func_list: the name of functions used in FC-KAN (default='dog,rbf'). Other functions are bs and base.
- combined_type: the type of data combination used in the output (default='quadratic', others are sum, product, sum_product, concat, max, min, mean). We are developing other combinations.
See run.sh or run_fc.sh (bash run.sh or bash run_fc.sh in BASH) for details. We trained the models on GeForce RTX 3060 Ti (with other default parameters). For example, FC-KAN models (Difference of Gaussians + B-splines) can be trained on MNIST with different output combinations.
python run.py --mode "train" --model_name "fc_kan" --epochs 25 --batch_size 64 --n_input 784 --n_hidden 64 --n_output 10 --ds_name "mnist" --note "full_0" --n_part 0 --func_list "dog,bs" --combined_type "sum"
python run.py --mode "train" --model_name "fc_kan" --epochs 25 --batch_size 64 --n_input 784 --n_hidden 64 --n_output 10 --ds_name "mnist" --note "full_0" --n_part 0 --func_list "dog,bs" --combined_type "product"
python run.py --mode "train" --model_name "fc_kan" --epochs 25 --batch_size 64 --n_input 784 --n_hidden 64 --n_output 10 --ds_name "mnist" --note "full_0" --n_part 0 --func_list "dog,bs" --combined_type "sum_product"
python run.py --mode "train" --model_name "fc_kan" --epochs 25 --batch_size 64 --n_input 784 --n_hidden 64 --n_output 10 --ds_name "mnist" --note "full_0" --n_part 0 --func_list "dog,bs" --combined_type "quadratic"
python run.py --mode "train" --model_name "fc_kan" --epochs 25 --batch_size 64 --n_input 784 --n_hidden 64 --n_output 10 --ds_name "mnist" --note "full_0" --n_part 0 --func_list "dog,bs" --combined_type "concat"
- https://github.com/hoangthangta/BSRBF_KAN
- https://github.com/Blealtan/efficient-kan
- https://github.com/AthanasiosDelis/faster-kan
- https://github.com/ZiyaoLi/fast-kan/
- https://github.com/zavareh1/Wav-KAN
- https://github.com/seydi1370/Basis_Functions
- https://github.com/KindXiaoming/pykan (the original KAN)
Cite our work if this paper is helpful for you.
@misc{ta2024fckan,
title={FC-KAN: Function Combinations in Kolmogorov-Arnold Networks},
author={Hoang-Thang Ta and Duy-Quy Thai and Abu Bakar Siddiqur Rahman and Grigori Sidorov and Alexander Gelbukh},
year={2024},
eprint={2409.01763},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
Also, give me a star if you like this repo. Thanks!
If you have any questions, please contact: tahoangthang@gmail.com. If you want to know more about me, please visit website: https://tahoangthang.com.