Kolmogorov–Arnold Networks

https://github.com/zhangyanbo/mlp-kan
Admin
Kolmogorov–Arnold Networks with modified activation (using MLP to represent the activation)
Tokens:664
Snippets:6
Trust Score:8.8
License:MIT
Update:1 year ago
Show doc for...
Tokens:
Raw
### Define Interpretable MLP-KAN Model Architecture

Source: https://github.com/zhangyanbo/mlp-kan/blob/main/README.md

Configures a specific MLP-KAN model architecture with reduced layers for enhanced interpretability. This setup is used to demonstrate the network's ability to learn target functions precisely.

```python
dims = [2, 1, 1]
model = nn.Sequential(
    KANLayer(dims[0], dims[1]),
    KANLayer(dims[1], dims[2])
)
```

--------------------------------

### Run Experiment Script

Source: https://github.com/zhangyanbo/mlp-kan/blob/main/README.md

Executes the main experiment script for the MLP-KAN project. This is a command-line operation and does not require specific inputs or outputs beyond the script's execution.

```bash
python experiment.py
```

--------------------------------

### Define and Use KANInterpoLayer Model in PyTorch

Source: https://github.com/zhangyanbo/mlp-kan/blob/main/README.md

Illustrates the usage of KANInterpoLayer for creating sequential models in PyTorch. Similar to KANLayer, it defines a model, generates random input, and performs a forward pass. The description notes potential training difficulties.

```python
from kan_layer import KANInterpoLayer

model = nn.Sequential(
        KANInterpoLayer(2, 5),
        KANInterpoLayer(5, 1)
    )

x = torch.randn(16, 2)
y = model(x)
# y.shape = (16, 1)
```

--------------------------------

### Define and Use KANLayer Model in PyTorch

Source: https://github.com/zhangyanbo/mlp-kan/blob/main/README.md

Demonstrates how to define a simple sequential model using KANLayer from the kan_layer module in PyTorch. It shows model instantiation, input data generation, and forward pass computation. The expected output shape is also noted.

```python
from kan_layer import KANLayer

model = nn.Sequential(
        KANLayer(2, 5),
        KANLayer(5, 1)
    )

x = torch.randn(16, 2)
y = model(x)
# y.shape = (16, 1)
```

--------------------------------

### Define MLP-KAN Model Architecture in PyTorch

Source: https://github.com/zhangyanbo/mlp-kan/blob/main/README.md

Sets up a sequential model architecture using KANLayer with specified input and output dimensions. This code snippet is used to define the network structure before training and visualization.

```python
dims = [2, 5, 1]
model = nn.Sequential(
    KANLayer(dims[0], dims[1]),
    KANLayer(dims[1], dims[2])
)
```

--------------------------------

### Define Target Function for Visualization

Source: https://github.com/zhangyanbo/mlp-kan/blob/main/README.md

Defines a Python function `target_fn` that computes the objective function f(x,y)=exp(sin(pi * x) + y^2). It handles both single-input (scalar) and multi-input (tensor) cases, returning a PyTorch tensor.

```python
def target_fn(input):
    # f(x,y)=exp(sin(pi * x) + y^2)
    if len(input.shape) == 1:
        x, y = input
    else:
        x, y = input[:, 0], input[:, 1]
    return torch.exp(torch.sin(torch.pi * x) + y**2)
```
Tokens:
Raw