rusticsom

neighbourhood_function, optional, is also a function pointer that accepts functions that take 3 parameters, a tuple of type (usize, usize) representing the size of the SOM, another tuple of type (usize, usize) representing the position of the winner neuron, and an f32 representing sigma; and returns a 2D Array containing weights of the neighbours of the winning neuron, i.e, centered at winner. By default, the Gaussian function will be used, which returns a "Gaussian centered at the winner neuron".

    pub fn from_json(serialized: &str,  decay_function: Option<fn(f32, u32, u32) -> f64>, neighbourhood_function: Option<fn((usize, usize), (usize, usize), f32) -> Array2<f64>>) -> serde_json::Result<SOM> { ... }

This function allows to create a SOM from a previously exported SOM json data using SOM::to_json().

Use SOM_Object.train_random() to train the SOM with the input dataset, where samples from the input dataset are picked in a random order.

pub fn train_random(&mut self, data: Array2<f64>, iterations: u32) { ... }

Samples (rows) from the 2D Array data are picked randomly and the SOM is trained for iterations iterations!

Use SOM_Object.train_batch() to train the SOM with the input dataset, where samples from the input dataset are picked in a sequential order.

pub fn train_batch(&mut self, data: Array2<f64>, iterations: u32) { ... }

Samples (rows) from the 2D Array data are picked sequentially and the SOM is trained for iterations iterations!

Use SOM_Object.winner() to find the winning neuron for a given sample.

pub fn winner(&mut self, elem: Array1<f64>) -> (usize, usize) { ... }

This function must be called with an SOM object.

Requires one parameter, a 1D Array of f64s representing the input sample.

Returns a tuple (usize, usize) representing the x and y coordinates of the winning neuron in the SOM.

Use SOM_Object.winner_dist() to find the winning neuron for a given sample, and it's distance from this winner neuron.

pub fn winner_dist(&mut self, elem: Array1<f64>) -> ((usize, usize), f64) { ... }

This function must be called with an SOM object.

Requires one parameter, a 1D Array of f64s representing the input sample.

Returns a tuple (usize, usize) representing the x and y coordinates of the winning neuron in the SOM.

Also returns an f64 representing the distance of the input sample from this winner neuron.

pub fn activation_response(&self) -> ArrayView2<usize> { ... }

This function returns the activation map of the SOM. The activation map is a 2D Array where each cell at (i, j) represents the number of times the (i, j) cell of the SOM was picked to be the winner neuron.

pub fn get_size(&self) -> (usize, usize)

This function returns a tuple representing the size of the SOM. Format is (length, breadth).

pub fn distance_map(self) -> Array2<f64> { ... }

Returns the distance map of the SOM, i.e, the normalized distance of every neuron with every other neuron.

pub fn to_json(&self) -> serde_json::Result<String> { ... }

Returns the internal SOM data as pretty printed json (using serde_json).

Primary Contributors

Example

We've tested this crate on the famous iris dataset (present in csv format in the extras folder).

The t_full_test function in /tests/test.rs was used to produce the required output. The following plots were obtained using matplotlib for Python.

Using a 5 x 5 SOM, trained for 250 iterations :

Using a 10 x 10 SOM, trained for 1000 iterations :