mask.MultiIO

Description

This is to support the senario when we need to project different inputs to different areas and readout from different areas. This will generate a multi-area RNN. This class primarily serves to generate the input and readout masks.

Parameters

• input_dims (list, default: [dims[0]]):
  A list denoting the dimensions of each group of input signals.
  • Example: For a 1-dimensional olfactory signal and a 100-dimensional visual signal, use [1, 100].
  • Must sum to the total input dimension specified in dims[0].
• readout_dims (list, default: [dims[2]]):
  A list denoting the dimensions of each group of readout signals.
  • Example: For a 1-dimensional olfactory signal and a 100-dimensional visual signal, use [1, 100].
  • Must sum to the total readout dimension specified in dims[2].
• input_table (np.ndarray, default: all ones):
  A table specifying whether an input group is projected to a hidden layer node.
  • Shape: (n_input_groups, hidden_size).
  • Values: 0 (no connection) or 1 (connection).
• readout_table (np.ndarray, default: all ones):
  A table specifying whether a hidden layer node contributes to a readout group.
  • Shape: (n_readout_groups, hidden_size).
  • Values: 0 (no contribution) or 1 (contribution).

Methods

get_specs()

Returns the specifications of the network, including:

• "dims"
• "hidden_size"
• "input_dim"
• "readout_dim"
• "input_dims"
• "readout_dims"
• "input_table_shape"
• "readout_table_shape"

Returns:

• dict: Specifications of the network.

Usage:

specs = multi_io.get_specs()

Example

import numpy as np
from nn4n.mask import MultiIO

input_table = np.ones((2, 200))
readout_table = np.ones((2, 200))
input_table[0, 0:100] = 0
input_table[1, 100:150] = 0
readout_table[0, 100:200] = 0

mask_params = {
    "dims": [20, 200, 10],
    "input_dims": [5, 15],
    "readout_dims": [5, 5],
    "input_table": input_table,
    "readout_table": readout_table,
}

mask = MultiIO(**mask_params)
mask.print_specs()
mask.plot_masks()

Output:

This mask does not change the hidden layer connectivity. Skipped.

MultiIO Specs: 
   | dims:               [20, 200, 10]
   | hidden_size:        200
   | input_dim:          20
   | readout_dim:         10
   | input_dims:         [5, 15]
   | readout_dims:        [5, 5]
   | input_table_shape:  (2, 200)
   | readout_table_shape: (2, 200)