Crossover implementation

Crossover implementation#

class DAGTwoPoint(size=10)[source]

Bases: object

Implementation of a two-point crossover on an array. A DAG-based crossover is used if some of the exchanged values are DAGs.

Parameters:

size (int, default=10) – Maximum size a DAG may have.

Examples

>>> from dragon.search_operators.crossover import DAGTwoPoint
>>> from dragon.search_space.base_variables import Constant, FloatVar, IntVar, CatVar,  ArrayVar
>>> arr = ArrayVar(IntVar("Number of features", 1, 10), CatVar("Optimizer", features=['Adam', 'SGD', 'AdamMax']), FloatVar('learning rate', 0.001, 0.5), Constant('Seed', value=0))
>>> parent_1 = arr.random()
>>> parent_2 = arr.random()
>>> crossover = DAGTwoPoint(arr)
>>> print(parent_1)
[5, 'AdamMax', 0.16718361674068502, 0]
>>> print(parent_2)
[9, 'SGD', 0.28364322926906005, 0]
>>> print(crossover(parent_1, parent_2))
([5, 'AdamMax', 0.28364322926906005, 0], [9, 'SGD', 0.16718361674068502, 0])

>>> from dragon.search_space.dag_variables import HpVar, NodeVariable, EvoDagVariable
>>> from dragon.search_space.bricks import MLP, MaxPooling1D, AVGPooling1D
>>> from dragon.search_space.base_variables import DynamicBlock
>>> from dragon.search_space.bricks_variables import activation_var
>>> mlp = HpVar("Operation", Constant("MLP operation", MLP), hyperparameters={"out_channels": IntVar("out_channels", 1, 10)})
>>> pooling = HpVar("Operation", CatVar("Pooling operation", [MaxPooling1D, AVGPooling1D]), hyperparameters={"pool_size": IntVar("pool_size", 1, 5)})
>>> candidates = NodeVariable(label = "Candidates",
...                           combiner=CatVar("Combiner", features=['add', 'concat']),
...                           operation=CatVar("Candidates", [mlp, pooling]),
...                           activation_function=activation_var("Activation"))
>>> operations = DynamicBlock("Operations", candidates, repeat=3)
>>> dag = EvoDagVariable(label="DAG", operations=operations)
>>> arr = ArrayVar(dag, dag, Constant('Seed', value=0))
>>> parent_1 = arr.random()
>>> parent_2 = arr.random()
>>> crossover = DAGTwoPoint(arr)
>>> print(parent_1)
[NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Identity() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.pooling.AVGPooling1D'> -- (hp) {'pool_size': 3} -- (activation) ELU(alpha=1.0) -- ] | MATRIX:[[0, 1], [0, 0]], NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Identity() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.basics.MLP'> -- (hp) {'out_channels': 10} -- (activation) SiLU() -- ,
(combiner) add -- (name) <class 'dragon.search_space.bricks.pooling.MaxPooling1D'> -- (hp) {'pool_size': 4} -- (activation) Sigmoid() -- ] | MATRIX:[[0, 1, 1], [0, 0, 1], [0, 0, 0]], 0]
>>> print(parent_2)
[NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Identity() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.pooling.AVGPooling1D'> -- (hp) {'pool_size': 1} -- (activation) LeakyReLU(negative_slope=0.01) -- ] | MATRIX:[[0, 1], [0, 0]], NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Identity() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.basics.MLP'> -- (hp) {'out_channels': 9} -- (activation) ELU(alpha=1.0) -- ] | MATRIX:[[0, 1], [0, 0]], 0]
>>> print(crossover(parent_1, parent_2))
([NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Identity() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.pooling.AVGPooling1D'> -- (hp) {'pool_size': 3} -- (activation) ELU(alpha=1.0) -- ] | MATRIX:[[0, 1], [0, 0]], NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Identity() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.basics.MLP'> -- (hp) {'out_channels': 10} -- (activation) SiLU() -- ,
(combiner) add -- (name) <class 'dragon.search_space.bricks.pooling.MaxPooling1D'> -- (hp) {'pool_size': 4} -- (activation) Sigmoid() -- ] | MATRIX:[[0, 1, 1], [0, 0, 1], [0, 0, 0]], 0], [NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Identity() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.pooling.AVGPooling1D'> -- (hp) {'pool_size': 1} -- (activation) LeakyReLU(negative_slope=0.01) -- ] | MATRIX:[[0, 1], [0, 0]], NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Identity() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.basics.MLP'> -- (hp) {'out_channels': 9} -- (activation) ELU(alpha=1.0) -- ] | MATRIX:[[0, 1], [0, 0]], 0])
adj_matrix_crossover(p1, p2)[source]
DAG-based crossover

  • Select the index of the operations that would be exchange in each graphs.

  • Remove the corresponding lines and columns from the adjacency matrices

  • Compute the index where the new nodes will be inserted

  • Insert the new rows and columns within the adjacency matrices

  • Make sure no nodes without incoming or outgoing connections are remaining within the matrices

  • Make sure the new matrices are upper-triangular

  • Recreate the list of nodes

  • Create new AdjMatrix variables with the new nodes and matrices

Parameters:

  • p1 (AdjMatrix) – First parent

  • p2 (`AdjMatrix) – Second parent

  • size (int, default=10) – Maximum number of nodes that an offspring graph could have

Returns:

  • f1 (AdjMatrix) – First offspring

  • f2 (AdjMatrix) – Second offspring

Examples

>>> import numpy as np
>>> from dragon.search_space.dag_variables import AdjMatrix
>>> import torch.nn as nn
>>> from dragon.search_space.bricks import MLP, Identity
>>> from dragon.search_space.dag_variables import Node
>>> node_1 = Node(combiner="add", operation=MLP, hp={"out_channels": 10}, activation=nn.ReLU())
>>> node_2 = Node(combiner="add", operation=MLP, hp={"out_channels": 5}, activation=nn.ReLU())
>>> node_3 = Node(combiner="concat", operation=Identity, hp={}, activation=nn.Softmax())
>>> op1 = [node_1, node_2, node_3]
>>> op2 = [node_1, node_3]
>>> m1 = np.array([[0, 1, 1],
...                 [0, 0, 1],
...                 [0, 0, 0]])
>>> m2 = np.array([[0, 1],
...                 [0, 0]])
>>> print(AdjMatrix(op1, m1))
NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.MLP'> -- (hp) {'out_channels': 10} -- (activation) ReLU() -- ,
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.MLP'> -- (hp) {'out_channels': 5} -- (activation) ReLU() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Softmax(dim=None) -- ] | MATRIX:[[0, 1, 1], [0, 0, 1], [0, 0, 0]]
>>> print(AdjMatrix(op2, m2))
NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.MLP'> -- (hp) {'out_channels': 10} -- (activation) ReLU() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Softmax(dim=None) -- ] | MATRIX:[[0, 1], [0, 0]]
>>> f1, f2 = adj_matrix_crossover(AdjMatrix(op1, m1), AdjMatrix(op2, m2))
>>> print(f1)
NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.MLP'> -- (hp) {'out_channels': 10} -- (activation) ReLU() -- ,
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.MLP'> -- (hp) {'out_channels': 5} -- (activation) ReLU() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Softmax(dim=None) -- ] | MATRIX:[[0, 1, 0], [0, 0, 1], [0, 0, 0]]
>>> print(f2)
NODES: [
(combiner) add -- (name) <class 'dragon.search_space.bricks.basics.MLP'> -- (hp) {'out_channels': 10} -- (activation) ReLU() -- ,
(combiner) concat -- (name) <class 'dragon.search_space.bricks.basics.Identity'> -- (hp) {} -- (activation) Softmax(dim=None) -- ] | MATRIX:[[0, 1], [0, 0]]