Module degann.search_algorithms.nn_code
Expand source code
from typing import Tuple, List
from degann.networks import imodel
act_to_hex = {
"elu": "0",
"relu": "1",
"gelu": "2",
"selu": "3",
"exponential": "4",
"linear": "5",
"sigmoid": "6",
"hard_sigmoid": "7",
"swish": "8",
"tanh": "9",
"softplus": "a",
"softsign": "b",
"parabolic": "c",
}
hex_to_act = {v: k for k, v in act_to_hex.items()}
alph_n_full = "0123456789abcdef"
alph_n_div3 = "0369cf"
alph_n_div2 = "02468ace"
alph_n_div4 = "048c"
alphabet_activations_cut = "0689"
alphabet_activations = "0123456789abc"
def encode(nn: imodel.IModel, offset: int = None) -> str:
"""
Encode neural network to string
Parameters
----------
nn: IModel
Neural network for coding
offset: int
Minimum possible layer size
Returns
-------
code: str
Encoded neural network
"""
blocks = nn.get_shape
activations = nn.get_activations
res = ""
offset = min(blocks) if offset is None else offset
for layer, act in zip(blocks, activations):
curr = hex(layer - offset)[2:] + act_to_hex[act]
res += curr
return res
def decode(s: str, block_size: int = 1, offset: int = 0) -> Tuple[List[int], List[str]]:
"""
Decode neural network from string as a pair of shape and activations
Parameters
----------
s: str
Neural network as code
block_size: int
Number of letters allocated to encode the size of one layer
offset: int
Minimum possible layer size
Returns
-------
network: tuple[list[int], list[str]]
Pair of shape and activations for neural network
"""
blocks = []
activations = []
for block in range(0, len(s), block_size + 1):
temp = s[block : block + block_size]
blocks.append(int(temp, 16) + offset)
activations.append(hex_to_act[s[block + block_size]])
return blocks, activationsFunctions
def decode(s: str, block_size: int = 1, offset: int = 0) ‑> Tuple[List[int], List[str]]-
Decode neural network from string as a pair of shape and activations
Parameters
s:str- Neural network as code
block_size:int- Number of letters allocated to encode the size of one layer
offset:int- Minimum possible layer size
Returns
network:tuple[list[int], list[str]]- Pair of shape and activations for neural network
Expand source code
def decode(s: str, block_size: int = 1, offset: int = 0) -> Tuple[List[int], List[str]]: """ Decode neural network from string as a pair of shape and activations Parameters ---------- s: str Neural network as code block_size: int Number of letters allocated to encode the size of one layer offset: int Minimum possible layer size Returns ------- network: tuple[list[int], list[str]] Pair of shape and activations for neural network """ blocks = [] activations = [] for block in range(0, len(s), block_size + 1): temp = s[block : block + block_size] blocks.append(int(temp, 16) + offset) activations.append(hex_to_act[s[block + block_size]]) return blocks, activations def encode(nn: IModel, offset: int = None) ‑> str-
Encode neural network to string
Parameters
nn:IModel- Neural network for coding
offset:int- Minimum possible layer size
Returns
code:str- Encoded neural network
Expand source code
def encode(nn: imodel.IModel, offset: int = None) -> str: """ Encode neural network to string Parameters ---------- nn: IModel Neural network for coding offset: int Minimum possible layer size Returns ------- code: str Encoded neural network """ blocks = nn.get_shape activations = nn.get_activations res = "" offset = min(blocks) if offset is None else offset for layer, act in zip(blocks, activations): curr = hex(layer - offset)[2:] + act_to_hex[act] res += curr return res