3

Description


I am trying to implement a simple dataclass that just holds a few parameters.

@dataclass
class ReconstructionParameters:

    img_size: int
    CR: int
    denoise: bool
    epochs: int
    learning_rate: float
    step_size: int
    gamma: float
    batch_size: int
    regularization: float
    N0: float
    sig: float

    arch_name: InitVar[str]

    net_arch: int = field(init=False)


    def __post_init__(self, arch_name):
        self.net_arch = arch_name

    @property
    def net_arch(self):
        return str(netType(self.net_arch))

    @net_arch.setter
    def net_arch(self, arch_name):
        self.net_arch = int(netType[arch_name])
    

An user should be able to pass a string containing a string arch_name during initialization of the class, but the class should keep an equivalent integer defined in the following IntEnum:

class netType(IntEnum):
    
    c0mp = 0
    comp = 1
    pinv = 2
    free = 3

However, if the user wants obtain the net_arch contained in a previously created class, they should have access to the same string used for initialization instead of the integer representation. I am trying to use the @property decorator and a setter. Ideally, I would like to use a __post_init__() method to initialize net_arch that uses its setter.

Error


When running the following code I get an error:

a = ReconstructionParameters(
    img_size=64, 
    CR=1024,
    denoise=True,
    epochs=20, 
    learning_rate=1e-6,
    step_size=10,
    gamma=1e-7,
    batch_size=256,
    regularization=1e-6, 
    N0=2500,
    sig=0.5,
    arch_name='c0mp')
Traceback (most recent call last):

  File "C:\spas\Programs\Python\test.py", line 51, in <module>
    a = ReconstructionParameters(

  File "<string>", line 14, in __init__

  File "C:\spas\Programs\Python\test.py", line 48, in net_arch
    self.net_arch = int(netType[arch_name])

  File "C:\Users\user\Anaconda3\envs\singlepixelenv\lib\enum.py", line 349, in __getitem__
    return cls._member_map_[name]

KeyError: <property object at 0x000001CF7C26BA40>
Guilherme B.M.
  • 65
  • 1
  • 10

2 Answers2

2

The dataclass field and the property cannot have the same name. But you can add a leading underscore to the field, then the property will work.

from dataclasses import InitVar, dataclass, field
from enum import IntEnum


@dataclass
class ReconstructionParameters:
    img_size: int
    CR: int
    denoise: bool
    epochs: int
    learning_rate: float
    step_size: int
    gamma: float
    batch_size: int
    regularization: float
    N0: float
    sig: float

    arch_name: InitVar[str]

    _net_arch: int = field(init=False)

    def __post_init__(self, arch_name):
        self.net_arch = arch_name

    @property
    def net_arch(self):
        return str(netType(self._net_arch))

    @net_arch.setter
    def net_arch(self, arch_name):
        self._net_arch = int(netType[arch_name])


class netType(IntEnum):
    c0mp = 0
    comp = 1
    pinv = 2
    free = 3


a = ReconstructionParameters(
    img_size=64,
    CR=1024,
    denoise=True,
    epochs=20,
    learning_rate=1e-6,
    step_size=10,
    gamma=1e-7,
    batch_size=256,
    regularization=1e-6,
    N0=2500,
    sig=0.5,
    arch_name='c0mp')

print(a)
# ReconstructionParameters(img_size=64, CR=1024, denoise=True, epochs=20, learning_rate=1e-06, step_size=10, gamma=1e-07, batch_size=256, regularization=1e-06, N0=2500, sig=0.5, _net_arch=0)
print(a.net_arch)
# netType.c0mp
Sven Eberth
  • 3,057
  • 12
  • 24
  • 29
  • Thanks, that's closer to what I wanted. Now, I've been thinking maybe it would make even more sense to have a `arch_name` property. – Guilherme B.M. Jun 05 '21 at 11:12
  • I think that there's a mistake because the body of the `__post_init__` method should have `self._net_arch = arch_name` (note the underscore). – Boštjan Mejak Jun 09 '22 at 23:39
1

I might be misunderstanding something, but isn't this here what you want, without needing properties?

from dataclasses import dataclass, field


@dataclass
class ReconstructionParameters:

    img_size: int
    CR: int
    denoise: bool
    epochs: int
    learning_rate: float
    step_size: int
    gamma: float
    batch_size: int
    regularization: float
    N0: float
    sig: float
    arch_name: str

    net_arch: int = field(init=False)

    def __post_init__(self):
        self.net_arch = getattr(netType, arch_name)
Arne
  • 17,706
  • 5
  • 83
  • 99