Skip to content

設定

設定クラスを提供します。

MLPConfig

MLPConfig dataclass 

MLPConfig(hidden_dim, n_layers, output_activation, linear_cfg)

Multi-layer perceptron configuration.

Attributes:

Name Type Description
hidden_dim int

Number of hidden units.
n_layers int

Number of layers.
output_activation str

Activation function for output layer.
linear_cfg LinearConfig

Linear layer configuration.

Attributes

hidden_dim instance-attribute 

hidden_dim

linear_cfg instance-attribute 

linear_cfg

n_layers instance-attribute 

n_layers

output_activation instance-attribute 

output_activation

Functions

dictcfg2dict 

dictcfg2dict()

Convert dictConfig to dict for MLPConfig.

Source code in src/ml_networks/config.py 
389
390
391
392
393
394
def dictcfg2dict(self) -> None:
    """Convert dictConfig to dict for `MLPConfig`."""
    self.linear_cfg.dictcfg2dict()
    for key, value in self.__dict__.items():
        if isinstance(value, DictConfig | ListConfig | list | tuple | dict):
            setattr(self, key, convert_dictconfig_to_dict(value))

LinearConfig

LinearConfig dataclass 

LinearConfig(activation, norm='none', norm_cfg=dict(), dropout=0.0, norm_first=False, bias=True)

A linear layer configuration.

Attributes:

Name Type Description
activation str

Activation function.
norm Literal['layer', 'rms', 'none']

Normalization layer. If it's set to "none", normalization is not applied. Default is "none".
norm_cfg dict

Normalization layer configuration. Default is {}.
dropout float

Dropout rate. If it's set to 0.0, dropout is not applied. Default is 0.0.
norm_first bool

Whether to apply normalization before linear layer. Default is False.
bias bool

Whether to use bias. Default is True.

Attributes

activation instance-attribute 

activation

bias class-attribute instance-attribute 

bias = True

dropout class-attribute instance-attribute 

dropout = 0.0

norm class-attribute instance-attribute 

norm = 'none'

norm_cfg class-attribute instance-attribute 

norm_cfg = field(default_factory=dict)

norm_first class-attribute instance-attribute 

norm_first = False

Functions

__post_init__ 

__post_init__()

Set norm_cfg.

Source code in src/ml_networks/config.py 
351
352
353
354
355
356
def __post_init__(self) -> None:
    """Set `norm_cfg`."""
    if self.norm == "none":
        self.norm_cfg = {}
    else:
        self.norm_cfg = dict(self.norm_cfg)

dictcfg2dict 

dictcfg2dict()

Convert dictConfig to dict for LinearConfig.

Source code in src/ml_networks/config.py 
358
359
360
361
362
363
def dictcfg2dict(self) -> None:
    """Convert dictConfig to dict for `LinearConfig`."""
    self.norm_cfg = dict(self.norm_cfg)
    for key, value in self.__dict__.items():
        if isinstance(value, DictConfig | ListConfig | list | tuple | dict):
            setattr(self, key, convert_dictconfig_to_dict(value))

ConvConfig

ConvConfig dataclass 

ConvConfig(activation, kernel_size, stride, padding, output_padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros', dropout=0.0, norm='none', norm_cfg=dict(), norm_first=False, scale_factor=0)

A convolutional layer configuration.

Attributes:

Name Type Description
activation str

Activation function.
kernel_size int

Kernel size.
stride int

Stride.
padding int

Padding.
output_padding int

Output padding, especially for transposed convolution. Default is 0.
dilation int

Dilation.
groups int

Number of groups. Default is 1. See https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html.
bias bool

Whether to use bias. Default is True.
dropout float

Dropout rate. If it's set to 0.0, dropout is not applied. Default is 0.0.
norm Literal['batch', 'group', 'none']

Normalization layer. If it's set to "none", normalization is not applied. Default is "none".
norm_cfg dict

Normalization layer configuration. If you want to use Instance, Layer, or Group normalization, set norm to "group" and set norm_cfg with "num_groups=$in_channel, 1, or any value". Default is {}.
scale_factor int

Scale factor for upsample, especially for PixelShuffle or PixelUnshuffle. If it's set to >0, upsample is applied. If it's set to <0 downsample is applied. Otherwise, no upsample or downsample is applied. Default is 0.

Attributes

activation instance-attribute 

activation

bias class-attribute instance-attribute 

bias = True

dilation class-attribute instance-attribute 

dilation = 1

dropout class-attribute instance-attribute 

dropout = 0.0

groups class-attribute instance-attribute 

groups = 1

kernel_size instance-attribute 

kernel_size

norm class-attribute instance-attribute 

norm = 'none'

norm_cfg class-attribute instance-attribute 

norm_cfg = field(default_factory=dict)

norm_first class-attribute instance-attribute 

norm_first = False

output_padding class-attribute instance-attribute 

output_padding = 0

padding instance-attribute 

padding

padding_mode class-attribute instance-attribute 

padding_mode = 'zeros'

scale_factor class-attribute instance-attribute 

scale_factor = 0

stride instance-attribute 

stride

Functions

__post_init__ 

__post_init__()

Set .norm_cfg.

Source code in src/ml_networks/config.py 
151
152
153
154
155
156
def __post_init__(self) -> None:
    """Set `.norm_cfg`."""
    if self.norm == "none":
        self.norm_cfg = {}
    else:
        self.norm_cfg = dict(self.norm_cfg)

dictcfg2dict 

dictcfg2dict()

Convert dictConfig to dict for ConvConfig.

Source code in src/ml_networks/config.py 
158
159
160
161
162
163
164
def dictcfg2dict(self) -> None:
    """Convert dictConfig to dict for `ConvConfig`."""
    self.norm_cfg = dict(self.norm_cfg)

    for key, value in self.__dict__.items():
        if isinstance(value, DictConfig | ListConfig | list | tuple | dict):
            setattr(self, key, convert_dictconfig_to_dict(value))

ConvNetConfig

ConvNetConfig dataclass 

ConvNetConfig(channels, conv_cfgs, attention=None, init_channel=16)

Convolutional neural network layers configuration.

Attributes:

Name Type Description
channels Tuple[int, ...]

Number of channels for each layer.
conv_cfgs Tuple[ConvConfig, ...]

Convolutional layer configurations. The length of conv_cfgs should be the same as the length of channels.
init_channel int

Initial number of channels, especially for transposed convolution.

Attributes

attention class-attribute instance-attribute 

attention = None

channels instance-attribute 

channels

conv_cfgs instance-attribute 

conv_cfgs

init_channel class-attribute instance-attribute 

init_channel = 16

Functions

__post_init__ 

__post_init__()

Set channels and conv_cfgs as tuple.

Source code in src/ml_networks/config.py 
188
189
190
191
def __post_init__(self) -> None:
    """Set `channels` and `conv_cfgs` as tuple."""
    self.conv_cfgs = tuple(self.conv_cfgs)
    self.channels = tuple(self.channels)

dictcfg2dict 

dictcfg2dict()

Convert dictConfig to dict for ConvNetConfig.

Source code in src/ml_networks/config.py 
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
def dictcfg2dict(self) -> None:
    """Convert dictConfig to dict for `ConvNetConfig`."""
    self.channels = tuple(self.channels)
    conv_cfgs = []
    for cfg_item in self.conv_cfgs:
        conv_cfg = cfg_item
        if isinstance(conv_cfg, DictConfig):
            conv_cfg_dict = convert_dictconfig_to_dict(conv_cfg)
            conv_cfg = ConvConfig(**conv_cfg_dict)
        conv_cfg.dictcfg2dict()
        conv_cfgs.append(conv_cfg)
    self.conv_cfgs = tuple(conv_cfgs)
    for key, value in self.__dict__.items():
        if isinstance(value, DictConfig | ListConfig | list | tuple | dict):
            setattr(self, key, convert_dictconfig_to_dict(value))

ResNetConfig

ResNetConfig dataclass 

ResNetConfig(conv_channel, conv_kernel, f_kernel, conv_activation, out_activation, n_res_blocks, scale_factor=2, n_scaling=2, norm='none', norm_cfg=dict(), dropout=0.0, init_channel=16, padding_mode='zeros', attention=None)

Residual neural network layers configuration.

Attributes:

Name Type Description
conv_channel int

Number of channels for convolutional layer. In ResNet, common number of channels is used for all layers.
conv_kernel int

Kernel size for convolutional layer. In ResNet, common kernel size is used for all layers.
f_kernel int

Kernel size for final or first convolutional layer. This depends on whether PixelShuffle or PixelUnshuffle is used.
conv_activation str

Activation function for convolutional layer.
out_activation str

Activation function for output layer.
n_res_blocks int

Number of residual blocks.
scale_factor int

Scale factor for upsample, especially for PixelShuffle or PixelUnshuffle.
n_scaling int

Number of upsample or downsample layers. The image size is scaled by scalefactor^n_scaling.
norm Literal['batch', 'group', 'none']

Normalization layer. If it's set to "none", normalization is not applied. Default is "none".
norm_cfg dict

Normalization layer configuration. If you want to use Instance, Layer, or Group normalization, set norm to "group" and set norm_cfg with "num_groups=$in_channel, 1, or any value". Default is {}.
dropout float

Dropout rate. If it's set to 0.0, dropout is not applied. Default is 0.0.
init_channel int

Initial number of channels, especially for decoder.

Attributes

attention class-attribute instance-attribute 

attention = None

conv_activation instance-attribute 

conv_activation

conv_channel instance-attribute 

conv_channel

conv_kernel instance-attribute 

conv_kernel

dropout class-attribute instance-attribute 

dropout = 0.0

f_kernel instance-attribute 

f_kernel

init_channel class-attribute instance-attribute 

init_channel = 16

n_res_blocks instance-attribute 

n_res_blocks

n_scaling class-attribute instance-attribute 

n_scaling = 2

norm class-attribute instance-attribute 

norm = 'none'

norm_cfg class-attribute instance-attribute 

norm_cfg = field(default_factory=dict)

out_activation instance-attribute 

out_activation

padding_mode class-attribute instance-attribute 

padding_mode = 'zeros'

scale_factor class-attribute instance-attribute 

scale_factor = 2

Functions

__post_init__ 

__post_init__()

Set norm_cfg.

Source code in src/ml_networks/config.py 
262
263
264
265
266
267
def __post_init__(self) -> None:
    """Set `norm_cfg`."""
    if self.norm == "none":
        self.norm_cfg = {}
    else:
        self.norm_cfg = dict(self.norm_cfg)

dictcfg2dict 

dictcfg2dict()

Convert dictConfig to dict for ResNetConfig.

Source code in src/ml_networks/config.py 
269
270
271
272
273
274
def dictcfg2dict(self) -> None:
    """Convert dictConfig to dict for `ResNetConfig`."""
    self.norm_cfg = dict(self.norm_cfg)
    for key, value in self.__dict__.items():
        if isinstance(value, DictConfig | ListConfig | list | tuple | dict):
            setattr(self, key, convert_dictconfig_to_dict(value))

EncoderConfig

EncoderConfig dataclass 

EncoderConfig(backbone, full_connection)

Encoder configuration.

Attributes:

Name Type Description
backbone Union[ConvNetConfig, ResNetConfig]

Backbone configuration.
full_connection Union[MLPConfig, LinearConfig, SpatialSoftmaxConfig]

Full connection configuration.

Attributes

backbone instance-attribute 

backbone

full_connection instance-attribute 

full_connection

Functions

dictcfg2dict 

dictcfg2dict()

Convert dictConfig to dict for EncoderConfig.

Source code in src/ml_networks/config.py 
522
523
524
525
526
527
528
529
530
def dictcfg2dict(self) -> None:
    """Convert dictConfig to dict for `EncoderConfig`."""
    if hasattr(self.backbone, "dictcfg2dict"):
        self.backbone.dictcfg2dict()
    if hasattr(self.full_connection, "dictcfg2dict"):
        self.full_connection.dictcfg2dict()
    for key, value in self.__dict__.items():
        if isinstance(value, DictConfig | ListConfig | list | tuple | dict):
            setattr(self, key, convert_dictconfig_to_dict(value))

DecoderConfig

DecoderConfig dataclass 

DecoderConfig(backbone, full_connection)

Decoder configuration.

Attributes:

Name Type Description
backbone Union[ConvNetConfig, ResNetConfig]

Backbone configuration.
full_connection Union[MLPConfig, LinearConfig, SpatialSoftmaxConfig]

Full connection configuration.

Attributes

backbone instance-attribute 

backbone

full_connection instance-attribute 

full_connection

Functions

dictcfg2dict 

dictcfg2dict()

Convert dictConfig to dict for DecoderConfig.

Source code in src/ml_networks/config.py 
549
550
551
552
553
554
555
556
def dictcfg2dict(self) -> None:
    """Convert dictConfig to dict for `DecoderConfig`."""
    self.backbone.dictcfg2dict()
    if hasattr(self.full_connection, "dictcfg2dict"):
        self.full_connection.dictcfg2dict()
    for key, value in self.__dict__.items():
        if isinstance(value, DictConfig | ListConfig | list | tuple | dict):
            setattr(self, key, convert_dictconfig_to_dict(value))

ViTConfig

ViTConfig dataclass 

ViTConfig(patch_size, transformer_cfg, cls_token=True, init_channel=16, unpatchify=False)

Vision Transformer configuration.

Attributes:

Name Type Description
patch_size int

Patch size.
transformer_cfg TransformerConfig

Transformer configuration.
cls_token bool

Whether to use class token. Default is True.
init_channel int

Initial number of channels. Default is 16.

Attributes

cls_token class-attribute instance-attribute 

cls_token = True

init_channel class-attribute instance-attribute 

init_channel = 16

patch_size instance-attribute 

patch_size

transformer_cfg instance-attribute 

transformer_cfg

unpatchify class-attribute instance-attribute 

unpatchify = False

Functions

UNetConfig

UNetConfig dataclass 

UNetConfig(channels, conv_cfg, cond_pred_scale=False, nhead=None, has_attn=False, use_shuffle=False, use_hypernet=False, hyper_mlp_cfg=None)

UNet configuration.

Attributes:

Name Type Description
channels Tuple[int, ...]

Number of channels for each layer.
conv_cfg ConvConfig

Convolutional layer configuration.
cond_cfg MLPConfig

Conditional configuration for UNet.
cond_pred_scale bool

Whether to scale the conditional prediction. Default is False.
nhead Optional[int]

Number of heads for attention mechanism. If it's set to None, attention is not applied. Default is None.
has_attn bool

Whether to use attention mechanism. Default is False.
use_shuffle bool

Whether to use PixelShuffle or PixelUnshuffle. Default is False.
use_hypernet bool

Whether to use hypernetwork. Default is False.
hyper_mlp_cfg Optional[MLPConfig]

Hypernetwork configuration. If it's set to None, hypernetwork is not used. Default is None.

Attributes

channels instance-attribute 

channels

cond_pred_scale class-attribute instance-attribute 

cond_pred_scale = False

conv_cfg instance-attribute 

conv_cfg

has_attn class-attribute instance-attribute 

has_attn = False

hyper_mlp_cfg class-attribute instance-attribute 

hyper_mlp_cfg = None

nhead class-attribute instance-attribute 

nhead = None

use_hypernet class-attribute instance-attribute 

use_hypernet = False

use_shuffle class-attribute instance-attribute 

use_shuffle = False

Functions

__post_init__ 

__post_init__()

Post-initialization processing.

Source code in src/ml_networks/config.py 
315
316
317
318
319
320
def __post_init__(self) -> None:
    """Post-initialization processing."""
    if self.has_attn:
        assert self.nhead is not None, "nhead must be specified when has_attn is True."
    if isinstance(self.channels, list | ListConfig):
        self.channels = tuple(self.channels)

SpatialSoftmaxConfig

SpatialSoftmaxConfig dataclass 

SpatialSoftmaxConfig(temperature=1.0, eps=1e-06, is_argmax=False, is_straight_through=False, additional_layer=None)

Spatial softmax configuration.

Attributes:

Name Type Description
temperature float

Softmax temperature. If it's set to 0.0, the layer outputs the coordinates of the maximum value. Otherwise, the layer outputs the expectation of the coordinates with softmax function. Default is 0.0.
eps float

Epsilon value for numerical stability in softmax. Default is 1e-6.
is_argmax bool

Whether to use argmax instead of softmax. Default is False.
is_straight_through bool

Whether to use straight-through estimator for backpropagation. Default is False.
additional_layer Optional[Union[MLPConfig, LinearConfig]]

Additional layer configuration. If it's set to None, no additional layer is applied. Default is None.

Attributes

additional_layer class-attribute instance-attribute 

additional_layer = None

eps class-attribute instance-attribute 

eps = 1e-06

is_argmax class-attribute instance-attribute 

is_argmax = False

is_straight_through class-attribute instance-attribute 

is_straight_through = False

temperature class-attribute instance-attribute 

temperature = 1.0

Functions

AttentionConfig

AttentionConfig dataclass 

AttentionConfig(nhead, patch_size)

Attention configuration.

Attributes:

Name Type Description
nhead int

Number of heads.
patch_size int

Patch size.

Attributes

nhead instance-attribute 

nhead

patch_size instance-attribute 

patch_size

Functions

TransformerConfig

TransformerConfig dataclass 

TransformerConfig(d_model, nhead, dim_ff, n_layers, dropout=0.1, hidden_activation='GELU', output_activation='GeLU')

Transformer configuration.

Attributes:

Name Type Description
d_model int

Dimension of model.
nhead int

Number of heads.
dim_ff int

Dimension of feedforward network.
n_layers int

Number of layers.
dropout float

Dropout rate. Default is 0.1.
hidden_activation Literal['ReLU', 'GELU']

Activation function for hidden layer. Default is "GELU".
output_activation str

Activation function for output layer. Default is "GeLU".

Attributes

d_model instance-attribute 

d_model

dim_ff instance-attribute 

dim_ff

dropout class-attribute instance-attribute 

dropout = 0.1

hidden_activation class-attribute instance-attribute 

hidden_activation = 'GELU'

n_layers instance-attribute 

n_layers

nhead instance-attribute 

nhead

output_activation class-attribute instance-attribute 

output_activation = 'GeLU'

Functions

AdaptiveAveragePoolingConfig

AdaptiveAveragePoolingConfig dataclass 

AdaptiveAveragePoolingConfig(output_size=(1, 1), additional_layer=None)

Adaptive average pooling configuration.

Attributes:

Name Type Description
output_size Union[int, Tuple[int, ...]]

Output size of the pooling layer. If it's an integer, it will be used for both height and width. If it's a tuple, it should contain two integers for height and width.

Attributes

additional_layer class-attribute instance-attribute 

additional_layer = None

output_size class-attribute instance-attribute 

output_size = (1, 1)

Functions

__post_init__ 

__post_init__()

Ensure output_size is a tuple.

Source code in src/ml_networks/config.py 
469
470
471
472
473
474
def __post_init__(self) -> None:
    """Ensure output_size is a tuple."""
    if isinstance(self.output_size, int):
        self.output_size = (self.output_size, self.output_size)
    elif isinstance(self.output_size, list | ListConfig):
        self.output_size = tuple(self.output_size)

SoftmaxTransConfig

SoftmaxTransConfig dataclass 

SoftmaxTransConfig(vector, sigma, n_ignore=0, max=1.0, min=-1.0)

Softmax transformation configuration.

Attributes:

Name Type Description
vector int

Vector size.
sigma float

Sigma value.
n_ignore int

Number of ignored elements. Default is 0.
max float

Maximum value. Default is 1.0.
min float

Minimum value. Default is -1.0.

Attributes

max class-attribute instance-attribute 

max = 1.0

min class-attribute instance-attribute 

min = -1.0

n_ignore class-attribute instance-attribute 

n_ignore = 0

sigma instance-attribute 

sigma

vector instance-attribute 

vector

Functions

ContrastiveLearningConfig

ContrastiveLearningConfig dataclass 

ContrastiveLearningConfig(dim_feature, eval_func, dim_input2=None, cross_entropy_like=False)

Contrastive learning configuration.

Attributes

cross_entropy_like class-attribute instance-attribute 

cross_entropy_like = False

dim_feature instance-attribute 

dim_feature

dim_input2 class-attribute instance-attribute 

dim_input2 = None

eval_func instance-attribute 

eval_func

Functions