Mask2FormerConfig

class lucid.models.Mask2FormerConfig(num_labels: int, feature_size: int = 256, mask_feature_size: int = 256, hidden_dim: int = 256, backbone_config: dict | None = None, num_channels: int = 3, num_queries: int = 100, encoder_layers: int = 6, encoder_feedforward_dim: int = 1024, decoder_layers: int = 10, dim_feedforward: int = 2048, num_attention_heads: int = 8, feature_strides: list[int] = <factory>, common_stride: int = 4, enforce_input_projection: bool = False, activation_function: str = 'relu', pre_norm: bool = False, dropout: float = 0.0, init_std: float = 0.02, init_xavier_std: float = 1.0, dilation: bool = False, class_weight: float = 2.0, mask_weight: float = 5.0, dice_weight: float = 5.0, no_object_weight: float = 0.1, train_num_points: int = 12544, oversample_ratio: float = 3.0, importance_sample_ratio: float = 0.75, use_auxiliary_loss: bool = True, output_auxiliary_logits: bool | None = None, output_attentions: bool = False, output_hidden_states: bool = False)

Mask2FormerConfig stores the complete model setup used by lucid.models.Mask2Former.

Class Signature

@dataclass
class Mask2FormerConfig:
    num_labels: int
    feature_size: int = 256
    mask_feature_size: int = 256
    hidden_dim: int = 256
    backbone_config: dict | None = None
    num_channels: int = 3
    num_queries: int = 100
    encoder_layers: int = 6
    encoder_feedforward_dim: int = 1024
    decoder_layers: int = 10
    dim_feedforward: int = 2048
    num_attention_heads: int = 8
    feature_strides: list[int] = [4, 8, 16, 32]
    common_stride: int = 4
    enforce_input_projection: bool = False
    activation_function: str = "relu"
    pre_norm: bool = False
    dropout: float = 0.0
    init_std: float = 0.02
    init_xavier_std: float = 1.0
    dilation: bool = False
    class_weight: float = 2.0
    mask_weight: float = 5.0
    dice_weight: float = 5.0
    no_object_weight: float = 0.1
    train_num_points: int = 12544
    oversample_ratio: float = 3.0
    importance_sample_ratio: float = 0.75
    use_auxiliary_loss: bool = True
    output_auxiliary_logits: bool | None = None
    output_attentions: bool = False
    output_hidden_states: bool = False

Usage

import lucid.models as models

cfg = models.Mask2FormerConfig(
    num_labels=150,
    hidden_dim=256,
    feature_size=256,
    mask_feature_size=256,
    backbone_config={
        "model_type": "swin",
        "embed_dim": 96,
        "depths": [2, 2, 18, 2],
        "num_heads": [3, 6, 12, 24],
        "image_size": 224,
        "window_size": 7,
    },
)

model = models.Mask2Former(cfg)