Class IMaskDINODecoder

semantic_sam/body/decoder/interactive_mask_dino.py:24–664

from ...utils import configurable

class IMaskDINODecoder(nn.Module):
    @configurable
    def __init__(
            self,
            lang_encoder: nn.Module,
            in_channels,
            mask_classification=True,
            *,
            num_classes: int,
            hidden_dim: int,
            dim_proj: int,
            num_queries: int,
            nheads: int,
            dim_feedforward: int,
            dec_layers: int,
            mask_dim: int,
            enforce_input_project: bool,
            two_stage: bool,
            dn: str,
            noise_scale:float,
            dn_num:int,
            initialize_box_type:bool,
            initial_pred:bool,
            learn_tgt: bool,
            total_num_feature_levels: int = 4,
            dropout: float = 0.0,
            activation: str = 'relu',
            nhead: int = 8,
            dec_n_points: int = 4,
            return_intermediate_dec: bool = True,
            query_dim: int = 4,
            dec_layer_share: bool = False,
            semantic_ce_loss: bool = False,
            num_mask_tokens: int = 3,
    ):
        """
        NOTE: this interface is experimental.
        Args:
            in_channels: channels of the input features
            mask_classification: whether to add mask classifier or not
            num_classes: number of classes
            hidden_dim: Transformer feature dimension
            num_queries: number of queries
            nheads: number of heads
            dim_feedforward: feature dimension in feedforward network
            enc_layers: number of Transformer encoder layers
            dec_layers: number of Transformer decoder layers
            pre_norm: whether to use pre-LayerNorm or not
            mask_dim: mask feature dimension
            enforce_input_project: add input project 1x1 conv even if input
                channels and hidden dim is identical
            d_model: transformer dimension
            dropout: dropout rate
            activation: activation function
            nhead: num heads in multi-head attention
            dec_n_points: number of sampling points in decoder
            return_intermediate_dec: return the intermediate results of decoder
            query_dim: 4 -> (x, y, w, h)