"""RT-DETRv2 configuration dataclasses. Fields mirror the HuggingFace `RTDetrV2Config` schema (`transformers.models.rt_detr_v2.configuration_rt_detr_v2`). `from_dict` flattens an HF `config.json` straight into `ModelConfig`; the sub-configs for backbone / hybrid encoder / decoder are constructed in `__post_init__`. """ from dataclasses import dataclass, field from typing import List, Optional, Union from ..base import BaseModelConfig @dataclass class RTDetrResNetConfig(BaseModelConfig): """ResNet-vd backbone configuration. The `vd` variant has a 3-stage stem (3x3 -> 3x3 -> 3x3, stride 2/1/1) followed by a 3x3 stride-2 max-pool, and uses `AvgPool2x2 stride 2 + 1x1 conv` for downsampling shortcuts rather than a stride-2 1x1. Depths default to ResNet-50 (`[3, 4, 6, 3]`); ResNet-101 is `[3, 4, 23, 3]`. BatchNorms are typically frozen at inference. """ model_type: str = "rt_detr_resnet" depths: List[int] = field(default_factory=lambda: [3, 4, 6, 3]) downsample_in_bottleneck: bool = False downsample_in_first_stage: bool = False embedding_size: int = 64 hidden_act: str = "relu" hidden_sizes: List[int] = field(default_factory=lambda: [256, 512, 1024, 2048]) layer_type: str = "bottleneck" num_channels: int = 3 out_features: List[str] = field( default_factory=lambda: ["stage2", "stage3", "stage4"] ) out_indices: List[int] = field(default_factory=lambda: [2, 3, 4]) stage_names: List[str] = field( default_factory=lambda: ["stem", "stage1", "stage2", "stage3", "stage4"] ) @dataclass class RTDetrV2HybridEncoderConfig(BaseModelConfig): """Hybrid encoder (AIFI + FPN + PAN) configuration.""" model_type: str = "rt_detr_v2_hybrid_encoder" encoder_hidden_dim: int = 256 encoder_in_channels: List[int] = field(default_factory=lambda: [512, 1024, 2048]) feat_strides: List[int] = field(default_factory=lambda: [8, 16, 32]) encoder_layers: int = 1 encoder_ffn_dim: int = 1024 encoder_attention_heads: int = 8 encoder_activation_function: str = "gelu" encode_proj_layers: List[int] = field(default_factory=lambda: [2]) positional_encoding_temperature: int = 10000 activation_function: str = "silu" normalize_before: bool = False layer_norm_eps: float = 1e-5 hidden_expansion: float = 1.0 batch_norm_eps: float = 1e-5 eval_size: Optional[List[int]] = None @dataclass class RTDetrV2TransformerConfig(BaseModelConfig): """Decoder + query-selection configuration.""" model_type: str = "rt_detr_v2_transformer" d_model: int = 256 decoder_layers: int = 6 decoder_attention_heads: int = 8 decoder_ffn_dim: int = 1024 decoder_in_channels: List[int] = field(default_factory=lambda: [256, 256, 256]) decoder_activation_function: str = "relu" decoder_method: str = "default" decoder_n_levels: int = 3 decoder_n_points: int = 4 decoder_offset_scale: float = 0.5 num_feature_levels: int = 3 num_queries: int = 300 num_labels: int = 17 learn_initial_query: bool = False layer_norm_eps: float = 1e-5 with_box_refine: bool = True use_focal_loss: bool = True @dataclass class ModelConfig(BaseModelConfig): """Top-level RT-DETRv2 config. HF stores backbone/encoder/decoder fields flat in `config.json`; `__post_init__` rebuilds the sub-configs from these flat fields.""" model_type: str = "rt_detr_v2" image_size: int = 640 num_labels: int = 17 id2label: Optional[dict] = None label2id: Optional[dict] = None backbone_config: Optional[Union[dict, RTDetrResNetConfig]] = None d_model: int = 256 encoder_hidden_dim: int = 256 encoder_in_channels: List[int] = field(default_factory=lambda: [512, 1024, 2048]) feat_strides: List[int] = field(default_factory=lambda: [8, 16, 32]) encoder_layers: int = 1 encoder_ffn_dim: int = 1024 encoder_attention_heads: int = 8 encoder_activation_function: str = "gelu" encode_proj_layers: List[int] = field(default_factory=lambda: [2]) positional_encoding_temperature: int = 10000 activation_function: str = "silu" normalize_before: bool = False layer_norm_eps: float = 1e-5 hidden_expansion: float = 1.0 batch_norm_eps: float = 1e-5 eval_size: Optional[List[int]] = None decoder_layers: int = 6 decoder_attention_heads: int = 8 decoder_ffn_dim: int = 1024 decoder_in_channels: List[int] = field(default_factory=lambda: [256, 256, 256]) decoder_activation_function: str = "relu" decoder_method: str = "default" decoder_n_levels: int = 3 decoder_n_points: int = 4 decoder_offset_scale: float = 0.5 num_feature_levels: int = 3 num_queries: int = 300 learn_initial_query: bool = False with_box_refine: bool = True use_focal_loss: bool = True freeze_backbone_batch_norms: bool = True def __post_init__(self): if self.backbone_config is None: self.backbone_config = RTDetrResNetConfig() elif isinstance(self.backbone_config, dict): self.backbone_config = RTDetrResNetConfig.from_dict(self.backbone_config) self._hybrid_encoder_config = RTDetrV2HybridEncoderConfig( encoder_hidden_dim=self.encoder_hidden_dim, encoder_in_channels=self.encoder_in_channels, feat_strides=self.feat_strides, encoder_layers=self.encoder_layers, encoder_ffn_dim=self.encoder_ffn_dim, encoder_attention_heads=self.encoder_attention_heads, encoder_activation_function=self.encoder_activation_function, encode_proj_layers=self.encode_proj_layers, positional_encoding_temperature=self.positional_encoding_temperature, activation_function=self.activation_function, normalize_before=self.normalize_before, layer_norm_eps=self.layer_norm_eps, hidden_expansion=self.hidden_expansion, batch_norm_eps=self.batch_norm_eps, eval_size=self.eval_size, ) self._transformer_config = RTDetrV2TransformerConfig( d_model=self.d_model, decoder_layers=self.decoder_layers, decoder_attention_heads=self.decoder_attention_heads, decoder_ffn_dim=self.decoder_ffn_dim, decoder_in_channels=self.decoder_in_channels, decoder_activation_function=self.decoder_activation_function, decoder_method=self.decoder_method, decoder_n_levels=self.decoder_n_levels, decoder_n_points=self.decoder_n_points, decoder_offset_scale=self.decoder_offset_scale, num_feature_levels=self.num_feature_levels, num_queries=self.num_queries, num_labels=self.num_labels, learn_initial_query=self.learn_initial_query, layer_norm_eps=self.layer_norm_eps, with_box_refine=self.with_box_refine, use_focal_loss=self.use_focal_loss, ) # Framework compatibility: sanitize_weights accesses these. self.text_config = None self.vision_config = None