# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """RT-DETR model with HF-compatible parameter naming. Architecture mirrors HuggingFace transformers' RTDetrForObjectDetection so that both ``paddle_dynamic`` and ``transformers`` backends can load the same HF-format safetensors without heavy runtime conversion. Key naming conventions (matching HF transformers): - model.backbone.model.* (HGNetV2 backbone) - model.encoder_input_proj.* (encoder input projections) - model.encoder.aifi.* (AIFI transformer encoder layers) - model.encoder.{lateral,fpn,downsample,pan}_* (hybrid FPN/PAN) - model.decoder.layers.* (decoder transformer layers) - model.decoder.class_embed.* (per-layer class prediction) - model.decoder.bbox_embed.* (per-layer bbox prediction) A lightweight ``set_hf_state_dict`` handles minor key differences between the safetensors on-disk format and the model's parameter names (see ``_apply_rt_detr_key_conversion`` in pp_doclayout_v2.py). """ from __future__ import absolute_import, division, print_function import paddle import paddle.nn as nn import paddle.nn.functional as F from ...common.transformers.activations import ACT2CLS, ACT2FN from ...common.transformers.transformers import ( BatchNormHFStateDictMixin, PretrainedModel, ) from ...image_classification.modeling.hgnetv2 import HGNetV2Backbone from ._config_rt_detr import RTDETRConfig from .pp_doclayout_v2 import ( _apply_rt_detr_key_conversion, _reverse_rt_detr_key_conversion, ) __all__ = ["RTDETR"] def bbox_cxcywh_to_xyxy(x): cxcy, wh = paddle.split(x, 2, axis=-1) return paddle.concat([cxcy - 0.5 * wh, cxcy + 0.5 * wh], axis=-1) def inverse_sigmoid(x, eps=1e-5): x = x.clip(min=0, max=1) x1 = x.clip(min=eps) x2 = (1 - x).clip(min=eps) return paddle.log(x1 / x2) class RTDETRFrozenBatchNorm2d(nn.Layer): """BatchNorm2d where the batch statistics and the affine parameters are fixed.""" def __init__(self, n): super().__init__() self.register_buffer("weight", paddle.ones([n])) self.register_buffer("bias", paddle.zeros([n])) self.register_buffer("_mean", paddle.zeros([n])) self.register_buffer("_variance", paddle.ones([n])) def forward(self, x): weight = self.weight.reshape([1, -1, 1, 1]) bias = self.bias.reshape([1, -1, 1, 1]) running_var = self._variance.reshape([1, -1, 1, 1]) running_mean = self._mean.reshape([1, -1, 1, 1]) epsilon = 1e-5 scale = weight * (running_var + epsilon).rsqrt() bias = bias - running_mean * scale return x * scale + bias class RTDETRMLPPredictionHead(nn.Layer): """Simple multi-layer perceptron for bbox/class prediction.""" def __init__(self, input_dim, hidden_dim, output_dim, num_layers): super().__init__() self.num_layers = num_layers h = [hidden_dim] * (num_layers - 1) self.layers = nn.LayerList( nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]) ) def forward(self, x): for i, layer in enumerate(self.layers): x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x) return x class RTDETTMLP(nn.Layer): """Feed-forward MLP used in encoder and decoder layers.""" def __init__(self, config, hidden_size, intermediate_size, activation_function): super().__init__() self.fc1 = nn.Linear(hidden_size, intermediate_size) self.fc2 = nn.Linear(intermediate_size, hidden_size) self.activation_fn = ACT2FN[activation_function] self.activation_dropout = config.activation_dropout self.dropout = config.dropout def forward(self, hidden_states): hidden_states = self.activation_fn(self.fc1(hidden_states)) hidden_states = F.dropout( hidden_states, p=self.activation_dropout, training=self.training ) hidden_states = self.fc2(hidden_states) hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training) return hidden_states class RTDETRSelfAttention(nn.Layer): """Multi-headed self-attention. Position embeddings added to queries and keys.""" def __init__( self, config, hidden_size, num_attention_heads, dropout=0.0, bias=True ): super().__init__() self.head_dim = hidden_size // num_attention_heads self.num_heads = num_attention_heads self.scaling = self.head_dim**-0.5 self.attention_dropout = dropout self.k_proj = nn.Linear(hidden_size, hidden_size, bias_attr=bias) self.v_proj = nn.Linear(hidden_size, hidden_size, bias_attr=bias) self.q_proj = nn.Linear(hidden_size, hidden_size, bias_attr=bias) self.o_proj = nn.Linear(hidden_size, hidden_size, bias_attr=bias) def forward(self, hidden_states, attention_mask=None, position_embeddings=None): batch_size, seq_len, _ = hidden_states.shape query_key_input = ( hidden_states + position_embeddings if position_embeddings is not None else hidden_states ) query_states = ( self.q_proj(query_key_input) .reshape([batch_size, seq_len, self.num_heads, self.head_dim]) .transpose([0, 2, 1, 3]) ) key_states = ( self.k_proj(query_key_input) .reshape([batch_size, seq_len, self.num_heads, self.head_dim]) .transpose([0, 2, 1, 3]) ) value_states = ( self.v_proj(hidden_states) .reshape([batch_size, seq_len, self.num_heads, self.head_dim]) .transpose([0, 2, 1, 3]) ) attn_weights = ( paddle.matmul(query_states, key_states.transpose([0, 1, 3, 2])) * self.scaling ) if attention_mask is not None: attn_weights = attn_weights + attention_mask attn_weights = F.softmax(attn_weights, axis=-1) attn_weights = F.dropout( attn_weights, p=self.attention_dropout, training=self.training ) attn_output = paddle.matmul(attn_weights, value_states) attn_output = attn_output.transpose([0, 2, 1, 3]).reshape( [batch_size, seq_len, -1] ) attn_output = self.o_proj(attn_output) return attn_output, attn_weights class RTDETRConvNormLayer(nn.Layer): """Conv layer with conv/norm attribute names.""" def __init__( self, config, in_channels, out_channels, kernel_size, stride, padding=None, activation=None, ): super().__init__() self.conv = nn.Conv2D( in_channels, out_channels, kernel_size, stride, padding=(kernel_size - 1) // 2 if padding is None else padding, bias_attr=False, ) self.norm = nn.BatchNorm2D(out_channels, epsilon=config.batch_norm_eps) self.activation = nn.Identity() if activation is None else ACT2CLS[activation]() def forward(self, hidden_state): hidden_state = self.conv(hidden_state) hidden_state = self.norm(hidden_state) hidden_state = self.activation(hidden_state) return hidden_state class RTDETREncoderLayer(nn.Layer): def __init__(self, config): super().__init__() self.normalize_before = config.normalize_before self.hidden_size = config.encoder_hidden_dim self.self_attn = RTDETRSelfAttention( config=config, hidden_size=self.hidden_size, num_attention_heads=config.encoder_attention_heads, dropout=config.dropout, ) self.self_attn_layer_norm = nn.LayerNorm( self.hidden_size, epsilon=config.layer_norm_eps ) self.dropout = config.dropout self.mlp = RTDETTMLP( config, self.hidden_size, config.encoder_ffn_dim, config.encoder_activation_function, ) self.final_layer_norm = nn.LayerNorm( self.hidden_size, epsilon=config.layer_norm_eps ) def forward(self, hidden_states, attention_mask, spatial_position_embeddings=None): residual = hidden_states if self.normalize_before: hidden_states = self.self_attn_layer_norm(hidden_states) hidden_states, _ = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, position_embeddings=spatial_position_embeddings, ) hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states if not self.normalize_before: hidden_states = self.self_attn_layer_norm(hidden_states) if self.normalize_before: hidden_states = self.final_layer_norm(hidden_states) residual = hidden_states hidden_states = self.mlp(hidden_states) hidden_states = residual + hidden_states if not self.normalize_before: hidden_states = self.final_layer_norm(hidden_states) return hidden_states class RTDETRRepVggBlock(nn.Layer): """RepVGG architecture block.""" def __init__(self, config): super().__init__() activation = config.activation_function hidden_channels = int(config.encoder_hidden_dim * config.hidden_expansion) self.conv1 = RTDETRConvNormLayer( config, hidden_channels, hidden_channels, 3, 1, padding=1 ) self.conv2 = RTDETRConvNormLayer( config, hidden_channels, hidden_channels, 1, 1, padding=0 ) self.activation = nn.Identity() if activation is None else ACT2CLS[activation]() def forward(self, x): y = self.conv1(x) + self.conv2(x) return self.activation(y) class RTDETRCSPRepLayer(nn.Layer): """Cross Stage Partial (CSP) network layer with RepVGG blocks.""" def __init__(self, config): super().__init__() in_channels = config.encoder_hidden_dim * 2 out_channels = config.encoder_hidden_dim num_blocks = 3 activation = config.activation_function hidden_channels = int(out_channels * config.hidden_expansion) self.conv1 = RTDETRConvNormLayer( config, in_channels, hidden_channels, 1, 1, activation=activation ) self.conv2 = RTDETRConvNormLayer( config, in_channels, hidden_channels, 1, 1, activation=activation ) self.bottlenecks = nn.Sequential( *[RTDETRRepVggBlock(config) for _ in range(num_blocks)] ) if hidden_channels != out_channels: self.conv3 = RTDETRConvNormLayer( config, hidden_channels, out_channels, 1, 1, activation=activation ) else: self.conv3 = nn.Identity() def forward(self, hidden_state): hidden_state_1 = self.conv1(hidden_state) hidden_state_1 = self.bottlenecks(hidden_state_1) hidden_state_2 = self.conv2(hidden_state) return self.conv3(hidden_state_1 + hidden_state_2) class RTDETRSinePositionEmbedding(nn.Layer): """2D sinusoidal position embedding.""" def __init__(self, embed_dim=256, temperature=10000): super().__init__() self.embed_dim = embed_dim self.temperature = temperature def forward(self, width, height, dtype): grid_w = paddle.arange(width).astype(dtype) grid_h = paddle.arange(height).astype(dtype) grid_h, grid_w = paddle.meshgrid(grid_h, grid_w) if self.embed_dim % 4 != 0: raise ValueError( "Embed dimension must be divisible by 4 for 2D sin-cos position embedding" ) pos_dim = self.embed_dim // 4 omega = paddle.arange(pos_dim).astype(dtype) / pos_dim omega = 1.0 / (self.temperature**omega) out_w = grid_w.flatten().unsqueeze(-1) @ omega.unsqueeze(0) out_h = grid_h.flatten().unsqueeze(-1) @ omega.unsqueeze(0) return paddle.concat( [out_h.sin(), out_h.cos(), out_w.sin(), out_w.cos()], axis=1 ).unsqueeze(0) class RTDETRAIFILayer(nn.Layer): """AIFI (Attention-based Intra-scale Feature Interaction) layer.""" def __init__(self, config): super().__init__() self.config = config self.encoder_hidden_dim = config.encoder_hidden_dim self.eval_size = config.eval_size self.position_embedding = RTDETRSinePositionEmbedding( embed_dim=self.encoder_hidden_dim, temperature=config.positional_encoding_temperature, ) self.layers = nn.LayerList( [RTDETREncoderLayer(config) for _ in range(config.encoder_layers)] ) def forward(self, hidden_states): batch_size = hidden_states.shape[0] height, width = hidden_states.shape[2:] hidden_states = hidden_states.flatten(2).transpose([0, 2, 1]) if self.training or self.eval_size is None: pos_embed = self.position_embedding( width=width, height=height, dtype=hidden_states.dtype, ) else: pos_embed = None for layer in self.layers: hidden_states = layer( hidden_states, attention_mask=None, spatial_position_embeddings=pos_embed, ) hidden_states = hidden_states.transpose([0, 2, 1]).reshape( [batch_size, self.encoder_hidden_dim, height, width] ) return hidden_states class RTDETRHybridEncoder(nn.Layer): """Hybrid encoder: AIFI layers + top-down FPN + bottom-up PAN.""" def __init__(self, config): super().__init__() self.config = config self.in_channels = config.encoder_in_channels self.feat_strides = config.feat_strides self.encoder_hidden_dim = config.encoder_hidden_dim self.encode_proj_layers = config.encode_proj_layers self.eval_size = config.eval_size self.num_fpn_stages = len(self.in_channels) - 1 self.num_pan_stages = len(self.in_channels) - 1 # AIFI layers self.aifi = nn.LayerList( [RTDETRAIFILayer(config) for _ in range(len(self.encode_proj_layers))] ) # top-down FPN self.lateral_convs = nn.LayerList() self.fpn_blocks = nn.LayerList() for _ in range(self.num_fpn_stages): lateral_conv = RTDETRConvNormLayer( config, in_channels=self.encoder_hidden_dim, out_channels=self.encoder_hidden_dim, kernel_size=1, stride=1, activation=config.activation_function, ) fpn_block = RTDETRCSPRepLayer(config) self.lateral_convs.append(lateral_conv) self.fpn_blocks.append(fpn_block) # bottom-up PAN self.downsample_convs = nn.LayerList() self.pan_blocks = nn.LayerList() for _ in range(self.num_pan_stages): downsample_conv = RTDETRConvNormLayer( config, in_channels=self.encoder_hidden_dim, out_channels=self.encoder_hidden_dim, kernel_size=3, stride=2, activation=config.activation_function, ) pan_block = RTDETRCSPRepLayer(config) self.downsample_convs.append(downsample_conv) self.pan_blocks.append(pan_block) def forward(self, feature_maps): # AIFI: Apply transformer encoder to specified feature levels if self.config.encoder_layers > 0: for i, enc_ind in enumerate(self.encode_proj_layers): feature_maps[enc_ind] = self.aifi[i](feature_maps[enc_ind]) # top-down FPN fpn_feature_maps = [feature_maps[-1]] for idx, (lateral_conv, fpn_block) in enumerate( zip(self.lateral_convs, self.fpn_blocks) ): backbone_feature_map = feature_maps[self.num_fpn_stages - idx - 1] top_fpn_feature_map = fpn_feature_maps[-1] top_fpn_feature_map = lateral_conv(top_fpn_feature_map) fpn_feature_maps[-1] = top_fpn_feature_map top_fpn_feature_map = F.interpolate( top_fpn_feature_map, scale_factor=2.0, mode="nearest" ) fused_feature_map = paddle.concat( [top_fpn_feature_map, backbone_feature_map], axis=1 ) new_fpn_feature_map = fpn_block(fused_feature_map) fpn_feature_maps.append(new_fpn_feature_map) fpn_feature_maps.reverse() # bottom-up PAN pan_feature_maps = [fpn_feature_maps[0]] for idx, (downsample_conv, pan_block) in enumerate( zip(self.downsample_convs, self.pan_blocks) ): top_pan_feature_map = pan_feature_maps[-1] fpn_feature_map = fpn_feature_maps[idx + 1] downsampled_feature_map = downsample_conv(top_pan_feature_map) fused_feature_map = paddle.concat( [downsampled_feature_map, fpn_feature_map], axis=1 ) new_pan_feature_map = pan_block(fused_feature_map) pan_feature_maps.append(new_pan_feature_map) return pan_feature_maps class MultiScaleDeformableAttention(nn.Layer): """Eager fallback implementation using grid_sample.""" def forward( self, value, value_spatial_shapes, value_spatial_shapes_list, level_start_index, sampling_locations, attention_weights, im2col_step, ): batch_size, _, num_heads, hidden_dim = value.shape _, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape value_list = paddle.split( value, [height * width for height, width in value_spatial_shapes_list], axis=1, ) sampling_grids = 2 * sampling_locations - 1 sampling_value_list = [] for level_id, (height, width) in enumerate(value_spatial_shapes_list): value_l_ = ( value_list[level_id] .flatten(2) .transpose([0, 2, 1]) .reshape([batch_size * num_heads, hidden_dim, height, width]) ) sampling_grid_l_ = ( sampling_grids[:, :, :, level_id] .transpose([0, 2, 1, 3]) .reshape([batch_size * num_heads, num_queries, num_points, 2]) ) sampling_value_l_ = F.grid_sample( value_l_, sampling_grid_l_, mode="bilinear", padding_mode="zeros", align_corners=False, ) sampling_value_list.append(sampling_value_l_) attention_weights = attention_weights.transpose([0, 2, 1, 3, 4]).reshape( [batch_size * num_heads, 1, num_queries, num_levels * num_points] ) output = ( (paddle.stack(sampling_value_list, axis=-2).flatten(-2) * attention_weights) .sum(-1) .reshape([batch_size, num_heads * hidden_dim, num_queries]) ) return output.transpose([0, 2, 1]) class RTDETRMultiscaleDeformableAttention(nn.Layer): """Multiscale deformable attention as proposed in Deformable DETR.""" def __init__(self, config, num_heads, n_points): super().__init__() self.attn = MultiScaleDeformableAttention() self.d_model = config.d_model self.n_levels = config.num_feature_levels self.n_heads = num_heads self.n_points = n_points self.sampling_offsets = nn.Linear( config.d_model, num_heads * self.n_levels * n_points * 2 ) self.attention_weights = nn.Linear( config.d_model, num_heads * self.n_levels * n_points ) self.value_proj = nn.Linear(config.d_model, config.d_model) self.output_proj = nn.Linear(config.d_model, config.d_model) self.im2col_step = 64 def forward( self, hidden_states, attention_mask=None, encoder_hidden_states=None, position_embeddings=None, reference_points=None, spatial_shapes=None, spatial_shapes_list=None, level_start_index=None, ): if position_embeddings is not None: hidden_states = hidden_states + position_embeddings batch_size, num_queries, _ = hidden_states.shape batch_size, sequence_length, _ = encoder_hidden_states.shape value = self.value_proj(encoder_hidden_states) if attention_mask is not None: value = paddle.where( attention_mask.unsqueeze(-1).astype("bool"), value, paddle.zeros_like(value), ) value = value.reshape( [batch_size, sequence_length, self.n_heads, self.d_model // self.n_heads] ) sampling_offsets = self.sampling_offsets(hidden_states).reshape( [batch_size, num_queries, self.n_heads, self.n_levels, self.n_points, 2] ) attn_weights = self.attention_weights(hidden_states).reshape( [batch_size, num_queries, self.n_heads, self.n_levels * self.n_points] ) attn_weights = F.softmax(attn_weights, axis=-1).reshape( [batch_size, num_queries, self.n_heads, self.n_levels, self.n_points] ) num_coordinates = reference_points.shape[-1] if num_coordinates == 2: offset_normalizer = paddle.stack( [spatial_shapes[..., 1], spatial_shapes[..., 0]], axis=-1 ) sampling_locations = ( reference_points[:, :, None, :, None, :] + sampling_offsets / offset_normalizer[None, None, None, :, None, :] ) elif num_coordinates == 4: sampling_locations = ( reference_points[:, :, None, :, None, :2] + sampling_offsets / self.n_points * reference_points[:, :, None, :, None, 2:] * 0.5 ) else: raise ValueError( f"Last dim of reference_points must be 2 or 4, but got {reference_points.shape[-1]}" ) output = self.attn( value, spatial_shapes, spatial_shapes_list, level_start_index, sampling_locations, attn_weights, self.im2col_step, ) output = self.output_proj(output) return output, attn_weights class RTDETRDecoderLayer(nn.Layer): def __init__(self, config): super().__init__() self.hidden_size = config.d_model self.self_attn = RTDETRSelfAttention( config=config, hidden_size=self.hidden_size, num_attention_heads=config.decoder_attention_heads, dropout=config.attention_dropout, ) self.dropout = config.dropout self.self_attn_layer_norm = nn.LayerNorm( self.hidden_size, epsilon=config.layer_norm_eps ) self.encoder_attn = RTDETRMultiscaleDeformableAttention( config, num_heads=config.decoder_attention_heads, n_points=config.decoder_n_points, ) self.encoder_attn_layer_norm = nn.LayerNorm( self.hidden_size, epsilon=config.layer_norm_eps ) self.mlp = RTDETTMLP( config, self.hidden_size, config.decoder_ffn_dim, config.decoder_activation_function, ) self.final_layer_norm = nn.LayerNorm( self.hidden_size, epsilon=config.layer_norm_eps ) def forward( self, hidden_states, object_queries_position_embeddings=None, reference_points=None, spatial_shapes=None, spatial_shapes_list=None, level_start_index=None, encoder_hidden_states=None, encoder_attention_mask=None, ): residual = hidden_states # Self Attention hidden_states, _ = self.self_attn( hidden_states=hidden_states, attention_mask=encoder_attention_mask, position_embeddings=object_queries_position_embeddings, ) hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states hidden_states = self.self_attn_layer_norm(hidden_states) residual = hidden_states # Cross-Attention hidden_states, _ = self.encoder_attn( hidden_states=hidden_states, encoder_hidden_states=encoder_hidden_states, position_embeddings=object_queries_position_embeddings, reference_points=reference_points, spatial_shapes=spatial_shapes, spatial_shapes_list=spatial_shapes_list, level_start_index=level_start_index, ) hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states hidden_states = self.encoder_attn_layer_norm(hidden_states) # Fully Connected residual = hidden_states hidden_states = self.mlp(hidden_states) hidden_states = residual + hidden_states hidden_states = self.final_layer_norm(hidden_states) return hidden_states class RTDETRDecoder(nn.Layer): def __init__(self, config): super().__init__() self.dropout = config.dropout self.layers = nn.LayerList( [RTDETRDecoderLayer(config) for _ in range(config.decoder_layers)] ) self.query_pos_head = RTDETRMLPPredictionHead( 4, 2 * config.d_model, config.d_model, num_layers=2 ) # Per-layer prediction heads (with_box_refine=True) self.class_embed = nn.LayerList( [ nn.Linear(config.d_model, config.num_labels) for _ in range(config.decoder_layers) ] ) self.bbox_embed = nn.LayerList( [ RTDETRMLPPredictionHead(config.d_model, config.d_model, 4, num_layers=3) for _ in range(config.decoder_layers) ] ) self.num_queries = config.num_queries def forward( self, inputs_embeds=None, encoder_hidden_states=None, encoder_attention_mask=None, reference_points=None, spatial_shapes=None, spatial_shapes_list=None, level_start_index=None, ): hidden_states = inputs_embeds intermediate_reference_points = [] intermediate_logits = [] reference_points = F.sigmoid(reference_points) for idx, decoder_layer in enumerate(self.layers): reference_points_input = reference_points.unsqueeze(2) object_queries_position_embeddings = self.query_pos_head(reference_points) hidden_states = decoder_layer( hidden_states, object_queries_position_embeddings=object_queries_position_embeddings, encoder_hidden_states=encoder_hidden_states, reference_points=reference_points_input, spatial_shapes=spatial_shapes, spatial_shapes_list=spatial_shapes_list, level_start_index=level_start_index, encoder_attention_mask=encoder_attention_mask, ) # Per-layer bbox refinement predicted_corners = self.bbox_embed[idx](hidden_states) new_reference_points = F.sigmoid( predicted_corners + inverse_sigmoid(reference_points) ) reference_points = new_reference_points.detach() intermediate_reference_points.append(new_reference_points) intermediate_logits.append(self.class_embed[idx](hidden_states)) intermediate_reference_points = paddle.stack( intermediate_reference_points, axis=1 ) intermediate_logits = paddle.stack(intermediate_logits, axis=1) return { "intermediate_logits": intermediate_logits, "intermediate_reference_points": intermediate_reference_points, } def replace_batch_norm(model): """Recursively replace all nn.BatchNorm2D with RTDETRFrozenBatchNorm2d.""" for name, module in model.named_children(): if isinstance(module, nn.BatchNorm2D): new_module = RTDETRFrozenBatchNorm2d(module._num_features) new_module.weight.set_value(module.weight) new_module.bias.set_value(module.bias) new_module._mean.set_value(module._mean) new_module._variance.set_value(module._variance) setattr(model, name, new_module) if len(list(module.children())) > 0: replace_batch_norm(module) class RTDETRConvEncoder(nn.Layer): """Convolutional backbone using HGNetV2Backbone.""" def __init__(self, config): super().__init__() backbone = HGNetV2Backbone(config.backbone_config) if config.freeze_backbone_batch_norms: with paddle.no_grad(): replace_batch_norm(backbone) self.model = backbone self.intermediate_channel_sizes = backbone.out_channels def forward(self, pixel_values): features = self.model(pixel_values) return features class RTDETRModel(nn.Layer): def __init__(self, config): super().__init__() self.config = config # Create backbone self.backbone = RTDETRConvEncoder(config) # Create encoder input projection layers (channels from config, not backbone) encoder_input_proj_list = [] for in_channels in config.encoder_in_channels: encoder_input_proj_list.append( nn.Sequential( nn.Conv2D( in_channels, config.encoder_hidden_dim, kernel_size=1, bias_attr=False, ), nn.BatchNorm2D(config.encoder_hidden_dim), ) ) self.encoder_input_proj = nn.LayerList(encoder_input_proj_list) # Create encoder self.encoder = RTDETRHybridEncoder(config) # Denoising embedding (HF convention: num_labels+1 with padding_idx) self.denoising_class_embed = nn.Embedding( config.num_labels + 1, config.d_model, padding_idx=config.num_labels ) # Decoder embedding if config.learn_initial_query: self.weight_embedding = nn.Embedding(config.num_queries, config.d_model) # Encoder head self.enc_output = nn.Sequential( nn.Linear(config.d_model, config.d_model), nn.LayerNorm(config.d_model, epsilon=config.layer_norm_eps), ) self.enc_score_head = nn.Linear(config.d_model, config.num_labels) self.enc_bbox_head = RTDETRMLPPredictionHead( config.d_model, config.d_model, 4, num_layers=3 ) # Create decoder input projection layers num_backbone_outs = len(config.decoder_in_channels) decoder_input_proj_list = [] for i in range(num_backbone_outs): in_channels = config.decoder_in_channels[i] decoder_input_proj_list.append( nn.Sequential( nn.Conv2D( in_channels, config.d_model, kernel_size=1, bias_attr=False ), nn.BatchNorm2D(config.d_model, epsilon=config.batch_norm_eps), ) ) for _ in range(config.num_feature_levels - num_backbone_outs): decoder_input_proj_list.append( nn.Sequential( nn.Conv2D( in_channels, config.d_model, kernel_size=3, stride=2, padding=1, bias_attr=False, ), nn.BatchNorm2D(config.d_model, epsilon=config.batch_norm_eps), ) ) in_channels = config.d_model self.decoder_input_proj = nn.LayerList(decoder_input_proj_list) # Decoder self.decoder = RTDETRDecoder(config) def generate_anchors(self, spatial_shapes, grid_size=0.05, dtype="float32"): anchors = [] for level, (height, width) in enumerate(spatial_shapes): grid_y, grid_x = paddle.meshgrid( paddle.arange(end=height).astype(dtype), paddle.arange(end=width).astype(dtype), ) grid_xy = paddle.stack([grid_x, grid_y], axis=-1) grid_xy = grid_xy.unsqueeze(0) + 0.5 grid_xy[..., 0] /= width grid_xy[..., 1] /= height wh = paddle.ones_like(grid_xy) * grid_size * (2.0**level) anchors.append( paddle.concat([grid_xy, wh], axis=-1).reshape([-1, height * width, 4]) ) eps = 1e-2 anchors = paddle.concat(anchors, axis=1) valid_mask = ((anchors > eps) & (anchors < 1 - eps)).all(axis=-1, keepdim=True) anchors = paddle.log(anchors / (1 - anchors)) anchors = paddle.where( valid_mask, anchors, paddle.full_like(anchors, float("inf")) ) return anchors, valid_mask def forward(self, pixel_values, pixel_mask=None): batch_size = pixel_values.shape[0] # Backbone — may return more stages than encoder needs; take last N features = self.backbone(pixel_values) num_proj = len(self.encoder_input_proj) features = features[-num_proj:] proj_feats = [ self.encoder_input_proj[level](source) for level, source in enumerate(features) ] # Encoder (hybrid FPN/PAN) encoder_outputs = self.encoder(proj_feats) # Prepare decoder inputs sources = [] for level, source in enumerate(encoder_outputs): sources.append(self.decoder_input_proj[level](source)) if self.config.num_feature_levels > len(sources): _len_sources = len(sources) sources.append(self.decoder_input_proj[_len_sources](encoder_outputs[-1])) for i in range(_len_sources + 1, self.config.num_feature_levels): sources.append(self.decoder_input_proj[i](encoder_outputs[-1])) # Flatten sources for decoder source_flatten = [] spatial_shapes_list = [] spatial_shapes = paddle.empty([len(sources), 2], dtype="int64") for level, source in enumerate(sources): h, w = source.shape[-2:] spatial_shapes[level, 0] = h spatial_shapes[level, 1] = w spatial_shapes_list.append((h, w)) source = source.flatten(2).transpose([0, 2, 1]) source_flatten.append(source) source_flatten = paddle.concat(source_flatten, axis=1) level_start_index = paddle.concat( [paddle.zeros([1], dtype="int64"), spatial_shapes.prod(1).cumsum(0)[:-1]] ) dtype = source_flatten.dtype spatial_shapes_tuple = tuple(spatial_shapes_list) anchors, valid_mask = self.generate_anchors(spatial_shapes_tuple, dtype=dtype) memory = valid_mask.astype(source_flatten.dtype) * source_flatten output_memory = self.enc_output(memory) enc_outputs_class = self.enc_score_head(output_memory) enc_outputs_coord_logits = self.enc_bbox_head(output_memory) + anchors _, topk_ind = paddle.topk( enc_outputs_class.max(-1), self.config.num_queries, axis=1 ) reference_points_unact = paddle.take_along_axis( enc_outputs_coord_logits, topk_ind.unsqueeze(-1).expand([-1, -1, enc_outputs_coord_logits.shape[-1]]), axis=1, ) enc_topk_bboxes = F.sigmoid(reference_points_unact) if self.config.learn_initial_query: target = self.weight_embedding.weight.unsqueeze(0).expand( [batch_size, -1, -1] ) else: target = paddle.take_along_axis( output_memory, topk_ind.unsqueeze(-1).expand([-1, -1, output_memory.shape[-1]]), axis=1, ) target = target.detach() init_reference_points = reference_points_unact.detach() # Decoder decoder_outputs = self.decoder( inputs_embeds=target, encoder_hidden_states=source_flatten, encoder_attention_mask=None, reference_points=init_reference_points, spatial_shapes=spatial_shapes, spatial_shapes_list=spatial_shapes_list, level_start_index=level_start_index, ) return decoder_outputs class DETRPostProcess(object): __shared__ = ["num_classes", "use_focal_loss"] __inject__ = [] def __init__( self, num_classes=80, num_top_queries=100, use_focal_loss=False, bbox_decode_type="origin", ): super(DETRPostProcess, self).__init__() assert bbox_decode_type in ["origin", "pad"] self.num_classes = num_classes self.num_top_queries = num_top_queries self.use_focal_loss = use_focal_loss self.bbox_decode_type = bbox_decode_type def __call__(self, head_out, im_shape, scale_factor, pad_shape): bboxes, logits = head_out bbox_pred = bbox_cxcywh_to_xyxy(bboxes) origin_shape = paddle.floor(im_shape / scale_factor + 0.5) img_h, img_w = paddle.split(origin_shape, 2, axis=-1) if self.bbox_decode_type == "pad": out_shape = pad_shape / im_shape * origin_shape out_shape = out_shape.flip(1).tile([1, 2]).unsqueeze(1) elif self.bbox_decode_type == "origin": out_shape = origin_shape.flip(1).tile([1, 2]).unsqueeze(1) else: raise Exception(f"Wrong `bbox_decode_type`: {self.bbox_decode_type}.") bbox_pred *= out_shape scores = ( F.sigmoid(logits) if self.use_focal_loss else F.softmax(logits)[:, :, :-1] ) if not self.use_focal_loss: scores, labels = scores.max(-1), scores.argmax(-1) if scores.shape[1] > self.num_top_queries: scores, index = paddle.topk(scores, self.num_top_queries, axis=-1) batch_ind = ( paddle.arange(end=scores.shape[0]) .unsqueeze(-1) .tile([1, self.num_top_queries]) ) index = paddle.stack([batch_ind, index], axis=-1) labels = paddle.gather_nd(labels, index) bbox_pred = paddle.gather_nd(bbox_pred, index) else: scores, index = paddle.topk( scores.flatten(1), self.num_top_queries, axis=-1 ) labels = index % self.num_classes index = index // self.num_classes batch_ind = ( paddle.arange(end=scores.shape[0]) .unsqueeze(-1) .tile([1, self.num_top_queries]) ) index = paddle.stack([batch_ind, index], axis=-1) bbox_pred = paddle.gather_nd(bbox_pred, index) bbox_pred = paddle.concat( [labels.unsqueeze(-1).astype("float32"), scores.unsqueeze(-1), bbox_pred], axis=-1, ) bbox_num = paddle.to_tensor(self.num_top_queries, dtype="int32").tile( [bbox_pred.shape[0]] ) bbox_pred = bbox_pred.reshape([-1, 6]) return bbox_pred, bbox_num class RTDETR(BatchNormHFStateDictMixin, PretrainedModel): config_class = RTDETRConfig _keys_to_ignore_on_load_unexpected = ["num_batches_tracked"] def __init__(self, config): super(RTDETR, self).__init__(config) self.config = config self.model = RTDETRModel(config) self.post_process = DETRPostProcess( num_classes=config.num_labels, num_top_queries=config.num_queries, use_focal_loss=config.use_focal_loss, ) def forward(self, inputs): pixel_values = paddle.to_tensor(inputs[1]) im_shape = paddle.to_tensor(inputs[0]) scale_factor = paddle.to_tensor(inputs[2]) decoder_outputs = self.model(pixel_values) intermediate_reference_points = decoder_outputs["intermediate_reference_points"] intermediate_logits = decoder_outputs["intermediate_logits"] # Take last layer outputs pred_boxes = intermediate_reference_points[:, -1] logits = intermediate_logits[:, -1] head_out = (pred_boxes, logits) pad_shape = paddle.to_tensor( [[pixel_values.shape[2], pixel_values.shape[3]]] * pixel_values.shape[0], dtype="float32", ) bbox, bbox_num = self.post_process( head_out, im_shape, scale_factor, pad_shape, ) output = [bbox, bbox_num] return output def get_transpose_weight_keys(self): t_layers = [ "fc", "o_proj", "out_proj", "output_proj", "q_proj", "k_proj", "v_proj", "enc_bbox_head", "enc_output", "query_pos_head", "enc_score_head", "encoder_attn", "decoder.bbox_embed", "decoder.class_embed", "sampling_offsets", "attention_weights", "value_proj", ] keys = [] for key, _ in self.get_hf_state_dict().items(): for t_layer in t_layers: if ( t_layer in key and key.endswith("weight") and "norm" not in key and "LayerNorm" not in key and "layer_norm" not in key and "enc_output.1" not in key ): keys.append(key) return keys def set_hf_state_dict(self, state_dict, *args, **kwargs): converted = _apply_rt_detr_key_conversion(state_dict) return super().set_hf_state_dict(converted, *args, **kwargs) def get_hf_state_dict(self, *args, **kwargs): return _reverse_rt_detr_key_conversion( super().get_hf_state_dict(*args, **kwargs) )