src/zoo/dfine/dfine_decoder.py

"""
D-FINE: Redefine Regression Task of DETRs as Fine-grained Distribution Refinement
Copyright (c) 2024 The D-FINE Authors. All Rights Reserved.
---------------------------------------------------------------------------------
Modified from RT-DETR (https://github.com/lyuwenyu/RT-DETR)
Copyright (c) 2023 lyuwenyu. All Rights Reserved.
"""

import copy
import functools
import math
from collections import OrderedDict
from typing import List

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.init as init

from ...core import register
from .denoising import get_contrastive_denoising_training_group
from .dfine_utils import distance2bbox, weighting_function
from .utils import (
    bias_init_with_prob,
    deformable_attention_core_func_v2,
    get_activation,
    inverse_sigmoid,
)

__all__ = ["DFINETransformer"]


class MLP(nn.Module):
    def __init__(self, input_dim, hidden_dim, output_dim, num_layers, act="relu"):
        super().__init__()
        self.num_layers = num_layers
        h = [hidden_dim] * (num_layers - 1)
        self.layers = nn.ModuleList(
            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
        )
        self.act = get_activation(act)

    def forward(self, x):
        for i, layer in enumerate(self.layers):
            x = self.act(layer(x)) if i < self.num_layers - 1 else layer(x)
        return x


class MSDeformableAttention(nn.Module):
    def __init__(
        self,
        embed_dim=256,
        num_heads=8,
        num_levels=4,
        num_points=4,
        method="default",
        offset_scale=0.5,
    ):
        """Multi-Scale Deformable Attention"""
        super(MSDeformableAttention, self).__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.num_levels = num_levels
        self.offset_scale = offset_scale

        if isinstance(num_points, list):
            assert len(num_points) == num_levels, ""
            num_points_list = num_points
        else:
            num_points_list = [num_points for _ in range(num_levels)]

        self.num_points_list = num_points_list

        num_points_scale = [1 / n for n in num_points_list for _ in range(n)]
        self.register_buffer(
            "num_points_scale", torch.tensor(num_points_scale, dtype=torch.float32)
        )

        self.total_points = num_heads * sum(num_points_list)
        self.method = method

        self.head_dim = embed_dim // num_heads
        assert (
            self.head_dim * num_heads == self.embed_dim
        ), "embed_dim must be divisible by num_heads"

        self.sampling_offsets = nn.Linear(embed_dim, self.total_points * 2)
        self.attention_weights = nn.Linear(embed_dim, self.total_points)

        self.ms_deformable_attn_core = functools.partial(
            deformable_attention_core_func_v2, method=self.method
        )

        self._reset_parameters()

        if method == "discrete":
            for p in self.sampling_offsets.parameters():
                p.requires_grad = False

    def _reset_parameters(self):
        # sampling_offsets
        init.constant_(self.sampling_offsets.weight, 0)
        thetas = torch.arange(self.num_heads, dtype=torch.float32) * (
            2.0 * math.pi / self.num_heads
        )
        grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
        grid_init = grid_init / grid_init.abs().max(-1, keepdim=True).values
        grid_init = grid_init.reshape(self.num_heads, 1, 2).tile([1, sum(self.num_points_list), 1])
        scaling = torch.concat([torch.arange(1, n + 1) for n in self.num_points_list]).reshape(
            1, -1, 1
        )
        grid_init *= scaling
        self.sampling_offsets.bias.data[...] = grid_init.flatten()

        # attention_weights
        init.constant_(self.attention_weights.weight, 0)
        init.constant_(self.attention_weights.bias, 0)

    def forward(
        self,
        query: torch.Tensor,
        reference_points: torch.Tensor,
        value: torch.Tensor,
        value_spatial_shapes: List[int],
    ):
        """
        Args:
            query (Tensor): [bs, query_length, C]
            reference_points (Tensor): [bs, query_length, n_levels, 2], range in [0, 1], top-left (0,0),
                bottom-right (1, 1), including padding area
            value (Tensor): [bs, value_length, C]
            value_spatial_shapes (List): [n_levels, 2], [(H_0, W_0), (H_1, W_1), ..., (H_{L-1}, W_{L-1})]

        Returns:
            output (Tensor): [bs, Length_{query}, C]
        """
        bs, Len_q = query.shape[:2]

        sampling_offsets: torch.Tensor = self.sampling_offsets(query)
        sampling_offsets = sampling_offsets.reshape(
            bs, Len_q, self.num_heads, sum(self.num_points_list), 2
        )

        attention_weights = self.attention_weights(query).reshape(
            bs, Len_q, self.num_heads, sum(self.num_points_list)
        )
        attention_weights = F.softmax(attention_weights, dim=-1)

        if reference_points.shape[-1] == 2:
            offset_normalizer = torch.tensor(value_spatial_shapes)
            offset_normalizer = offset_normalizer.flip([1]).reshape(1, 1, 1, self.num_levels, 1, 2)
            sampling_locations = (
                reference_points.reshape(bs, Len_q, 1, self.num_levels, 1, 2)
                + sampling_offsets / offset_normalizer
            )
        elif reference_points.shape[-1] == 4:
            # reference_points [8, 480, None, 1,  4]
            # sampling_offsets [8, 480, 8,    12, 2]
            num_points_scale = self.num_points_scale.to(dtype=query.dtype).unsqueeze(-1)
            offset = (
                sampling_offsets
                * num_points_scale
                * reference_points[:, :, None, :, 2:]
                * self.offset_scale
            )
            sampling_locations = reference_points[:, :, None, :, :2] + offset
        else:
            raise ValueError(
                "Last dim of reference_points must be 2 or 4, but get {} instead.".format(
                    reference_points.shape[-1]
                )
            )

        output = self.ms_deformable_attn_core(
            value, value_spatial_shapes, sampling_locations, attention_weights, self.num_points_list
        )

        return output


class TransformerDecoderLayer(nn.Module):
    def __init__(
        self,
        d_model=256,
        n_head=8,
        dim_feedforward=1024,
        dropout=0.0,
        activation="relu",
        n_levels=4,
        n_points=4,
        cross_attn_method="default",
        layer_scale=None,
    ):
        super(TransformerDecoderLayer, self).__init__()
        if layer_scale is not None:
            dim_feedforward = round(layer_scale * dim_feedforward)
            d_model = round(layer_scale * d_model)

        # self attention
        self.self_attn = nn.MultiheadAttention(d_model, n_head, dropout=dropout, batch_first=True)
        self.dropout1 = nn.Dropout(dropout)
        self.norm1 = nn.LayerNorm(d_model)

        # cross attention
        self.cross_attn = MSDeformableAttention(
            d_model, n_head, n_levels, n_points, method=cross_attn_method
        )
        self.dropout2 = nn.Dropout(dropout)

        # gate
        self.gateway = Gate(d_model)

        # ffn
        self.linear1 = nn.Linear(d_model, dim_feedforward)
        self.activation = get_activation(activation)
        self.dropout3 = nn.Dropout(dropout)
        self.linear2 = nn.Linear(dim_feedforward, d_model)
        self.dropout4 = nn.Dropout(dropout)
        self.norm3 = nn.LayerNorm(d_model)

        self._reset_parameters()

    def _reset_parameters(self):
        init.xavier_uniform_(self.linear1.weight)
        init.xavier_uniform_(self.linear2.weight)

    def with_pos_embed(self, tensor, pos):
        return tensor if pos is None else tensor + pos

    def forward_ffn(self, tgt):
        return self.linear2(self.dropout3(self.activation(self.linear1(tgt))))

    def forward(
        self, target, reference_points, value, spatial_shapes, attn_mask=None, query_pos_embed=None
    ):
        # self attention
        q = k = self.with_pos_embed(target, query_pos_embed)

        target2, _ = self.self_attn(q, k, value=target, attn_mask=attn_mask)
        target = target + self.dropout1(target2)
        target = self.norm1(target)

        # cross attention
        target2 = self.cross_attn(
            self.with_pos_embed(target, query_pos_embed), reference_points, value, spatial_shapes
        )

        target = self.gateway(target, self.dropout2(target2))

        # ffn
        target2 = self.forward_ffn(target)
        target = target + self.dropout4(target2)
        target = self.norm3(target.clamp(min=-65504, max=65504))

        return target


class Gate(nn.Module):
    def __init__(self, d_model):
        super(Gate, self).__init__()
        self.gate = nn.Linear(2 * d_model, 2 * d_model)
        bias = bias_init_with_prob(0.5)
        init.constant_(self.gate.bias, bias)
        init.constant_(self.gate.weight, 0)
        self.norm = nn.LayerNorm(d_model)

    def forward(self, x1, x2):
        gate_input = torch.cat([x1, x2], dim=-1)
        gates = torch.sigmoid(self.gate(gate_input))
        gate1, gate2 = gates.chunk(2, dim=-1)
        return self.norm(gate1 * x1 + gate2 * x2)


class Integral(nn.Module):
    """
    A static layer that calculates integral results from a distribution.

    This layer computes the target location using the formula: `sum{Pr(n) * W(n)}`,
    where Pr(n) is the softmax probability vector representing the discrete
    distribution, and W(n) is the non-uniform Weighting Function.

    Args:
        reg_max (int): Max number of the discrete bins. Default is 32.
                       It can be adjusted based on the dataset or task requirements.
    """

    def __init__(self, reg_max=32):
        super(Integral, self).__init__()
        self.reg_max = reg_max

    def forward(self, x, project):
        shape = x.shape
        x = F.softmax(x.reshape(-1, self.reg_max + 1), dim=1)
        x = F.linear(x, project.to(x.device)).reshape(-1, 4)
        return x.reshape(list(shape[:-1]) + [-1])


class LQE(nn.Module):
    def __init__(self, k, hidden_dim, num_layers, reg_max):
        super(LQE, self).__init__()
        self.k = k
        self.reg_max = reg_max
        self.reg_conf = MLP(4 * (k + 1), hidden_dim, 1, num_layers)
        init.constant_(self.reg_conf.layers[-1].bias, 0)
        init.constant_(self.reg_conf.layers[-1].weight, 0)

    def forward(self, scores, pred_corners):
        B, L, _ = pred_corners.size()
        prob = F.softmax(pred_corners.reshape(B, L, 4, self.reg_max + 1), dim=-1)
        prob_topk, _ = prob.topk(self.k, dim=-1)
        stat = torch.cat([prob_topk, prob_topk.mean(dim=-1, keepdim=True)], dim=-1)
        quality_score = self.reg_conf(stat.reshape(B, L, -1))
        return scores + quality_score


class TransformerDecoder(nn.Module):
    """
    Transformer Decoder implementing Fine-grained Distribution Refinement (FDR).

    This decoder refines object detection predictions through iterative updates across multiple layers,
    utilizing attention mechanisms, location quality estimators, and distribution refinement techniques
    to improve bounding box accuracy and robustness.
    """

    def __init__(
        self,
        hidden_dim,
        decoder_layer,
        decoder_layer_wide,
        num_layers,
        num_head,
        reg_max,
        reg_scale,
        up,
        eval_idx=-1,
        layer_scale=2,
    ):
        super(TransformerDecoder, self).__init__()
        self.hidden_dim = hidden_dim
        self.num_layers = num_layers
        self.layer_scale = layer_scale
        self.num_head = num_head
        self.eval_idx = eval_idx if eval_idx >= 0 else num_layers + eval_idx
        self.up, self.reg_scale, self.reg_max = up, reg_scale, reg_max
        self.layers = nn.ModuleList(
            [copy.deepcopy(decoder_layer) for _ in range(self.eval_idx + 1)]
            + [copy.deepcopy(decoder_layer_wide) for _ in range(num_layers - self.eval_idx - 1)]
        )
        self.lqe_layers = nn.ModuleList(
            [copy.deepcopy(LQE(4, 64, 2, reg_max)) for _ in range(num_layers)]
        )

    def value_op(self, memory, value_proj, value_scale, memory_mask, memory_spatial_shapes):
        """
        Preprocess values for MSDeformableAttention.
        """
        value = value_proj(memory) if value_proj is not None else memory
        value = F.interpolate(memory, size=value_scale) if value_scale is not None else value
        if memory_mask is not None:
            value = value * memory_mask.to(value.dtype).unsqueeze(-1)
        value = value.reshape(value.shape[0], value.shape[1], self.num_head, -1)
        split_shape = [h * w for h, w in memory_spatial_shapes]
        return value.permute(0, 2, 3, 1).split(split_shape, dim=-1)

    def convert_to_deploy(self):
        self.project = weighting_function(self.reg_max, self.up, self.reg_scale, deploy=True)
        self.layers = self.layers[: self.eval_idx + 1]
        self.lqe_layers = nn.ModuleList(
            [nn.Identity()] * (self.eval_idx) + [self.lqe_layers[self.eval_idx]]
        )

    def forward(
        self,
        target,
        ref_points_unact,
        memory,
        spatial_shapes,
        bbox_head,
        score_head,
        query_pos_head,
        pre_bbox_head,
        integral,
        up,
        reg_scale,
        attn_mask=None,
        memory_mask=None,
        dn_meta=None,
    ):
        output = target
        output_detach = pred_corners_undetach = 0
        value = self.value_op(memory, None, None, memory_mask, spatial_shapes)

        dec_out_bboxes = []
        dec_out_logits = []
        dec_out_pred_corners = []
        dec_out_refs = []
        if not hasattr(self, "project"):
            project = weighting_function(self.reg_max, up, reg_scale)
        else:
            project = self.project

        ref_points_detach = F.sigmoid(ref_points_unact)

        for i, layer in enumerate(self.layers):
            ref_points_input = ref_points_detach.unsqueeze(2)
            query_pos_embed = query_pos_head(ref_points_detach).clamp(min=-10, max=10)

            # TODO Adjust scale if needed for detachable wider layers
            if i >= self.eval_idx + 1 and self.layer_scale > 1:
                query_pos_embed = F.interpolate(query_pos_embed, scale_factor=self.layer_scale)
                value = self.value_op(
                    memory, None, query_pos_embed.shape[-1], memory_mask, spatial_shapes
                )
                output = F.interpolate(output, size=query_pos_embed.shape[-1])
                output_detach = output.detach()

            output = layer(
                output, ref_points_input, value, spatial_shapes, attn_mask, query_pos_embed
            )

            if i == 0:
                # Initial bounding box predictions with inverse sigmoid refinement
                pre_bboxes = F.sigmoid(pre_bbox_head(output) + inverse_sigmoid(ref_points_detach))
                pre_scores = score_head[0](output)
                ref_points_initial = pre_bboxes.detach()

            # Refine bounding box corners using FDR, integrating previous layer's corrections
            pred_corners = bbox_head[i](output + output_detach) + pred_corners_undetach
            inter_ref_bbox = distance2bbox(
                ref_points_initial, integral(pred_corners, project), reg_scale
            )

            if self.training or i == self.eval_idx:
                scores = score_head[i](output)
                # Lqe does not affect the performance here.
                scores = self.lqe_layers[i](scores, pred_corners)
                dec_out_logits.append(scores)
                dec_out_bboxes.append(inter_ref_bbox)
                dec_out_pred_corners.append(pred_corners)
                dec_out_refs.append(ref_points_initial)

                if not self.training:
                    break

            pred_corners_undetach = pred_corners
            ref_points_detach = inter_ref_bbox.detach()
            output_detach = output.detach()

        return (
            torch.stack(dec_out_bboxes),
            torch.stack(dec_out_logits),
            torch.stack(dec_out_pred_corners),
            torch.stack(dec_out_refs),
            pre_bboxes,
            pre_scores,
        )


@register()
class DFINETransformer(nn.Module):
    __share__ = ["num_classes", "eval_spatial_size"]

    def __init__(
        self,
        num_classes=80,
        hidden_dim=256,
        num_queries=300,
        feat_channels=[512, 1024, 2048],
        feat_strides=[8, 16, 32],
        num_levels=3,
        num_points=4,
        nhead=8,
        num_layers=6,
        dim_feedforward=1024,
        dropout=0.0,
        activation="relu",
        num_denoising=100,
        label_noise_ratio=0.5,
        box_noise_scale=1.0,
        learn_query_content=False,
        eval_spatial_size=None,
        eval_idx=-1,
        eps=1e-2,
        aux_loss=True,
        cross_attn_method="default",
        query_select_method="default",
        reg_max=32,
        reg_scale=4.0,
        layer_scale=1,
    ):
        super().__init__()
        assert len(feat_channels) <= num_levels
        assert len(feat_strides) == len(feat_channels)

        for _ in range(num_levels - len(feat_strides)):
            feat_strides.append(feat_strides[-1] * 2)

        self.hidden_dim = hidden_dim
        scaled_dim = round(layer_scale * hidden_dim)
        self.nhead = nhead
        self.feat_strides = feat_strides
        self.num_levels = num_levels
        self.num_classes = num_classes
        self.num_queries = num_queries
        self.eps = eps
        self.num_layers = num_layers
        self.eval_spatial_size = eval_spatial_size
        self.aux_loss = aux_loss
        self.reg_max = reg_max

        assert query_select_method in ("default", "one2many", "agnostic"), ""
        assert cross_attn_method in ("default", "discrete"), ""
        self.cross_attn_method = cross_attn_method
        self.query_select_method = query_select_method

        # backbone feature projection
        self._build_input_proj_layer(feat_channels)

        # Transformer module
        self.up = nn.Parameter(torch.tensor([0.5]), requires_grad=False)
        self.reg_scale = nn.Parameter(torch.tensor([reg_scale]), requires_grad=False)
        decoder_layer = TransformerDecoderLayer(
            hidden_dim,
            nhead,
            dim_feedforward,
            dropout,
            activation,
            num_levels,
            num_points,
            cross_attn_method=cross_attn_method,
        )
        decoder_layer_wide = TransformerDecoderLayer(
            hidden_dim,
            nhead,
            dim_feedforward,
            dropout,
            activation,
            num_levels,
            num_points,
            cross_attn_method=cross_attn_method,
            layer_scale=layer_scale,
        )
        self.decoder = TransformerDecoder(
            hidden_dim,
            decoder_layer,
            decoder_layer_wide,
            num_layers,
            nhead,
            reg_max,
            self.reg_scale,
            self.up,
            eval_idx,
            layer_scale,
        )
        # denoising
        self.num_denoising = num_denoising
        self.label_noise_ratio = label_noise_ratio
        self.box_noise_scale = box_noise_scale
        if num_denoising > 0:
            self.denoising_class_embed = nn.Embedding(
                num_classes + 1, hidden_dim, padding_idx=num_classes
            )
            init.normal_(self.denoising_class_embed.weight[:-1])

        # decoder embedding
        self.learn_query_content = learn_query_content
        if learn_query_content:
            self.tgt_embed = nn.Embedding(num_queries, hidden_dim)
        self.query_pos_head = MLP(4, 2 * hidden_dim, hidden_dim, 2)

        # if num_select_queries != self.num_queries:
        #     layer = TransformerEncoderLayer(hidden_dim, nhead, dim_feedforward, activation='gelu')
        #     self.encoder = TransformerEncoder(layer, 1)

        self.enc_output = nn.Sequential(
            OrderedDict(
                [
                    ("proj", nn.Linear(hidden_dim, hidden_dim)),
                    (
                        "norm",
                        nn.LayerNorm(
                            hidden_dim,
                        ),
                    ),
                ]
            )
        )

        if query_select_method == "agnostic":
            self.enc_score_head = nn.Linear(hidden_dim, 1)
        else:
            self.enc_score_head = nn.Linear(hidden_dim, num_classes)

        self.enc_bbox_head = MLP(hidden_dim, hidden_dim, 4, 3)

        # decoder head
        self.eval_idx = eval_idx if eval_idx >= 0 else num_layers + eval_idx
        self.dec_score_head = nn.ModuleList(
            [nn.Linear(hidden_dim, num_classes) for _ in range(self.eval_idx + 1)]
            + [nn.Linear(scaled_dim, num_classes) for _ in range(num_layers - self.eval_idx - 1)]
        )
        self.pre_bbox_head = MLP(hidden_dim, hidden_dim, 4, 3)
        self.dec_bbox_head = nn.ModuleList(
            [
                MLP(hidden_dim, hidden_dim, 4 * (self.reg_max + 1), 3)
                for _ in range(self.eval_idx + 1)
            ]
            + [
                MLP(scaled_dim, scaled_dim, 4 * (self.reg_max + 1), 3)
                for _ in range(num_layers - self.eval_idx - 1)
            ]
        )
        self.integral = Integral(self.reg_max)

        # init encoder output anchors and valid_mask
        if self.eval_spatial_size:
            anchors, valid_mask = self._generate_anchors()
            self.register_buffer("anchors", anchors)
            self.register_buffer("valid_mask", valid_mask)
        # init encoder output anchors and valid_mask
        if self.eval_spatial_size:
            self.anchors, self.valid_mask = self._generate_anchors()

        self._reset_parameters(feat_channels)

    def convert_to_deploy(self):
        self.dec_score_head = nn.ModuleList(
            [nn.Identity()] * (self.eval_idx) + [self.dec_score_head[self.eval_idx]]
        )
        self.dec_bbox_head = nn.ModuleList(
            [
                self.dec_bbox_head[i] if i <= self.eval_idx else nn.Identity()
                for i in range(len(self.dec_bbox_head))
            ]
        )

    def _reset_parameters(self, feat_channels):
        bias = bias_init_with_prob(0.01)
        init.constant_(self.enc_score_head.bias, bias)
        init.constant_(self.enc_bbox_head.layers[-1].weight, 0)
        init.constant_(self.enc_bbox_head.layers[-1].bias, 0)

        init.constant_(self.pre_bbox_head.layers[-1].weight, 0)
        init.constant_(self.pre_bbox_head.layers[-1].bias, 0)

        for cls_, reg_ in zip(self.dec_score_head, self.dec_bbox_head):
            init.constant_(cls_.bias, bias)
            if hasattr(reg_, "layers"):
                init.constant_(reg_.layers[-1].weight, 0)
                init.constant_(reg_.layers[-1].bias, 0)

        init.xavier_uniform_(self.enc_output[0].weight)
        if self.learn_query_content:
            init.xavier_uniform_(self.tgt_embed.weight)
        init.xavier_uniform_(self.query_pos_head.layers[0].weight)
        init.xavier_uniform_(self.query_pos_head.layers[1].weight)
        for m, in_channels in zip(self.input_proj, feat_channels):
            if in_channels != self.hidden_dim:
                init.xavier_uniform_(m[0].weight)

    def _build_input_proj_layer(self, feat_channels):
        self.input_proj = nn.ModuleList()
        for in_channels in feat_channels:
            if in_channels == self.hidden_dim:
                self.input_proj.append(nn.Identity())
            else:
                self.input_proj.append(
                    nn.Sequential(
                        OrderedDict(
                            [
                                ("conv", nn.Conv2d(in_channels, self.hidden_dim, 1, bias=False)),
                                (
                                    "norm",
                                    nn.BatchNorm2d(
                                        self.hidden_dim,
                                    ),
                                ),
                            ]
                        )
                    )
                )

        in_channels = feat_channels[-1]

        for _ in range(self.num_levels - len(feat_channels)):
            if in_channels == self.hidden_dim:
                self.input_proj.append(nn.Identity())
            else:
                self.input_proj.append(
                    nn.Sequential(
                        OrderedDict(
                            [
                                (
                                    "conv",
                                    nn.Conv2d(
                                        in_channels, self.hidden_dim, 3, 2, padding=1, bias=False
                                    ),
                                ),
                                ("norm", nn.BatchNorm2d(self.hidden_dim)),
                            ]
                        )
                    )
                )
                in_channels = self.hidden_dim

    def _get_encoder_input(self, feats: List[torch.Tensor]):
        # get projection features
        proj_feats = [self.input_proj[i](feat) for i, feat in enumerate(feats)]
        if self.num_levels > len(proj_feats):
            len_srcs = len(proj_feats)
            for i in range(len_srcs, self.num_levels):
                if i == len_srcs:
                    proj_feats.append(self.input_proj[i](feats[-1]))
                else:
                    proj_feats.append(self.input_proj[i](proj_feats[-1]))

        # get encoder inputs
        feat_flatten = []
        spatial_shapes = []
        for i, feat in enumerate(proj_feats):
            _, _, h, w = feat.shape
            # [b, c, h, w] -> [b, h*w, c]
            feat_flatten.append(feat.flatten(2).permute(0, 2, 1))
            # [num_levels, 2]
            spatial_shapes.append([h, w])

        # [b, l, c]
        feat_flatten = torch.concat(feat_flatten, 1)
        return feat_flatten, spatial_shapes

    def _generate_anchors(
        self, spatial_shapes=None, grid_size=0.05, dtype=torch.float32, device="cpu"
    ):
        if spatial_shapes is None:
            spatial_shapes = []
            eval_h, eval_w = self.eval_spatial_size
            for s in self.feat_strides:
                spatial_shapes.append([int(eval_h / s), int(eval_w / s)])

        anchors = []
        for lvl, (h, w) in enumerate(spatial_shapes):
            grid_y, grid_x = torch.meshgrid(torch.arange(h), torch.arange(w), indexing="ij")
            grid_xy = torch.stack([grid_x, grid_y], dim=-1)
            grid_xy = (grid_xy.unsqueeze(0) + 0.5) / torch.tensor([w, h], dtype=dtype)
            wh = torch.ones_like(grid_xy) * grid_size * (2.0**lvl)
            lvl_anchors = torch.concat([grid_xy, wh], dim=-1).reshape(-1, h * w, 4)
            anchors.append(lvl_anchors)

        anchors = torch.concat(anchors, dim=1).to(device)
        valid_mask = ((anchors > self.eps) * (anchors < 1 - self.eps)).all(-1, keepdim=True)
        anchors = torch.log(anchors / (1 - anchors))
        anchors = torch.where(valid_mask, anchors, torch.inf)

        return anchors, valid_mask

    def _get_decoder_input(
        self, memory: torch.Tensor, spatial_shapes, denoising_logits=None, denoising_bbox_unact=None
    ):
        # prepare input for decoder
        if self.training or self.eval_spatial_size is None:
            anchors, valid_mask = self._generate_anchors(spatial_shapes, device=memory.device)
        else:
            anchors = self.anchors
            valid_mask = self.valid_mask
        if memory.shape[0] > 1:
            anchors = anchors.repeat(memory.shape[0], 1, 1)

        # memory = torch.where(valid_mask, memory, 0)
        # TODO fix type error for onnx export
        memory = valid_mask.to(memory.dtype) * memory

        output_memory: torch.Tensor = self.enc_output(memory)
        enc_outputs_logits: torch.Tensor = self.enc_score_head(output_memory)

        enc_topk_bboxes_list, enc_topk_logits_list = [], []
        enc_topk_memory, enc_topk_logits, enc_topk_anchors = self._select_topk(
            output_memory, enc_outputs_logits, anchors, self.num_queries
        )

        enc_topk_bbox_unact: torch.Tensor = self.enc_bbox_head(enc_topk_memory) + enc_topk_anchors

        if self.training:
            enc_topk_bboxes = F.sigmoid(enc_topk_bbox_unact)
            enc_topk_bboxes_list.append(enc_topk_bboxes)
            enc_topk_logits_list.append(enc_topk_logits)

        # if self.num_select_queries != self.num_queries:
        #     raise NotImplementedError('')

        if self.learn_query_content:
            content = self.tgt_embed.weight.unsqueeze(0).tile([memory.shape[0], 1, 1])
        else:
            content = enc_topk_memory.detach()

        enc_topk_bbox_unact = enc_topk_bbox_unact.detach()

        if denoising_bbox_unact is not None:
            enc_topk_bbox_unact = torch.concat([denoising_bbox_unact, enc_topk_bbox_unact], dim=1)
            content = torch.concat([denoising_logits, content], dim=1)

        return content, enc_topk_bbox_unact, enc_topk_bboxes_list, enc_topk_logits_list

    def _select_topk(
        self,
        memory: torch.Tensor,
        outputs_logits: torch.Tensor,
        outputs_anchors_unact: torch.Tensor,
        topk: int,
    ):
        if self.query_select_method == "default":
            _, topk_ind = torch.topk(outputs_logits.max(-1).values, topk, dim=-1)

        elif self.query_select_method == "one2many":
            _, topk_ind = torch.topk(outputs_logits.flatten(1), topk, dim=-1)
            topk_ind = topk_ind // self.num_classes

        elif self.query_select_method == "agnostic":
            _, topk_ind = torch.topk(outputs_logits.squeeze(-1), topk, dim=-1)

        topk_ind: torch.Tensor

        topk_anchors = outputs_anchors_unact.gather(
            dim=1, index=topk_ind.unsqueeze(-1).repeat(1, 1, outputs_anchors_unact.shape[-1])
        )

        topk_logits = (
            outputs_logits.gather(
                dim=1, index=topk_ind.unsqueeze(-1).repeat(1, 1, outputs_logits.shape[-1])
            )
            if self.training
            else None
        )

        topk_memory = memory.gather(
            dim=1, index=topk_ind.unsqueeze(-1).repeat(1, 1, memory.shape[-1])
        )

        return topk_memory, topk_logits, topk_anchors

    def forward(self, feats, targets=None):
        # input projection and embedding
        memory, spatial_shapes = self._get_encoder_input(feats)

        # prepare denoising training
        if self.training and self.num_denoising > 0:
            denoising_logits, denoising_bbox_unact, attn_mask, dn_meta = (
                get_contrastive_denoising_training_group(
                    targets,
                    self.num_classes,
                    self.num_queries,
                    self.denoising_class_embed,
                    num_denoising=self.num_denoising,
                    label_noise_ratio=self.label_noise_ratio,
                    box_noise_scale=1.0,
                )
            )
        else:
            denoising_logits, denoising_bbox_unact, attn_mask, dn_meta = None, None, None, None

        init_ref_contents, init_ref_points_unact, enc_topk_bboxes_list, enc_topk_logits_list = (
            self._get_decoder_input(memory, spatial_shapes, denoising_logits, denoising_bbox_unact)
        )

        # decoder
        out_bboxes, out_logits, out_corners, out_refs, pre_bboxes, pre_logits = self.decoder(
            init_ref_contents,
            init_ref_points_unact,
            memory,
            spatial_shapes,
            self.dec_bbox_head,
            self.dec_score_head,
            self.query_pos_head,
            self.pre_bbox_head,
            self.integral,
            self.up,
            self.reg_scale,
            attn_mask=attn_mask,
            dn_meta=dn_meta,
        )

        if self.training and dn_meta is not None:
            dn_pre_logits, pre_logits = torch.split(pre_logits, dn_meta["dn_num_split"], dim=1)
            dn_pre_bboxes, pre_bboxes = torch.split(pre_bboxes, dn_meta["dn_num_split"], dim=1)
            dn_out_bboxes, out_bboxes = torch.split(out_bboxes, dn_meta["dn_num_split"], dim=2)
            dn_out_logits, out_logits = torch.split(out_logits, dn_meta["dn_num_split"], dim=2)

            dn_out_corners, out_corners = torch.split(out_corners, dn_meta["dn_num_split"], dim=2)
            dn_out_refs, out_refs = torch.split(out_refs, dn_meta["dn_num_split"], dim=2)

        if self.training:
            out = {
                "pred_logits": out_logits[-1],
                "pred_boxes": out_bboxes[-1],
                "pred_corners": out_corners[-1],
                "ref_points": out_refs[-1],
                "up": self.up,
                "reg_scale": self.reg_scale,
            }
        else:
            out = {"pred_logits": out_logits[-1], "pred_boxes": out_bboxes[-1]}

        if self.training and self.aux_loss:
            out["aux_outputs"] = self._set_aux_loss2(
                out_logits[:-1],
                out_bboxes[:-1],
                out_corners[:-1],
                out_refs[:-1],
                out_corners[-1],
                out_logits[-1],
            )
            out["enc_aux_outputs"] = self._set_aux_loss(enc_topk_logits_list, enc_topk_bboxes_list)
            out["pre_outputs"] = {"pred_logits": pre_logits, "pred_boxes": pre_bboxes}
            out["enc_meta"] = {"class_agnostic": self.query_select_method == "agnostic"}

            if dn_meta is not None:
                out["dn_outputs"] = self._set_aux_loss2(
                    dn_out_logits,
                    dn_out_bboxes,
                    dn_out_corners,
                    dn_out_refs,
                    dn_out_corners[-1],
                    dn_out_logits[-1],
                )
                out["dn_pre_outputs"] = {"pred_logits": dn_pre_logits, "pred_boxes": dn_pre_bboxes}
                out["dn_meta"] = dn_meta

        return out

    @torch.jit.unused
    def _set_aux_loss(self, outputs_class, outputs_coord):
        # this is a workaround to make torchscript happy, as torchscript
        # doesn't support dictionary with non-homogeneous values, such
        # as a dict having both a Tensor and a list.
        return [{"pred_logits": a, "pred_boxes": b} for a, b in zip(outputs_class, outputs_coord)]

    @torch.jit.unused
    def _set_aux_loss2(
        self,
        outputs_class,
        outputs_coord,
        outputs_corners,
        outputs_ref,
        teacher_corners=None,
        teacher_logits=None,
    ):
        # this is a workaround to make torchscript happy, as torchscript
        # doesn't support dictionary with non-homogeneous values, such
        # as a dict having both a Tensor and a list.
        return [
            {
                "pred_logits": a,
                "pred_boxes": b,
                "pred_corners": c,
                "ref_points": d,
                "teacher_corners": teacher_corners,
                "teacher_logits": teacher_logits,
            }
            for a, b, c, d in zip(outputs_class, outputs_coord, outputs_corners, outputs_ref)
        ]