navsim/agents/vadv2/vadv2_pdm_model_progress_ablate.py

from typing import Dict

import numpy as np
import torch
import torch.nn as nn

from navsim.agents.transfuser.transfuser_backbone import TransfuserBackbone
from navsim.agents.transfuser.transfuser_backbone_conv import TransfuserBackboneConv
from navsim.agents.transfuser.transfuser_backbone_moe import TransfuserBackboneMoe
from navsim.agents.transfuser.transfuser_backbone_moe_ult32 import TransfuserBackboneMoeUlt32
from navsim.agents.transfuser.transfuser_backbone_vit import TransfuserBackboneViT
from navsim.agents.transfuser.transfuser_model import AgentHead
from navsim.agents.utils.attn import MemoryEffTransformer
from navsim.agents.utils.nerf import nerf_positional_encoding
from navsim.agents.vadv2.vadv2_config import Vadv2Config


class Vadv2ModelPDMProgressAblate(nn.Module):
    def __init__(self, config: Vadv2Config):
        super().__init__()

        self._query_splits = [
            config.num_bounding_boxes,
        ]

        self._config = config
        assert config.backbone_type in ['vit', 'intern', 'vov', 'resnet', 'eva', 'moe', 'moe_ult32', 'swin']
        if config.backbone_type == 'vit' or config.backbone_type == 'eva':
            self._backbone = TransfuserBackboneViT(config)
        elif config.backbone_type == 'intern' or config.backbone_type == 'vov' or config.backbone_type == 'swin':
            self._backbone = TransfuserBackboneConv(config)
        elif config.backbone_type == 'moe':
            self._backbone = TransfuserBackboneMoe(config)
        elif config.backbone_type == 'moe_ult32':
            self._backbone = TransfuserBackboneMoeUlt32(config)
        else:
            self._backbone = TransfuserBackbone(config)

        bev_size = config.lidar_vert_anchors * config.lidar_horz_anchors
        bev_c = self._backbone.lidar_encoder.feature_info.info[4]['num_chs']

        self._keyval_embedding = nn.Embedding(
            bev_size, config.tf_d_model
        )  # 8x8 feature grid + trajectory
        self._query_embedding = nn.Embedding(sum(self._query_splits), config.tf_d_model)

        # usually, the BEV features are variable in size.
        self._bev_downscale = nn.Conv2d(bev_c, config.tf_d_model, kernel_size=1)
        # todo drop ego status like plantf
        # assert config.num_ego_status == 1
        # assert not config.use_nerf
        self._status_encoding = nn.Linear((4 + 2 + 2) * config.num_ego_status, config.tf_d_model)

        self._bev_semantic_head = nn.Sequential(
            nn.Conv2d(
                config.bev_features_channels,
                config.bev_features_channels,
                kernel_size=(3, 3),
                stride=1,
                padding=(1, 1),
                bias=True,
            ),
            nn.ReLU(inplace=True),
            nn.Conv2d(
                config.bev_features_channels,
                config.num_bev_classes,
                kernel_size=(1, 1),
                stride=1,
                padding=0,
                bias=True,
            ),
            nn.Upsample(
                size=(config.lidar_resolution_height // 2, config.lidar_resolution_width),
                mode="bilinear",
                align_corners=False,
            ),
        )

        tf_decoder_layer = nn.TransformerDecoderLayer(
            d_model=config.tf_d_model,
            nhead=config.tf_num_head,
            dim_feedforward=config.tf_d_ffn,
            dropout=config.tf_dropout,
            batch_first=True,
        )

        self._tf_decoder = nn.TransformerDecoder(tf_decoder_layer, config.tf_num_layers)
        self._agent_head = AgentHead(
            num_agents=config.num_bounding_boxes,
            d_ffn=config.tf_d_ffn,
            d_model=config.tf_d_model,
        )

        self._trajectory_head = Vadv2HeadPDMProgress(
            num_poses=config.trajectory_sampling.num_poses,
            d_ffn=config.tf_d_ffn,
            d_model=config.tf_d_model,
            nhead=config.vadv2_head_nhead,
            nlayers=config.vadv2_head_nlayers,
            vocab_path=config.vocab_path,
            config=config
        )

    def forward(self, features: Dict[str, torch.Tensor],
                interpolated_traj=None) -> Dict[str, torch.Tensor]:
        # Todo egostatus
        camera_feature: torch.Tensor = features["camera_feature"]
        lidar_feature: torch.Tensor = features["lidar_feature"]
        status_feature: torch.Tensor = features["status_feature"]

        batch_size = status_feature.shape[0]

        bev_feature_upscale, bev_feature, _ = self._backbone(camera_feature, lidar_feature)

        bev_feature = self._bev_downscale(bev_feature).flatten(-2, -1)
        bev_feature = bev_feature.permute(0, 2, 1)

        if self._config.num_ego_status == 1 and status_feature.shape[1] == 32:
            status_encoding = self._status_encoding(status_feature[:, :8])
        else:
            status_encoding = self._status_encoding(status_feature)

        keyval = bev_feature
        keyval += self._keyval_embedding.weight[None, ...]

        query = self._query_embedding.weight[None, ...].repeat(batch_size, 1, 1)
        agents_query = self._tf_decoder(query, keyval)

        bev_semantic_map = self._bev_semantic_head(bev_feature_upscale)

        output: Dict[str, torch.Tensor] = {"bev_semantic_map": bev_semantic_map}
        # 轨迹预测head
        trajectory = self._trajectory_head(keyval, status_encoding, interpolated_traj)
        output.update(trajectory)

        agents = self._agent_head(agents_query)
        output.update(agents)

        return output


class Vadv2HeadPDMProgress(nn.Module):
    def __init__(self, num_poses: int, d_ffn: int, d_model: int, vocab_path: str,
                 nhead: int, nlayers: int, config: Vadv2Config = None
                 ):
        super().__init__()
        self._num_poses = num_poses
        self.transformer = nn.TransformerDecoder(
            nn.TransformerDecoderLayer(
                d_model, nhead, d_ffn,
                dropout=0.0, batch_first=True
            ), nlayers
        )
        self.vocab = nn.Parameter(
            torch.from_numpy(np.load(vocab_path)),
            requires_grad=False
        )

        self.heads = nn.ModuleDict({
            'total': nn.Sequential(
                nn.Linear(d_model, d_ffn),
                nn.ReLU(),
                nn.Linear(d_ffn, 1),
            ),
            'imi': nn.Sequential(
                nn.Linear(d_model, d_ffn),
                nn.ReLU(),
                nn.Linear(d_ffn, d_ffn),
                nn.ReLU(),
                nn.Linear(d_ffn, 1),
            )
        })

        self.inference_imi_weight = config.inference_imi_weight
        self.inference_da_weight = config.inference_da_weight
        self.normalize_vocab_pos = config.normalize_vocab_pos
        if self.normalize_vocab_pos:
            self.encoder = MemoryEffTransformer(
                d_model=d_model,
                nhead=nhead,
                dim_feedforward=d_model * 4,
                dropout=0.0
            )
        self.use_nerf = config.use_nerf

        if self.use_nerf:
            self.pos_embed = nn.Sequential(
                nn.Linear(1040, d_ffn),
                nn.ReLU(),
                nn.Linear(d_ffn, d_model),
            )
        else:
            self.pos_embed = nn.Sequential(
                nn.Linear(num_poses * 3, d_ffn),
                nn.ReLU(),
                nn.Linear(d_ffn, d_model),
            )

    def forward(self, bev_feature, status_encoding, interpolated_traj) -> Dict[str, torch.Tensor]:
        # todo sinusoidal embedding
        # vocab: 4096, 40, 3
        # bev_feature: B, 32, C
        # embedded_vocab: B, 4096, C
        vocab = self.vocab.data
        L, HORIZON, _ = vocab.shape
        B = bev_feature.shape[0]
        if self.use_nerf:
            vocab = torch.cat(
                [
                    nerf_positional_encoding(vocab[..., :2]),
                    torch.cos(vocab[..., -1])[..., None],
                    torch.sin(vocab[..., -1])[..., None],
                ], dim=-1
            )

        if self.normalize_vocab_pos:
            embedded_vocab = self.pos_embed(vocab.view(L, -1))[None]
            embedded_vocab = self.encoder(embedded_vocab).repeat(B, 1, 1)
        else:
            embedded_vocab = self.pos_embed(vocab.view(L, -1))[None].repeat(B, 1, 1)
        tr_out = self.transformer(embedded_vocab, bev_feature)
        dist_status = tr_out + status_encoding.unsqueeze(1)
        result = {}
        # selected_indices: B,
        for k, head in self.heads.items():
            if k == 'imi':
                result[k] = head(dist_status).squeeze(-1)
            else:
                result[k] = head(dist_status).squeeze(-1).sigmoid()
        # how
        scores = (
            result['imi'].softmax(-1).log() + result['total'].log()
        )
        selected_indices = scores.argmax(1)
        result["trajectory"] = self.vocab.data[selected_indices]
        result["trajectory_vocab"] = self.vocab.data
        result["selected_indices"] = selected_indices
        return result