File size: 8,401 Bytes
da2e2ac |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 |
from typing import Dict
import numpy as np
import torch
import torch.nn as nn
from navsim.agents.dreamer.dreamer_network import DreamerNetwork
from navsim.agents.dreamer.hydra_dreamer_config import HydraDreamerConfig
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 HydraDreamerPlanningModel(nn.Module):
def __init__(self, config: HydraDreamerConfig):
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']
self.dreamer_network = DreamerNetwork(config)
img_num = 2 if config.use_back_view else 1
self._keyval_embedding = nn.Embedding(
config.img_vert_anchors * config.img_horz_anchors * img_num, config.tf_d_model
)
self._query_embedding = nn.Embedding(sum(self._query_splits), config.tf_d_model)
self.downscale_layer = nn.Conv2d(self.dreamer_network.fixed_vit.img_feat_c, config.tf_d_model, kernel_size=1)
self._status_encoding = nn.Linear((4 + 2 + 2) * config.num_ego_status, config.tf_d_model)
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 = HydraTrajDreamerHead(
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 img_feat_blc(self, camera_feature):
img_features = self.dreamer_network(camera_feature)['pred']
B, L, C = img_features.shape
img_features = img_features.view(B, self._config.img_vert_anchors, self._config.img_horz_anchors, C)
img_features = img_features.permute(0, 3, 1, 2)
img_features = self.downscale_layer(img_features).flatten(-2, -1)
img_features = img_features.permute(0, 2, 1)
return img_features
def forward(self, features: Dict[str, torch.Tensor],
interpolated_traj=None) -> Dict[str, torch.Tensor]:
status_feature: torch.Tensor = features["status_feature"]
batch_size = status_feature.shape[0]
img_features = self.img_feat_blc(features)
if self._config.use_back_view:
img_features_back = self.img_feat_blc(features["camera_feature_back"])
img_features = torch.cat([img_features, img_features_back], 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 = img_features
keyval += self._keyval_embedding.weight[None, ...]
query = self._query_embedding.weight[None, ...].repeat(batch_size, 1, 1)
agents_query = self._tf_decoder(query, keyval)
output: Dict[str, torch.Tensor] = {}
trajectory = self._trajectory_head(keyval, status_encoding, interpolated_traj)
output.update(trajectory)
agents = self._agent_head(agents_query)
output.update(agents)
return output
class HydraTrajDreamerHead(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({
'noc': nn.Sequential(
nn.Linear(d_model, d_ffn),
nn.ReLU(),
nn.Linear(d_ffn, 1),
),
'da':
nn.Sequential(
nn.Linear(d_model, d_ffn),
nn.ReLU(),
nn.Linear(d_ffn, 1),
),
'ttc': nn.Sequential(
nn.Linear(d_model, d_ffn),
nn.ReLU(),
nn.Linear(d_ffn, 1),
),
'comfort': nn.Sequential(
nn.Linear(d_model, d_ffn),
nn.ReLU(),
nn.Linear(d_ffn, 1),
),
'progress': 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=None) -> 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()
# imi_weight = 0.01
# noc_weight = 0.1
# da_weight = 0.5
# tpc_weight = 3.0
scores = (
0.01 * result['imi'].softmax(-1).log() +
0.1 * result['noc'].log() +
0.5 * result['da'].log() +
3.0 * (5 * result['ttc'] + 2 * result['comfort'] + 5 * result['progress']).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
|