File size: 3,657 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 |
from dataclasses import dataclass
from typing import Any, List, Tuple, Dict
from nuplan.common.maps.abstract_map import SemanticMapLayer
from nuplan.common.actor_state.tracked_objects_types import TrackedObjectType
from nuplan.planning.simulation.trajectory.trajectory_sampling import TrajectorySampling
@dataclass
class TransfuserConfig:
trajectory_sampling: TrajectorySampling = TrajectorySampling(
time_horizon=4, interval_length=0.5
)
image_architecture: str = "resnet34"
lidar_architecture: str = "resnet34"
max_height_lidar: float = 100.0
pixels_per_meter: float = 4.0
hist_max_per_pixel: int = 5
lidar_min_x: float = -32
lidar_max_x: float = 32
lidar_min_y: float = -32
lidar_max_y: float = 32
lidar_split_height: float = 0.2
use_ground_plane: bool = False
# new
lidar_seq_len: int = 1
camera_width: int = 1024
camera_height: int = 256
lidar_resolution_width = 256
lidar_resolution_height = 256
img_vert_anchors: int = 256 // 32
img_horz_anchors: int = 1024 // 32
lidar_vert_anchors: int = 256 // 32
lidar_horz_anchors: int = 256 // 32
block_exp = 4
n_layer = 2 # Number of transformer layers used in the vision backbone
n_head = 4
n_scale = 4
embd_pdrop = 0.1
resid_pdrop = 0.1
attn_pdrop = 0.1
# Mean of the normal distribution initialization for linear layers in the GPT
gpt_linear_layer_init_mean = 0.0
# Std of the normal distribution initialization for linear layers in the GPT
gpt_linear_layer_init_std = 0.02
# Initial weight of the layer norms in the gpt.
gpt_layer_norm_init_weight = 1.0
perspective_downsample_factor = 1
transformer_decoder_join = True
detect_boxes = True
use_bev_semantic = True
use_semantic = False
use_depth = False
add_features = True
# Transformer
tf_d_model: int = 256
tf_d_ffn: int = 1024
tf_num_layers: int = 3
tf_num_head: int = 8
tf_dropout: float = 0.0
# detection
num_bounding_boxes: int = 30
# loss weights
trajectory_imi_weight: float = 10.0
agent_class_weight: float = 10.0
agent_box_weight: float = 1.0
bev_semantic_weight: float = 10.0
# BEV mapping
bev_semantic_classes = {
1: ("polygon", [SemanticMapLayer.LANE, SemanticMapLayer.INTERSECTION]), # road
2: ("polygon", [SemanticMapLayer.WALKWAYS]), # walkways
3: ("linestring", [SemanticMapLayer.LANE, SemanticMapLayer.LANE_CONNECTOR]), # centerline
4: (
"box",
[
TrackedObjectType.CZONE_SIGN,
TrackedObjectType.BARRIER,
TrackedObjectType.TRAFFIC_CONE,
TrackedObjectType.GENERIC_OBJECT,
],
), # static_objects
5: ("box", [TrackedObjectType.VEHICLE]), # vehicles
6: ("box", [TrackedObjectType.PEDESTRIAN]), # pedestrians
}
bev_pixel_width: int = lidar_resolution_width
bev_pixel_height: int = lidar_resolution_height // 2
bev_pixel_size: float = 0.25
num_bev_classes = 7
bev_features_channels: int = 64
bev_down_sample_factor: int = 4
bev_upsample_factor: int = 2
@property
def bev_semantic_frame(self) -> Tuple[int, int]:
return (self.bev_pixel_height, self.bev_pixel_width)
@property
def bev_radius(self) -> float:
values = [self.lidar_min_x, self.lidar_max_x, self.lidar_min_y, self.lidar_max_y]
return max([abs(value) for value in values])
|