multicam_flosp_crp_stereodepth_cascadecls_2080ti.yaml

### dataset
#dataset: "NYU" # "kitti", "tartanair"
dataset: "NYU"

n_relations: 4

enable_log: true
data_root: '/data/dataset/NYU_dataset/depthbin/'
data_preprocess_root: '/data/dataset/NYU_dataset_preprocess'
data_stereo_depth_root: ''
data_lidar_depth_root: ''

logdir: 'logdir'

### training cfg
max_epochs: 30 # 训练 epoch 数量
gradient_clip_val: 35 # gradient clip val (0 FOR w/o gradient clip)
use_stereo_depth_gt: false # stereo depth 监督，只有 kitti 支持
use_lidar_depth_gt: false # lidar depth 监督，只有 kitti 支持
use_depth_gt: true # depth gt 监督，只有 nyu 支持
depth_loss_weight: 1.0 # depth loss 权重
deterministic: false # will increase gpu memory usage
use_strong_img_aug: false # 图像域增广
sem_step_decay_loss: false # sem loss step decay

share_2d_backbone_gradient: true #true: save gpu memory, false: higher scores

fp_loss: true
frustum_size: 8
batch_size_per_gpu: 1
n_gpus: 1
num_workers_per_gpu: 0
exp_prefix: "nyu_exp"
run: 1
lr: 2e-4
weight_decay: 1e-4

context_prior: false # crp loss

relation_loss: true
CE_ssc_loss: true
sem_scal_loss: true
geo_scal_loss: true

project_1_2: true
project_1_4: true
project_1_8: true

ckpt: ''

### MultiView
multi_view_mode: false # 多摄模式，只有 kitti 支持多摄
pattern_id: 0 # flosp 取点逻辑

### network config
full_scene_size: # 3d scene 尺寸定义
  - 60
  - 36
  - 60
project_scale: 1 # 投影到 3d voxel 的下采样倍数
feature: 100 # 3d 网络 channel 维度
feature_2d_oc: 100 # # 2d 网络输出 channel 维度
n_classes: 12 # 分类数目
backbone_2d_name: 'tf_efficientnet_b4_ns' # 2d backbone 定义
return_up_feats: 1 # choices:1, 2, 4, 8; help: 2d backbone 返回的最大特征尺寸设定
cascade_cls: false # 级联分类
occluded_cls: false # occluded 分类，只有 kitti 支持

### 2d->3d transformation
trans_2d_to_3d_debug: false
trans_2d_to_3d: "flosp" # "flosp",  "flosp_depth"