Why is my best IoU only 66.4%? Is there something wrong with the hyperparameter settings in my config file?

[Misswcnm]

Why is my best IoU only 664? Is there something wrong with the hyperparameter settings in my config file? I’m training on 4 A30s in a cluster

DATA:
data_name: semantic_kitti
data_root: /share/home/u2415283021/SemanticKitti/dataset # 数据路径
label_mapping: util/semantic-kitti.yaml
classes: 19
fea_dim: 6
voxel_size: [0.05, 0.05, 0.05]
voxel_max: 120000

TRAIN:

architecture

arch: unet_spherical_transformer
input_c: 4
m: 32
block_reps: 2
block_residual: True
layers: [32, 64, 128, 256, 256]
quant_size_scale: 24
patch_size: 1
window_size: 6
use_xyz: True
sync_bn: True
rel_query: True
rel_key: True
rel_value: True
drop_path_rate: 0.3
max_batch_points: 1000000
class_weight: [3.1557, 8.7029, 7.8281, 6.1354, 6.3161, 7.9937, 8.9704, 10.1922, 1.6155, 4.2187, 1.9385, 5.5455, 2.0198, 2.6261, 1.3212, 5.1102, 2.5492, 5.8585, 7.3929]
xyz_norm: False
pc_range: [[-51.2, -51.2, -4], [51.2, 51.2, 2.4]]
window_size_sphere: [2, 2, 80]
window_size_scale: [2.0, 1.5]
sphere_layers: [1,2,3,4,5]
grad_checkpoint_layers: []
a: 0.0125
loss_name: ce_loss
use_tta: False
vote_num: 4

training setup

aug: True
transformer_lr_scale: 0.1
scheduler_update: step
scheduler: Poly
power: 0.9
use_amp: True
train_gpu: [0,1,2,3] # 使用 4 张 GPU
workers: 16 # 可适当加大以充分利用 CPU
batch_size: 8
batch_size_val: 8
base_lr: 0.006
epochs: 50
start_epoch: 0
momentum: 0.9
weight_decay: 0.02
drop_rate: 0.5

ignore_label: 255
manual_seed: 123
print_freq: 100
save_freq: 1
save_path: runs/semantic_kitti_unet32_spherical_transformer
weight:
resume:
evaluate: True
eval_freq: 10
val: False

Distributed:
dist_url: env://
dist_backend: nccl
multiprocessing_distributed: True
world_size: 1 # ✅ 单节点，必须是 1
rank: 0