请问在训练的时候，指标的数值一直很高，该如何解决呢？

[Yhc-777]

Val criterion: 218.1800 mse: 161874.7692 best: 224.6491, time: 168.9916

你好打扰啦，请问我在训练qnrf和sta、stb的时候，print出来的指标总是很高，尝试下调整了超参数但是还是变化不大，请问我应该如何解决呢？谢谢！

[Shimmer93]

可以发一下训练的script、config file和log吗？

[Yhc-777]

log.txt

script：python main.py --task train --config configs/sta_train.yml

config：

seed: 2023 version: sta device: 'cuda:0' log_para: 1000 patch_size: 10000 mode: 'final' num_epochs: 180

checkpoint: null

checkpoint: /mnt/2/hc/MPCount/pretrained_models/sta.pth model: name: 'final' params: pretrained: True mem_size: 1024 mem_dim: 256 cls_thrs: 0.5 err_thrs: 0.5 den_dropout: 0.5 cls_dropout: 0.3 has_err_loss: False train_dataset: &train_dataset_params name: 'den_cls' params: root: 'data/sta' crop_size: 320 downsample: 1 is_grey: False unit_size: 16 pre_resize: 1 val_dataset: train_dataset_params test_dataset: name: 'den_cls' params: root: '/mnt/2/hc/MPCount/data/ShanghaiTech/part_B/test_data/images' crop_size: 320 downsample: 1 is_grey: False unit_size: 16 pre_resize: 1 train_loader: batch_size: 8 num_workers: 8 shuffle: True pin_memory: True val_loader: &val_loader_params batch_size: 1 num_workers: 4 shuffle: False pin_memory: False test_loader: val_loader_params optimizer: name: 'adamw' params: lr: &lr 0.001 weight_decay: 0.0001 scheduler: name: 'onecycle' params: max_lr: *lr epochs: 300 steps_per_epoch: 15 final_div_factor: 1000

辛苦帮忙查看下！谢谢！！！

[Shimmer93]

看起来没有问题，可以检查下testing的script和config file

[Yhc-777]

testing的script: python main.py --task test --config configs/sta_test_stb.yml

config file:保持和repo的一致没有改动。

刚刚又测试了下，效果是：

(mpcount) u@u:/mnt/2/hc/MPCount$ python main.py --task test --config configs/sta_test_stb.yml Start testing at 2024-08-15 16:24:27 Loading checkpoint from logs/sta/best.pth Testing... ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:37 Testing results: mae: 13.0754 mse: 574.5333 Testing results saved to logs/sta End testing at 2024-08-15 16:25:05

这个的效果是属于较好的吗？麻烦你啦实在是！

[Shimmer93]

这个效果没问题

[Yhc-777]

明白，辛苦啦！谢谢！