训练的时候报错tools/train_amp.py FAILED

[yangaiping]

我按照作者您提供的训练执行命令·： export CUDA_VISIBLE_DEVICES=0 NGPUS=1 cfg_file=configs/bisenetv2_coco.py torchrun --nproc_per_node=$NGPUS tools/train_amp.py --config $cfg_file 遇到了如下报错问题：

[CoinCheung]

Hi,

Would you please show me the full error message?

[yangaiping]

好的，谢谢您，这是完整报错信息

[CoinCheung]

Have you specified dataset correctly?

[yangaiping]

您好，这是我的数据集文件

[CoinCheung]

Did you generate train.txt with the method in README.txt

[yangaiping]

我刚刚发现当我执行这个代码时会生成train.txt和val.txt，但是这两个文件是空的

[yangaiping]

[CoinCheung]

What is in the folder of images and labels?

[yangaiping]

[CoinCheung]

Why are these label files in format of txt?

Are you using coco-stuff dataset?

[yangaiping]

我用的是coco2017labels-segments.zip数据集

[yangaiping]

我可能已经发现了我的问题，我再试试coco-stuff dataset数据集

[yangaiping]

我已经使用了正确的数据集，并且成功划分数据集，但是训练的时候仍然报错，请问这是为什么呢

[CoinCheung]

Seems that you have hidden files in your image/train2017 folder, and likely in your train.txt file, would you have a check of this?

[yangaiping]

我把隐藏的文件.ipynb_checkpoints删掉了重新执行python tools/gen_dataset_annos.py --dataset coco，然后再执行下列命令，又出现了一个新的错误

[CoinCheung]

python tools/check_dataset_info.py --im_root datasets/coco --im_anns datasets/coco/train.txt

What is the output of this?

[yangaiping]

[CoinCheung]

Why does your coco-stuff has 201 categories? I used coco-stuff with only 171 classes.

You can change n_cat in the config file into 202, if you would like to use your dataset.

[yangaiping]

谢谢您，可能是我的数据集问题，我改成202后重新执行训练命令后好像成功运行了，但是我还想问怎么修改iter: 400/180000，感觉180000很大，怎么调小这个参数呢

[CoinCheung]

You should modify in the config file https://github.com/CoinCheung/BiSeNet/blob/f2b901599752ce50656d2e50908acecd06f7eb47/configs/bisenetv2_coco.py#L10

[yangaiping]

真的非常感谢您，感谢您耐心地解答我的问题，再次感谢！