tensorboard

[kkkkkkkb]

请问如何显示tensorboard，train里面的代码没用相关函数～ 我按照vmunet里train.py修改了，但是出现的结果不太对。 我是这么修改的：

1. train.py：

   • 添加了： global writer writer = SummaryWriter(config.work_dir + 'summary')
   • 在train_one_epoch添加writer参数： train_one_epoch( train_loader, model, criterion, optimizer, scheduler, epoch, logger, config, writer, scaler=scaler )

2. engine.py：def train_one_epoch 函数添加writer参数，并在函数内加了句writer.add_scalar('loss', loss, epoch)

最后的结果如下图

[wurenkai]

Hi, according to the graph you have provided, Loss is decreasing at 0-50 epochs and is increasing at 50-100 epochs. 100 epochs is one cycle. According to the graph you have provided, Loss will further decrease at 100-150 epoch, such a trend is normal. This is because the scheduler with 'CosineAnnealingLR' is set by default.

What are the results you are referring to?

[kkkkkkkb]

Hi, according to the graph you have provided, Loss is decreasing at 0-50 epochs and is increasing at 50-100 epochs. 100 epochs is one cycle. According to the graph you have provided, Loss will further decrease at 100-150 epoch, such a trend is normal. This is because the scheduler with 'CosineAnnealingLR' is set by default.

What are the results you are referring to?

谢谢，原来这是正常的结果，一开始loss抖动太大所以我没看明白。

[kkkkkkkb]

Hi, according to the graph you have provided, Loss is decreasing at 0-50 epochs and is increasing at 50-100 epochs. 100 epochs is one cycle. According to the graph you have provided, Loss will further decrease at 100-150 epoch, such a trend is normal. This is because the scheduler with 'CosineAnnealingLR' is set by default.

What are the results you are referring to?

给作者一个反馈，我在ISIC2017上跑出了跟您论文上一样的结果，太优秀了！但是在自己数据集上跑完300个epoch，dice竟然0.02，我是一个多分类的任务，可能我的代码还有些问题，包括数据集的处理。