使用tensorboardX可视化Pytorch

可视化loss和acc

参考https://www.jianshu.com/p/46eb3004beca

1. 环境安装：

   conda activate xxx

   pip install tensorboardX

   pip install tensorflow

2. 代码：

   from tensorboardXimport SummaryWriter
   writer = SummaryWriter('runs/001')
       writer.add_scalar('Train/Loss', train_loss / batch_idx, epoch)
       writer.add_scalar('Train/Acc', 100.0 * correct / total, epoch)
   writer.close()

3. 服务器：

   conda activate xxx

   tensorboard --logdir=runs/001

4. 本地：

   终端上输入：ssh -p 22222 -L 6006:localhost:6006 yinwenbin@192.168.2.237

   浏览器上输入：localhost:6006

可视化模型

参考：https://blog.csdn.net/sunqiande88/article/details/80155925?utm_source=copy

import torchvision.models as models

from tensorboardX import SummaryWriter

import torch

model = models.resnet18()

dummy_input = torch.rand(13, 3, 224, 224)

with SummaryWriter(comment='resnet18') as w:

　　w.add_graph(model, (dummy_input, ))

conda activate xxx

tensorboard --logdir runs

浏览器上输入：localhost:6006

若提示错误：'torch._C.Value' object has no attribute 'debugName'

修改tensorboardX 1.9为tensorboardX 1.8

参考：https://blog.csdn.net/East_Plain/article/details/103073311