为什么 TensorBoard 摘要没有更新？

Question

I use tensorboard with pytorch1.1 to log loss values.

I use writer.add_scalar("loss", loss.item(), global_step) in every for- loop body.

However, the plotting graph does not update while the training is processing.

Every time I want to see the latest loss, I have to restart the tensorboard server.

The code is here

import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
from torch.utils.tensorboard import SummaryWriter
from torchvision import datasets, transforms

# Writer will output to ./runs/ directory by default
writer = SummaryWriter()

transform = transforms.Compose(
    [transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))]
)
trainset = datasets.MNIST("mnist_train", train=True, download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=64, shuffle=True)
model = torchvision.models.resnet50(False)
# Have ResNet model take in grayscale rather than RGB
model.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)
model.fc = nn.Linear(2048, 10, True)

criterion = nn.CrossEntropyLoss()

epochs = 100

opt = torch.optim.Adam(model.parameters())

niter = 0

for epoch in range(epochs):
    for step, (x, y) in enumerate(trainloader):
        yp = model(x)
        loss = criterion(yp, y)
        opt.zero_grad()
        loss.backward()
        opt.step()
        writer.add_scalar("loss", loss.item(), niter)
        niter += 1
        print(loss.item())

grid = torchvision.utils.make_grid(images)
writer.add_image("images", grid, 0)
writer.add_graph(model, images)
writer.close()

The training is still going on, and the global steps has already been 3594. However, the tensorboard still shows around 1900.

Answer 1

同样对于单次运行有多个事件日志文件的人，您需要使用--reload_multifile True启动张量--reload_multifile True

Answer 2

There is caching done internally on the logging side. To see if that is the issue, create your SummaryWriter with

writer = SummaryWriter(flush_secs=1)

and see if things update right away. If so, feel free to tune flush_secs (defaults to 120) for your case. From your description, though, this might be from the TensorBoard visualization side. If so, it must have something to do with the polling interval.

Does installing TensorFlow (which forces TensorBoard to use a different filesystem backend) change this behavior for you?