Skip to content

Logging Compatibility

Runicorn supports both its own logging APIs and several common ML logging patterns that people already use in training scripts.

Core options

Console capture

run = rn.init(path="demo", capture_console=True, tqdm_mode="smart")

This captures print() output and writes it to logs.txt while still showing it in the terminal.

That is different from run.log_text(...), which writes directly to logs.txt but does not print to the terminal by itself.

Python logging integration

import logging
import runicorn as rn

run = rn.init(path="demo")
logger = logging.getLogger(__name__)
logger.addHandler(run.get_logging_handler())
logger.info("training started")

Compatibility matrix

Pattern Import to use
torchvision MetricLogger from runicorn.log_compat.torchvision import MetricLogger
ImageNet AverageMeter from runicorn.log_compat.imagenet import AverageMeter, ProgressMeter
TensorBoard SummaryWriter from runicorn.log_compat.tensorboard import SummaryWriter
tensorboardX SummaryWriter from runicorn.log_compat.tensorboardX import SummaryWriter

torchvision MetricLogger

from runicorn.log_compat.torchvision import MetricLogger

metric_logger = MetricLogger(delimiter="  ")
for batch in metric_logger.log_every(dataloader, 10, header="Train"):
    loss = train_step(batch)
    metric_logger.update(loss=loss)

When an active Runicorn run exists, updates are routed into Runicorn automatically.

ImageNet-style meters

from runicorn.log_compat.imagenet import AverageMeter, ProgressMeter

losses = AverageMeter("Loss", ":.4f")
top1 = AverageMeter("Acc@1", ":.2f")
progress = ProgressMeter(len(loader), [losses, top1], prefix="Train: ")

for i, batch in enumerate(loader):
    loss, acc = train_step(batch)
    losses.update(loss)
    top1.update(acc)
    if i % 10 == 0:
        progress.display(i)

TensorBoard and tensorboardX

from runicorn.log_compat.tensorboard import SummaryWriter

writer = SummaryWriter()
writer.add_scalar("train/loss", 0.42, 1)
writer.add_text("notes", "warmup finished", 1)
writer.add_hparams({"lr": 1e-3}, {"val_acc": 0.88})

The tensorboardX-compatible module supports the same migration pattern:

from runicorn.log_compat.tensorboardX import SummaryWriter

When automatic routing works

The compatibility helpers write through to Runicorn when there is an active run:

import runicorn as rn

run = rn.init(path="demo")

If no active run exists, the compatibility layer cannot attach your metrics to a Runicorn run.

For most users:

  1. start a Runicorn run normally
  2. enable capture_console=True if you want console logs
  3. use the compat import that matches your existing training loop

Next steps