TensorBoard

TensorBoard is a way to visualize your model and various statistics during or after training.

CLI Usage

CLI

tensorboard --logdir [logs]

Flags

• --samples_per_plugin indices the number of samples to show for each tab. Non-scalar objects are sampled using reservoir sampling.
  • --samples_per_plugin images=10000 samples approximately 10000 samples.

Training Usage

If you're using a custom training loop (i.e. gradient tape), then you'll need to set everything up manually.

First create a SummaryWriter

train_log_dir = os.path.join(args.checkpoint_dir, "logs", "train")
train_summary_writer = tf.summary.create_file_writer(train_log_dir)

Scalars

Add scalars using tf.summary.scalar:

with train_summary_writer.as_default():
  tf.summary.scalar("training_loss", m_loss.numpy(), step=int(ckpt.step))

PyTorch

PyTorch also supports output tensorboard logs.
See https://pytorch.org/docs/stable/tensorboard.html.
There is also lanpa/tensorboardX but I haven't tried it.

from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter(log_dir="./runs")

writer.add_scalar("train_loss", loss_np, step)

# Optionally flush e.g. at checkpoints
writer.flush()

# Close the writer (will flush)
writer.close()

Resources

• Getting started with TensorBoard
• Hands-on TensorBoard (TensorFlow Dev Summit 2017)