TensorBoardX

Warnings

• breaking The `protobuf` dependency has undergone several version changes. Version 2.6.5 requires `protobuf>=5.29.6`. Earlier versions required `>=3.20`. Ensure your `protobuf` version is compatible with your `tensorboardX` installation to avoid `TypeError` or `ImportError` issues.
  Fix: Check the `tensorboardX` PyPI page or `pyproject.toml` for the exact `protobuf` requirement for your installed version and upgrade/downgrade `protobuf` accordingly (e.g., `pip install protobuf==X.Y.Z`).
• gotcha TensorBoardX only generates the log files; the `tensorboard` package itself must be installed separately to run the visualization server. Without it, you cannot view the generated logs.
  Fix: Install `tensorboard` alongside `tensorboardx`: `pip install tensorboardx tensorboard`. Then run the server with `tensorboard --logdir=<your_log_dir>`.
• gotcha Logging performance can be an issue, especially when displaying many experiments (3+) or logging high-resolution data (images, large histograms) with many points. While logging is cheap, displaying can be expensive and slow down TensorBoard.
  Fix: Consider reducing the frequency of logging expensive data types (images, histograms), grouping related scalars using `writer.add_scalars`, or using fewer experiments in the TensorBoard view.
• gotcha When logging PyTorch scalar tensors, you must extract the Python scalar value using `.item()` before passing it to `add_scalar()`, otherwise `tensorboardX` may complain.
  Fix: Instead of `writer.add_scalar('tag', torch_tensor, iteration)`, use `writer.add_scalar('tag', torch_tensor.item(), iteration)`.
• deprecated The `add_audio()` function was significantly optimized (200x speedup) starting from version 2.1, but this requires the `soundfile` package to be installed.
  Fix: For optimal performance with `add_audio()`, ensure `tensorboardX` is v2.1 or newer and install `soundfile`: `pip install soundfile`.
• breaking From `tensorboardX` v2.1, the `add_graph` function's underlying implementation was delegated to `torch.utils.tensorboard`. While the API might remain similar, internal behavior and potential compatibility nuances with specific PyTorch versions could change.
  Fix: If encountering issues with `add_graph` on older `tensorboardX` versions, consider upgrading to v2.1 or newer and ensuring PyTorch compatibility. Consult `torch.utils.tensorboard` documentation for specific graph tracing requirements.
• gotcha Importing `tensorboardX` after `tensorboard` (specifically `from tensorboard import main as tb`) in the same Python session can lead to `TypeError: Couldn't build proto file into descriptor pool!` due to conflicts in protobuf descriptor registration.
  Fix: Avoid importing `tensorboardX` after `tensorboard` in the same session. If both are needed, ensure `tensorboardX` is imported first, or consider running them in separate processes if strict isolation is required.

Install

• pip install tensorboardx Basic installation
• pip install tensorboardx crc32c soundfile tensorboard With optional speedups and TensorBoard for viewing

Imports

• SummaryWriter

  from tensorboardX import SummaryWriter

  This is the primary class for writing events to TensorBoard logs.

Quickstart

This quickstart demonstrates how to initialize a `SummaryWriter`, log scalar values over iterations, and log a PyTorch model graph. After running the script, a 'runs' directory will be created containing the event files. You then use the `tensorboard` command-line tool to launch the visualization server.

import torch
import numpy as np
from tensorboardX import SummaryWriter

# Create a writer instance, logs will be saved in 'runs/my_experiment'
writer = SummaryWriter('runs/my_experiment')

# Log scalar values
for i in range(100):
    writer.add_scalar('training/loss', 100 - i * 0.5 + np.random.rand(), i)
    writer.add_scalar('training/accuracy', 0.5 + i * 0.005 + np.random.rand() * 0.01, i)

# Log a Pytorch model graph (dummy model and input)
try:
    import torch.nn as nn
    class SimpleModel(nn.Module):
        def __init__(self):
            super().__init__()
            self.linear = nn.Linear(10, 2)
        def forward(self, x):
            return self.linear(x)
    dummy_input = torch.randn(1, 10)
    model = SimpleModel()
    writer.add_graph(model, dummy_input)
except ImportError:
    print("PyTorch not installed, skipping add_graph example.")

# Close the writer to flush all pending events to disk
writer.close()

print("TensorBoardX logs written to 'runs/my_experiment'.")
print("To view these logs, run the following command in your terminal:")
print("  tensorboard --logdir=runs")