TensorBoard

To launch TensorBoard in Google Colab, you must first load the TensorBoard notebook extension using the magic command `%load_ext tensorboard`. Then, you can start the dashboard by running `%tensorboard --logdir logs` in a new cell, pointing to your output directory. This embeds the interactive visualization interface directly within the Colab cell output.

Yes, TensorBoard is fully compatible with PyTorch via the `torch.utils.tensorboard.SummaryWriter` API. You can log scalar metrics like training loss, visualize model graphs, and inspect embeddings by writing data to a local event file. This integration works natively in PyTorch without needing to rely exclusively on TensorFlow or Keras.

To view your model architecture, you need to pass a dummy input tensor through your model and write it using `SummaryWriter.add_graph(model, input)`. Once TensorBoard loads the event files, you can navigate to the 'Graphs' tab on the top navigation bar to see a visual, node-based map of your layers. This is highly useful for debugging structural flaws in complex networks.

TensorBoard is an open-source, locally hosted tool built primarily for visualizing TensorFlow and PyTorch metrics during a single run or local session. Weights & Biases (W&B) is a commercial, cloud-based platform that focuses heavily on experiment tracking, team collaboration, model registry, and hyperparameter tuning across distributed environments. While TensorBoard is free and great for individuals, W&B excels in enterprise machine learning operations.