Foundations — TensorBoard - Weights & Biases logging

Before we can read a single line of the parent note, we need to make sure no word on this page is a mystery. Below, every symbol and term the parent uses is unpacked from zero: plain meaning → the picture it draws → why the topic needs it.

Start with the most basic object: a training loop

Everything logged comes out of a loop that repeats. If you have never seen one, here is the whole idea in a picture.

Look at the figure. The circle is one turn of the loop. Every turn spits out at least one number — the loss. That stream of numbers, turn after turn, is exactly what we want to record. Without the loop there is nothing to log; the loop is the source of all data.

The x-axis of everything: step and epoch

Every logged number needs an answer to "when did this happen?". Two words answer that.

In the figure, the long horizontal line is step — it never resets. Epochs are the coloured brackets grouping equal chunks of steps. The parent note computes:

global_step = epoch * len(train_loader) + batch_idx

Let's read that in plain words, because it is the single most important arithmetic on the parent page.

WHY multiply? Because $b$ alone restarts at $0$ every epoch. If we plotted against $b$, epoch 2's points would land on top of epoch 1's points — the graph would be a scribble. Multiplying $e\times B$ pushes each epoch's block of steps to a fresh, higher region of the x-axis so nothing overlaps. WHAT it looks like: a straight, always-increasing timeline (the flat line in the figure), never a loop-back.

The y-axis values: the four things we plot

The parent note logs four kinds of thing. Each is a different shape of picture. Here is what each word means before you ever meet its code.

1. Scalar — a single number

That is the plot you already imagined for loss. add_scalar(...) / wandb.log({...}) write these. The picture: a line chart, x = step, y = the number.

2. Histogram — the shape of a bag of numbers

A model does not hold one weight; a single layer holds thousands. To watch them all at once we need a picture of a whole collection of numbers.

Look at the figure: the same 5,000 weight values, drawn as a histogram. WHY the topic needs this: a single scalar can't tell you if a layer's weights are all collapsing to zero (dead neurons) or blowing up. The shape of the bag tells you. This is the direct link to vanishing / exploding gradients — a histogram of gradients that shrinks toward a spike at zero is what "vanishing" looks like.

3. Image — a grid of pixels

4. Graph — the model's wiring diagram

Where the numbers get written: the writer and the file

We have numbers and we have timestamps. Now: how do they leave the running program?

The figure shows the flow: loop → writer → event file → (later) the viewer draws the plot. Two crucial properties, both visible as one-way arrows:

• Append-only: new records are only ever added to the end, never rewritten. WHY? If training crashes at step 5000, everything up to 5000 is already safely on disk — nothing is lost. That is the "crash-safe" claim in the parent note.
• Decoupled: the viewer reads the file independently of the training program. WHY? You can open the plots while training still runs, or days after it finished, from a different terminal.

The vocabulary of "what we're tuning": hyperparameters, metrics, artifacts

The parent's definition of experiment tracking uses three category words. Pin them down.

Two words the parent leans on: reproducibility and provenance

This is the bridge to model versioning: an artifact (a checkpoint) plus its provenance (the exact run that made it) is a versioned model you can trust and retrieve.

How the pieces feed the topic

Read it top to bottom: the loop makes numbers, we tag them with when and what, a writer files them, and a viewer turns the file back into pictures. Add the categories and provenance, and you have full experiment tracking — the parent topic.

Equipment checklist

Cover the right side and answer. If any answer surprises you, re-read that section before the parent note.