Already using W&B, TensorBoard, or plain logs?¶
Epochix isn't a replacement for your experiment tracker. It answers a different question.
A tracker records what happened across many runs — every metric, every config, every sweep, kept so you can compare them later. Epochix reads one run and tells you, in words, what it means: where the model peaked, whether it is overfitting, what the best epoch actually was, and a letter grade with its reasoning attached.
You can keep your tracker and point Epochix at runs you already have.
From a plain log file — nothing installed, nothing changed¶
This is the main path and needs no integration at all. If your training script prints anything resembling metrics, Epochix reads it:
Already finished? Point it at the file:
This works on logs you did not write — a colleague's, a cluster job's, one from six months ago. There is no SDK call to add and nothing to instrument, which is the part a tracker cannot do retroactively.
Not sure whether your log is readable? Ask before committing to anything:
It reports what it can and cannot extract, and what to add to your prints if something is missing.
From TensorBoard¶
Reads the event files straight off disk. No account, no network, no login:
Point it at the directory containing events.out.tfevents.*. Every
sub-directory becomes its own run, so a logdir holding several experiments
imports all of them.
From Weights & Biases¶
From a run already on your disk — no account needed¶
Every W&B run writes a directory. Point Epochix at it:
That reads the run history straight off local disk — no login, no API key, no
network. It accepts your wandb/ directory, a single run directory inside it,
or the run-*.wandb file itself, and imports every run it finds.
This works for runs logged with WANDB_MODE=offline as well as ordinary ones,
because the local directory is written either way.
Requires pip install wandb — the file is a binary record log, and Epochix
uses W&B's own reader for it rather than guessing at the format.
From the W&B servers¶
If the run is not on this machine, fetch it by reference:
This form needs credentials
Anything that is not an existing path is treated as
entity/project/run_id and fetched through the W&B API, which needs an
API key — set WANDB_API_KEY or pass --api-key.
What you get that a tracker doesn't¶
A tracker draws the curve. Epochix reads it:
lower-lr finished ahead of baseline: 0.8970 against 0.8460 (val_accuracy). baseline peaked at 0.8650 on epoch 7 and ended worse, at 0.8460. Had it stopped at its best the gap would have been 0.0320 rather than 0.0510. lower-lr was still improving when it stopped, so its result is probably not its ceiling.
Plus a letter grade, an animated GIF you can drop into a slide or a PR, and a single-file HTML report that opens offline.
What it deliberately will not do¶
Epochix does not manage sweeps, host a model registry, track artifact lineage, or coordinate a team. Those are what a tracker is for, and W&B, MLflow and Neptune do them well.
It also refuses to state things it cannot support. A grade is compared against absolute per-task thresholds, so it cannot know that 85% on your dataset is excellent where 85% on MNIST is poor — and the dashboard says so on the card rather than leaving you to assume otherwise. Metric values that are impossible for the quantity in question are reported as impossible instead of narrated as fact.