Skip to content

Concepts

Matyan uses the same core concepts: runs, repo (backend), params, sequences, and context. Data lives server-side in FDB and S3, not in a local repo.

Run

A Run represents one experiment/training run. In code it’s matyan_client.Run. You use it to:

  • Set hyperparameters and other params (run["hparams"] = {...}).
  • Track metrics and objects (run.track(...)).
  • Add tags, set name/description, log artifacts and terminal output.

Runs are queryable and visible in the UI; each has a unique hash.

Repo (backend)

Repo is the set of all runs (and related data) served by one backend. The client’s Repo object is a handle to that backend:

from matyan_client import Repo

repo = Repo("http://localhost:53800")  # backend URL
for run in repo.iter_runs():
    ...

So “repo” in the docs and UI means “the set of runs (and related data) served by this backend.”

Run params

Each run has parameters: hyperparameters, config, custom keys. You set them with a dict-like interface:

run["hparams"] = {"lr": 0.001, "batch_size": 32}
run["custom_key"] = "value"

Params are stored in FDB and can be used in MatyanQL and for grouping in the UI.

Run sequence

A sequence is an ordered list of values of the same kind (e.g. a metric series, or a sequence of images). It’s bound to a run and identified by name and context. When you call run.track(value, name="loss", context={"subset": "train"}), you’re appending to (or creating) the sequence "loss" with context {"subset": "train"}.

Sequences are typed (metric, image, audio, etc.). The UI and API represent them accordingly (charts, galleries, etc.).

Sequence context

Context lets you have multiple sequences with the same name in one run. For example:

  • name="loss", context={"subset": "train"}
  • name="loss", context={"subset": "val"}

So the sequence is defined by (run, name, context). Context is a dict and is used in MatyanQL and the UI for filtering and grouping.