tb-like 0.2.0

A faster, columnar TensorBoard-style scalar viewer for many series and many runs.

tb_like

A faster, columnar TensorBoard-style scalar viewer — built for many series and many runs.

TensorBoard re-parses event files on demand and gets slow when a run has tens of thousands of scalar series across hundreds of experiments. tb_like instead converts TensorBoard event files into per-run Parquet once, then serves downsampled series on demand to a fast Plotly dashboard. New events are picked up incrementally in the background.

Why it's fast

• Convert once, read many. Each run's events.out.tfevents.* are parsed into a columnar Parquet file sorted by (tag, step), with row-group statistics so a query for a few tags only touches the matching row groups — even when a run has ~18k series.
• Incremental & idempotent. Ingestion tracks each event file's size and record count, so re-scans only parse new data. A background watcher keeps the cache in sync; parsing is parallelized across event files with joblib.
• Lazy, prioritized rendering. The dashboard renders charts only as they scroll into view, fetched through a priority queue (visible first, then neighbors, biased toward the scroll direction).
• LTTB downsampling keeps long curves cheap to draw without losing their shape.

Install

pip install tb-like
# or
uv tool install tb-like

Quick start

A "run" is a directory containing events.out.tfevents.* files. Point tb_like at a directory of runs and open the dashboard — that's it:

my_runs/
  run_a/  events.out.tfevents.*  config.yaml
  run_b/  events.out.tfevents.*
  ...

tblike my_runs --port 8000 --jobs 8
# open http://127.0.0.1:8000

The background watcher discovers runs under the folder, converts any that changed to Parquet (parsing event files across --jobs worker processes), keeps the cache in sync, and serves them — no separate build step. The cache lives in <runs_dir>/.tblike_cache by default (override with --cache-dir).

Dashboard features

• Hierarchical, searchable tag tree (regex filter) with smart grouping: path compression of a.b.c chains, numeric-enumeration collapsing (…expert_idx_∗), and layer indices kept as their own levels.
• Multi-run overlay, unified hover, EMA smoothing, log-y, step vs. relative-time x-axis, and outlier clipping by value percentiles.
• Collapsible per-group chart sections, resizable sidebar, and a one-click Refresh selected that re-ingests from disk and rebuilds the plots.
• Text diff panel (pinned at the bottom): compare logged text summaries — typically the run config — between any two runs/steps as a git-style diff.

Text summaries are ingested as event files are parsed, so new runs get them automatically. For runs converted by an older version (before text support), attach text without re-parsing scalars by running tblike backfill-text <runs_dir> — or just hit Refresh selected in the UI, which backfills any selected run that's missing text.

CLI

tblike <runs_dir> [--port P] [--host H] [--cache-dir D] [-j JOBS] [--no-watch]
                                            # the main command: serve + auto-ingest

Advanced / scriptable subcommands:

tblike convert RUN_DIR [RUN_ID] [-j JOBS]   # ingest one run into Parquet (one-off)
tblike scan                                 # one incremental ingest pass, no server
tblike backfill-text RUNS_DIR [-j JOBS]     # attach text (configs) to old caches

How it stores data

cache/<run_id>/
    data/seg-00000.parquet   # one immutable segment per ingest pass
    index.json               # tags, per-file ingest state, metadata
    meta.json                # tiny summary used for fast run listing

Reads union all segments and de-duplicate (tag, step) by latest wall_time.

License

MIT — see LICENSE.