Run dashboard¶
The dashboard reads history.jsonl to compare active and completed runs on one
screen. Streamlit and Altair remain optional, so the core install does not get
heavier.
Install and run locally¶
The safe default is http://127.0.0.1:8501, refreshing from disk every two
seconds. Disable automatic refresh when needed:
The dashboard discovers history.jsonl in immediate children of the run root.
Because Trainer flushes every row, an active run is visible through its latest completed point. Only an incomplete final JSONL fragment is deferred until the next refresh. A malformed newline-terminated line is shown as an error for that run while healthy runs keep rendering.
Reading the screen¶
- Toggle each run independently in the sidebar.
Focus runchooses the source for the KPI, latest-row, andmeta.jsonpanels.- Metric toggles select which charts are visible.
train_lossandval_lossshare one loss chart. Color identifies the run; solid and dashed lines identify training and validation.- Arbitrary scalars such as IoU and AP50 keep their names and receive separate charts.
- Epoch and step runs render in separate chart groups with separate x axes.
The displayed timestamp is the last history-file modification time, not an inferred live/completed state. The dashboard is read-only and cannot control training.
Viewing training over SSH¶
SSH local forwarding is the safest default. On your local machine:
Then, on the remote host through that SSH session:
Open http://127.0.0.1:8501 locally. Streamlit stays bound to the remote
loopback interface and SSH provides the encrypted path. If local port 8501 is
already occupied, change both forwarded endpoints:
ssh -L 9000:127.0.0.1:9000 user@training-host
# remote
deeptool-dashboard runs/ --port 9000 --no-browser
Direct network exposure¶
No authentication or TLS
deeptool does not add authentication or TLS. Binding to 0.0.0.0 lets
anyone who can reach that port see run paths, metadata, and metrics. Use an
SSH tunnel unless you have configured your own firewall and trusted private
network.
Next¶
- Trainer — JSONL recording and step/epoch training
- Best weights & early stopping — monitors to compare with Focus