Logging Metrics & Data

Creating a Tracker

from trailer import Tracker

t = Tracker(project="demo")
t = Tracker(project="demo", resume_from="run_a1b2c3")     # resume a previous run
t = Tracker(project="demo", config={"lr": 0.01})          # log hyperparameters
t = Tracker(project="demo", metric_directions={"accuracy": "max"})

Logging scalar metrics

t.log({"loss": 0.5, "accuracy": 0.91}, step=10)

• Log multiple metrics per call; dict keys are metric names
• step is optional; auto-increments when omitted
• TensorBoard-style grouping works: train/loss, val/accuracy — grouped automatically in the UI

Step semantics

• Explicit step: precise control of the x-axis
• Auto step: consecutive log() calls increment automatically
• Hardware-monitoring samples align with experiment steps; toggle Step / Wall Time views

Resuming a run

t = Tracker(project="demo", resume_from="run_a1b2c3")   # reuse the run_id, continue from last_step

• Pass the same run_id via resume_from to append data to a previous run instead of creating a new one. On first run, persist t.run_id (e.g. into a checkpoint file) so it can be restored later
• On resume the SDK asks the server for the run’s last step and auto-increments from last_step + 1; with explicit step= you control the x-axis yourself
• Resume negotiation always goes over HTTP (POST /api/v1/runs/{id}/resume + GET /api/v1/runs/{id}/last_step, default 127.0.0.1:5120) — even in local mode. Run trailer up so step continuation works; otherwise the negotiation fails silently and auto-steps restart from 0
• All three local storage backends support resuming: SQLite (storage="sqlite", default), file mode (storage="file", data_dir="data"), PostgreSQL (storage="postgres")

Lifecycle

t.finish()   # mark the run finished (state → finished)

• The SDK heartbeats every 30s in the background; after 60s without a heartbeat the server marks a run crashed
• finish() flushes the remaining buffer and joins the sampler thread, so the last step is never lost

Performance

• Writes go through a thread-safe RingBuffer (100k entries, P99 < 100µs) — log() never blocks training
• A background FlushThread batches (500 items / 1s) and writes in bulk; local mode goes through PyO3 with zero network