guide
From W&B or Neptune to GPU Zero
Coming from Weights & Biases, or from Neptune before its hosted service closed in March 2026? Bring your old runs over, change a few lines of logging, and gain checkpoints you can resume and fork.
1. Bring your old runs over
ravex import, in Ravex 0.11.0 and later, writes a run from elsewhere as a Ravex store of its own, which ravex ship then sends to the console. Only the scalars come across, a number per metric per step: images, tables and another tool's histograms do not, and an imported run can be compared and charted, not resumed, since its checkpoints were never Ravex's.
From W&B
With the run's path as W&B shows it, entity, project and run id. Ravex reads its full history through your own W&B login, and keeps the run's name, configuration and start time:
pip install wandb # the W&B client, to read your runs
wandb login
ravex import wandb my-team/my-project/1a2b3c4d --storage ./imported/baselineFrom Neptune
From Ravex 0.11.1. Export your project with Neptune's own neptune-exporter, then point Ravex at the Parquet files it writes. Ravex keeps every float series of the run, with its steps and times, takes the run's name and start from sys/, and its other single values as the configuration. Where the export holds several runs, --run picks one. A run logged at fractional steps, such as 0.5, has its steps numbered 0, 1, 2 in order, and the command says so:
pip install pyarrow # to read the exporter's Parquet files
ravex import neptune ./exports/data --run SAN-42 --storage ./imported/san-42From MLflow or anything else
From a CSV or JSONL file of metrics, such as one written from MLflow's metric history, in either of two shapes. Wide: a row per step, a step column and a column per metric. Long: a row per value, with the metric's name in metric or key, the number in value and its step. A time or timestamp column, in seconds, milliseconds or ISO 8601, is kept.
ravex import table metrics.csv --storage ./imported/run-42 --name run-42Then send it
With a machine token from the console, under Settings → Machine tokens:
export RAVEX_METRICS_TOKEN=gzat_...
ravex ship --storage ./imported/baseline --endpoint https://api.gpuzero.devThe run appears on the Metrics page, finished, with the name it had.
2. Change the logging
The loop goes inside one decorated function, and wandb.log becomes ravex.log_metrics. The step is Ravex's own count of optimizer steps, so it is not passed; the configuration is what the script was started with.
Before:
import wandb
wandb.init(project="shakespeare", config={"lr": 3e-4})
for step, batch in enumerate(loader):
loss = train_step(batch)
wandb.log({"train/loss": loss.item()}, step=step)After:
import ravex
@ravex.train_loop(max_steps=5000)
def train():
for batch in loader:
loss = train_step(batch)
ravex.log_metrics({"train/loss": loss})
train()A tensor with more than one element is kept as a histogram. GPU and CPU use are sampled on their own. Set RAVEX_METRICS_ENDPOINT and RAVEX_METRICS_TOKEN as in the docs and the run shows up while it trains. W&B can keep running beside it while you compare: Ravex does not touch its calls.
3. What you get that you did not have
- Resume and fork from any checkpoint, with the optimizer, the scheduler, the random generators and the position in the data, so the child continues the parent's curve.
- Preemption that costs a step: on SIGTERM the run saves and exits, and the same command carries on.
- GPUs in your own cloud account, started for a job and given back after it, under a budget.
- Your data stays yours: checkpoints in your own bucket once you connect one, and every run downloadable as one file. The libraries are open source, at github.com/gpuzero-dev.
Weights & Biases, W&B, Neptune and MLflow are trademarks of their respective owners, named here only to say which tool a guide is for. GPU Zero is not affiliated with them.