Weights & Biases
Migrate from Weights & Biases
We support migrating from existing experiment tracking tools, including Weights & Biases.
API Compatibility
The majority of APIs you use should be directly compatible with wandb's Python client at PyPI with the exception of a few that include performance improvements from mlop.
If you encounter any issues, please let us know via founders@mlop.ai or Discord.
Migrating existing code should be as simple as changing the import statements:
| Type | New | Deprecated |
|---|---|---|
| General | ||
| PyTorch Lightning | ||
| Hugging Face Transformers |
Migrating Existing Runs
We fully support migrating runs hosted on Weights & Biases to our platform. Everything you can (and cannot) see that is recorded by wandb should be viewable at mlop too.
We invest a lot of development time in transparency in data collection and ensures you are fully aware what data is collected and stored, despite potentially different practices from platforms you may be migrating from.
Overcoming Limitations
Our custom in-house client can be used for this process - this means the migration will not be subjected to the same content and rate limits as the public wandb client or wandb.Api().
At the moment, wandb's official client restrict downloading runs with a large number of files (images/vidoes) due to their API endpoint taking too long to respond before dropping the connection. Their support has acknowledged this issue but there is no timeline for a fix for now. There is also data that is recorded by wandb but simply not made available using the public wandb.Api().
Our own implementation supports all data types recorded by wandb and uses a different protected endpoint. As such, it is immune to any artificial limitations.
For enterprise customers, the migration can be self-hosted and is fully on-premise under your control over all code and data. For individuals, we recommend contacting us for support before proceeding.