Skip to main content
An ONNX model is uploaded once, described, and downloaded wherever it is needed. BlobHub stores it, describes it and hands it back; running it is up to you. Each model is stored as a versioned blob of type blobhub.model.onnx, and each revision of it holds one model. How BlobHub works covers blobs, revisions and the type behind them.

​
Uploading a model

A model travels as an archive, model.tar.gz, holding a single file named model.onnx. An archive of up to 4 MiB goes in the body of one upload request. A larger one goes in parts: the platform hands back a presigned URL for each 10 MiB part, you send the parts straight to storage, and one more call completes the upload. Either way, the call returns an operation id instead of waiting. The platform unpacks the archive in the background, and you poll the operation until it is completed, or failed when the archive does not hold a model that parses as ONNX. Uploads go to a draft revision, and committing the draft freezes its model. A new revision starts with a copy of the stored model, so a retrained model is uploaded to the new draft while the committed revision keeps the old one.

​
Topology

Once the archive is in, the platform extracts the model file and parses its graph into a JSON topology: the nodes, inputs, outputs and initializers ONNX defines, without the coefficients. It is the model’s shape without its weights, and it is how you look at a large model without fetching it. A revision therefore holds the model in three forms, each fetched on its own: the archive as uploaded (model.tar.gz), the extracted model file (model.onnx) and the topology (model.json).

​
Downloading

download returns one artifact’s bytes inline, in the response. An artifact larger than 4 MiB comes in ranged parts over presigned URLs instead. describe lists the three artifacts by name and size without fetching any of them, so you know which way to fetch one before you start.

​
From Python

The Python SDK’s onnx module does all of this in a call each. upload() packs a .onnx file, sends it in one request or in parts by its size, and waits on the operation. download() fetches the extracted model and caches it for a committed revision. topology() returns the parsed graph.
A public model needs no account. The web app’s Labs page lists the public ones: Object Detection, Super Resolution and BiDAF. Four notebooks ship in the blobhub package: three download those models and run them on your machine with ONNX Runtime, and the fourth exports a model with PyTorch and uploads it to a blob of yours.