Quickstart

A Sihl AI project is associated to a single dataset stored at any S3-compatible object storage.
We support cloud solutions like AWS S3 and on-premise solutions like MinIO. In either case, you retain full control over your data, and we only ever access it for model training.
By default, we randomly split off 10% of the dataset to validate the model during the training process. However, you can optionally link a project to explicit training and validation sets, which is useful to validate generalization ability.
We supply foundation datasets from which you can create pre-training projects easily. These datasets are made available for everyone to conveniently train on, but the underlying images cannot be downloaded for legal reasons.
Datasets must be annotated to train most tasks. Only self-supervised tasks like autoencoding or view-invariance learning support unlabeled datasets of images.
Right before a model starts training, the runner will download the dataset onto its local disk to accelerate the training process. Once the model is done training, the local copy is deleted. We will never keep, share, or sell your datasets.
A project admin (by default, its creator) can change the dataset storage details anytime.

Once a project is created, you can start training models by submitting training jobs, that one of our GPU runners will pick up and perform.
A training job is configured by hyperparameters, which have default values you can adjust to precisely control the training process.
Most of these are hidden under the collapsible fold, but the most important ones are always shown, and we recommend you always check these. They include:

• weight initialization (imagenet or a checkpoint from any of the projects you are a member of)
• training duration (in iterations and/or hours, to cap the cost of the job)
• image size (all training images need to be resized to be batched for training)
• backbone architecture (from small and fast to big and performant)

Your job will run immediately if you do not already have a running job in the same project, in which case the new job will be queued until the first one stops. This means that all members of a project can each have one job running concurrently. A given user can also run concurrent jobs in different projects. If you need to run concurrent jobs in a single project, let us know at support@sihl.ai. While the model trains, you can view its live-updated metrics and logs, and compare those with other models of the project. You can also "duplicate" the job, copying over all the hyperparameters to let you change just what you need before submitting.

Once a model is trained, you have several options to run it.
To quickly test it out, you can just drag-and-drop an image into your browser. Your browser will then automatically forward it to an endpoint serving your model and return its outputs. This is the easiest and quickest way to run inference, but it has a couple drawbacks. First, you cannot process a batch of images this way, it only supports single image inference. Second, the first call to this endpoint will incur extra latency due to needing to download your model before running it; subsequent calls will be much faster due to your model being cached. Alternatively, you can call the API endpoint programmatically, from a shell script or Python for example. The same limitations as above apply here.
If you need consistently low latency, batched inference, or to run inference at the edge, you can export the model and run your own inference server. We support ONNX (FP32 or INT8) and TensorRT (FP16) formats optimized for inference, and pytorch state dict (to be used with our open-source code) for other use-cases.
Exported model completely are license-free - you can use them however you want, including commercially!