The OCI portfolio allows us to consume models that have been prepared, to develop their own models or to manage infrastructure for specialized loads. The decision depends on how much work you have to control.
Objectives: To distinguish AI Services, Data Science and infrastructure recognition of the model cycle explain GPU and RDMA and apply AI owner principles to a particular case.
Key concepts
| Requirement | Retirement plate | The liability that you retain |
|---|---|---|
| Extracting information with a known capacity | AI Services | Prepare data, permissions and review results |
| Training and operation a model of your problem | OCI Data Science | Data, algorithm, evaluation and deployment |
| Control a dedicated compute load | AI infrastructure | Environment, distribution and operation by service |
The preset services are consumed with console, API, SDK or CLI. Some allow custom models. You don't have to train a network from scratch for every task. Digital Assistant organizes contiguous and skills experiences isn't another name for a GPU, a notebook or the Language service.
OCI Data Science and model cycle
Projects They have jobs. The notebook sessions provide a JupiterLab environment to explore and train. A Conda environment manages dependencies. He's not a model. The SDK ADS assistance with data cycle and model tasks.
The Model Catalog keeps artifacts and metadata for monitoring and collaboration. A Model Deploment Expand an inference model through an endpoint. A job performs a repeatable task, such as data production or training. A pipeline coordinates phases. Save in the catalogue does not automatically create an active endpoint.
To reproduce a result, preserve the data version, code, dependencies and configuration. In the Data Science demonstration, ADS prepares the artifact, checks its prediction capability and saves it; deployment and the prediction call are later steps. See the concepts of OCI Data Science.
Infrastructure and AI owner
One GPU conduct a number of similar operations at the same time, enabling master calculation of training and inference. Not every task improves by adding GPU: they influence data size, memory, transfers and software. The CPU continues to participate in other parts of the work.
For distributed training, communication between nodes can restrict performance. RDMA allows minor CPU intervention transfers. RoCE transport RDMA over Ethernet. The town reduces communication tours. Do not confuse links within the node with the network between nodes and a Supercluster with an application endpoint.
The material sums up AI reliable as lawful, ethical and robust. In practice, it defines responsibility, data controls, group assessment, appropriate explanations, human monitoring and monitoring after deployment. A mean accuracy can hide worse performance for a population or condition of use. A technically precise model can also be used with an inappropriate goal.
Step-by-step example
The company wants to estimate the repair time with its own history.
- Define the numerical output and the decision that it will support. check that the features exist before the repair.
- It organizes a Data Science project and opens a notebook session with the necessary environment.
- It trains a simple reference and evaluates later repairs without filtering future information.
- Prepare and check the artifact, register the model and its metadata in the catalogue.
- If an application needs online predictions, create an deployment if recalculates every night, consider a job.
- It monitors errors by type of equipment and preserves a human correction path.
If the problem changes to transcribe audios, start evaluating Speech. You don't have to have a model of your own just because you know Data Science.
Common errors
- To merge catalogue with deployment or a job with an always available endpoint.
- To believe that a notebook session removes the need for IAM and data access.
- To memorize GPU, date of availability or cluster sizes of a recording as current boundaries.
- Treat "managed" as an absence of cost or liability.
- Add an ethical statement without evaluating results and claim mechanisms.
Election exercise
A computer should transcribe 300 recordings, train an own classifier and conduct its appraisal every week. Assign a capacity to every need.
Solution
Evalúa Speech for in batches transcription uses Data Science to develop the classifier. uses jobs or a pipeline to repeat the evaluation. The catalogue retains the model. If you later require online inference, add a deployment. A Supercluster does not have to be administered to justify these tasks.
Sources and review
Check OCI AI Portfolio and display Data Science at MyLearning. Query AI infrastructure of Oracle for models of GPU and currently available. The figures for a demonstration are examples of its date.