5.1.1. My Model

My Models is used to manage models (for storage) that pass through the platform, and the current platform supports three types of model management (storage): base model (base file model), image model, and composite model. The composite model is obtained by combining multiple file models.

The left half of the model management page displays the filter criteria, including:

Modalities: Multimodality, Image, OCR, Text, 2D Table, Speech Recognition

Model Type: Mirror, Combination, Foundation

Model frameworks: ONNX, PMML, Transformer, PEFT, and Pickle

Model tags: user-defined tags, such as "subtasks"

Task Type: displays the task types in the selected mode, as detailed in the following table:

The right half of the atomic model page displays the model, which is available in two formats: cards and tables. Click the button in the upper right corner to switch the display format.

5.1.1.1. Export and import of models

After clicking "Export", you need to select the version you want to export in the pop-up window (you can select more than one), and finally click "Export" to get the model file zip package. If you want to export multiple models quickly, you can check the required models and click the "Batch Export" button to export all versions of the selected models.

Users can click on the "Import" button on the page and upload the zip package. Subsequently, the platform parses the zip package, displaying the model name, version name, whether there are duplicates, and the model type. If there are duplicate models, you need to overwrite or skip the duplicate models according to the situation.

5.1.1.2. Create a new file model

If you want to add a model, you need to click the "Add" button, fill in the model name, task type, and model type according to the model type (file or mirror), and then click "OK". You can also fill in the model description and model labels, and upload the model cover.

Thereinto:

The task type needs to be selected based on the actual functions of the model, which will affect the display of data dependencies and the testing functions of the model after deployment.

The model type needs to be selected according to the format of the model file. Currently, the model supports PMML, ONNX, Transformer, PEFT, and Pickle, and the PMML and ONNX models support automatic parsing of input and output parameters.

5.1.1.2.1. Create a new version

After creating a new model, you can click Add and Create Version to complete the creation of the first version of the model, if there is no version, the model will not be available. Go to the model details, switch to the Model Version tab, and click Create Version.

In the pop-up window, you can choose to create a new independent version or fork from an existing version (which means that the current model already has a version and can be incrementally updated based on the existing version). Currently, you can upload files from the file system, s3, HF, ModelScope, and local files, and you can choose different upload methods according to different file sizes.

If the model has some training data files or other attachments, you can click the attachments to upload them together.

When you create a new version, you need to enter the version alias, but you do not need to enter the version number (the version number is automatically generated in the format of V+ digits, for example, V2). Then you need to select the hardware framework, here if the model has no requirements for the hardware framework, then please select "General", if there are requirements, then please choose AMD64 and ARM64 according to the actual situation. Finally, upload the model file and fill in the input and output parameters. For different types of file models, the input parameters and output parameters are slightly different, and some additional parameters can be configured for large models, which will be explained in detail next

If the model type is File Model-PEFT, the model is fine-tuned based on the pre-trained model, and the model needs to depend on the pre-trained model. Therefore, when you create a PEFT model, you need to select a pre-trained model that depends on it. The rest of the parameters are consistent with the Transformer model.

Click "Add" when you're done, and you'll see the version in the list.

5.1.1.3. Create a mirror model

Click Create on the list page and fill in the basic information of the model when you create the file model, except that when you create an image model version, you can fill in the corresponding platform image source address of the model.

After the API is created, you can go to the interface configuration module to configure the interface for image inference.

5.1.1.4. Create a new composite model

Click New Model to select "Combination" model, fill in the model name and select the task type in turn, because the combined model is a combination of multiple basic models, you need to declare the input and output types of the model in advance, as shown in the following figure for the object detection model:

Thereinto:

Task types: You can choose from a variety of task types for the combined model. The choice here affects the presentation of data dependencies, the testing capabilities of the model after deployment, and the types of tasks that can be selected for the atomic model.

Task Input: You need to select one of the selected Task Types as input.

Task Output: You need to select one of the selected Task Types as the output.

After confirmation, select Create Model Version, and enter the version name and assembly method of the combined model. If the new version is modified based on an existing version, you can select Fork from an existing version, otherwise select Create Independent Version. Then fill in the version alias, select the hardware framework, and assembly method. In the Ensemble pattern, multiple atomic models are concatenated one after the other, and the two models are connected by a copy of python code. The BLS pattern is to call multiple atomic models with a single python code. Take ensemble as an example:

When you're done, you'll be redirected to the version editing page. Please add the atomic model and glue code (python code) to this page. Atomic models and glue codes can be dynamically added based on actual model needs.

After adding the atomic model, you can add the glue code (python code). Here you need to specify the folder name of the glue code, click "OK" after filling in, and you can see the folder on the left.

In the text editor, you can select "Compare" to edit diffs for multiple versions of the file to quickly locate the changes.

After the atomic model is added and the code is added and modified, you can click the "Save and Return" button in the upper right corner of the page. If you have completely modified everything and will not make any further changes, you can click the "Confirm Submission" button and mark the version as "Submitted". You will then be able to see the version in the list of versions.

5.1.1.5. Model Traceability

The platform supports managed model traceability management, and you can view the basic model trained by the current model and the information of the trained dataset through the relationship graph of model details (provided that the dataset and model are managed through the platform)

5.1.1.6. Resource Estimates

After the model file version is successfully created, the platform provides a resource plug-in to calculate the recommended resources consumed by the current model for fine-tuning training and deployment.

5.1.1.7. Model Deployment

Click the "Deploy" button in the model details version list to enter the deployment form. For details, see the Deployment Service section of the Model Service