System and method for heterogeneous model composition

1. A method comprising:

determining a set of connected heterogeneous models comprising a parent model and a set of child models, wherein the heterogeneous models are connected in a directed acyclic graph, wherein each model is represented as a subgraph within the directed acyclic graph, wherein each node of each subgraph is associated with a model layer;

receiving a selection of an output from a child model of the set;

identifying a path within the directed acyclic graph between the parent model and the child model, the path comprising the model layers of a subset of models between the parent model and the selected output;

receiving an input;

for each successive model in the subset of models:

a) converting the input into a standard format;

b) converting the standard-formatted input into a model-specific input (MSI) format for the respective model;

c) executing the model within the set of models using the MSI-formatted input data to generate an output in a model-specific output (MSO) format for the respective model;

d) converting the output from the MSO format to the standard format; and

wherein a)-d) are repeated for each successive child model of the subset of models, using the standard-formatted output from a prior model as the standard-formatted input data for the child model, wherein a)-d) are only performed using the identified model layers of the models of the subset; and

converting the standard-formatted output of the selected child model to an output datatype.

2. The method of claim 1 , further comprising:

converting the standard-formatted output for a model from the set to an output datatype; and

providing the output datatype as a final model output.

3. The method of claim 1 , wherein the set of heterogeneous models are each authored by different entities.

4. The method of claim 1 , wherein the set of models comprise at least one of neural network models, regression models, or ruleset models.

5. The method of claim 1 , wherein the standard format is a tensor.

6. The method of claim 1 , wherein the MSO format of the parent model is incompatible with an MSI format of a child model.

7. The method of claim 1 , wherein the selected output is an output variable, wherein the input is an input value, and wherein a)-d) are performed using input values and output values.

8. A heterogeneous model composition system, comprising:

a non-transitory computer readable medium; and

a processor coupled to the non-transitory computer readable medium, the processor configured to:

determine a model composition, comprising a series of heterogeneous models connected by inputs and outputs, wherein each model is associated with a handler configured to convert an input for the model from a standard object to a model-specific input (MSI) object and convert a set of outputs of the model from a model-specific output (MSO) object to the standard object, wherein the series of heterogeneous models cooperatively form a directed acyclic graph;

present the outputs of the models of the model composition to a user;

receive a selection of an output from an intermediary model of the series;

determine a subset of models connecting a root model of the model composition to the intermediary model and a subset of model layers connecting the root model to the selected output of the intermediary model;

facilitate execution of the model composition, wherein only the subset of models and the subset of layers are executed, wherein model composition execution comprises:

converting input data into a standard input object;

for each successive model:

converting the standard input object into the respective MSI object using the respective handler;

executing the model using the MSI object to generate a set of MSO objects; and

converting the set MSO objects to standard output objects using the respective handler; and

providing the standard output object from the selected intermediary model to an endpoint.

9. The system of claim 8 , wherein model composition execution further comprises converting the standard output object for a model to an output datatype.

10. The system of claim 8 , wherein determining the series of heterogeneous models comprises receiving an identifier for each model within the series and the connections between the inputs and the outputs from a user at a user interface.

11. The system of claim 8 , wherein each model within the series is authored by a different entity.

12. The system of claim 11 , wherein the series of heterogeneous models are stored in a model repository, and wherein the processor is further configured to access the model repository and execute the model composition.

13. The system of claim 8 , wherein the series of models comprises at least one of neural network models, regression models, or ruleset models.

14. The system of claim 8 , wherein the standard object is a tensor.

15. The system of claim 8 , wherein the MSO object of a preceding model in the series is incompatible with an MSI object of a succeeding model in the series.

16. The system of claim 8 , wherein the models are executed in a different computing environment from the processor, wherein facilitating execution of the model composition comprises, for each successive model in the series: sending an output identifier for the standard output object, output by a preceding model, to the computing environment in association with a model identifier for the successive model, wherein the successive model uses the output identifier to access the standard output object.