Building sequence to sequence (S2S) models using previously generated S2S models with similar use cases

1. A computer-implemented method comprising:

receiving, at a sequence to sequence (S2S) algorithm generator, via an application programing interface (API), a request to build a S2S model for a use case, wherein the request includes or identifies at least a training data set;

determining, by an input data analyzer of the S2S algorithm generator using the training data set, and using previous model data for previously generated S2S models, a set of use cases in previously generated S2S models that are similar to the use case of the request, wherein a use case at least defines network primitives;

receiving, by a parameter analyzer of the S2S algorithm generator, the set of use cases in the previously generated S2S models;

determining, by the parameter analyzer using the previous model data and the set of use cases in the previously generated S2S models, parameters that apply to the set of use cases in the previously generated S2S models;

receiving, by a hyperparameter recommendation engine of the S2S algorithm generator, the set of use cases in the previously generated S2S models;

determining, by the hyperparameter recommendation engine using the previous model data and the set of use cases in the previously generated S2S models, hyperparameters that apply to the set of use cases in the previously generated S2S models;

receiving, by an algorithm builder of the S2S algorithm generator, the parameters that apply to the set of use cases in the previously generated S2S models and the hyperparameters that apply to the set of use cases in the previously generated S2S models;

building, by the algorithm builder using the parameters that apply to the set of use cases in the previously generated S2S models and the hyperparameters that apply to the set of use cases in the previously generated S2S models, a S2S algorithm;

storing the S2S algorithm as a part of a container;

executing code in the container using the training data set to generate the S2S model; and

storing information regarding parameters, hyperparameters, and network primitives for the S2S model.

2. The computer-implemented method of claim 1 , wherein the network primitives include an encoder and a decoder, and at least one of the determined hyperparameters is related to a number of layers of the encoder and the decoder.

3. The computer-implemented method of claim 2 , wherein the network primitives further include one or more of an attention layer, a loss function, and an accuracy/quality metric.

4. A computer-implemented method comprising:

receiving, at a sequence to sequence (S2S) algorithm generator, via an application programing interface (API), a request to build a S2S model for a use case, wherein the request includes or identifies at least a training data set;

determining, by an input data analyzer of the S2S algorithm generator using the training data set, and using previous model data for previously generated S2S models, at least one use case in the previously generated S2S models that is similar to the use case of the request, wherein a use case at least defines network primitives;

determining, by a parameter analyzer of the S2S algorithm generator using the previous model data and the at least one use case in the previously generated S2S models, any parameters that apply to the at least one use case in the previously generated S2S models;

determining, by a hyperparameter recommendation engine of the S2S algorithm generator using the previous model data and the at least one use case in the previously generated S2S models, any hyperparameters that apply to the at least one use case in the previously generated S2S models;

receiving, by an algorithm builder of the S2S algorithm generator, the parameters that apply to the at least one use case in the previously generated S2S models and the hyperparameters that apply to the at least one use case in the previously generated S2S models;

generating, by the algorithm builder using the parameters that apply to the at least one use case in the previously generated S2S models and the hyperparameters that apply to the at least one use case in the previously generated S2S models, parts of a S2S algorithm; and

training another S2S algorithm built from the parts of the S2S algorithm using the training data set to generate the S2S model.

5. The computer-implemented method of claim 4 , further comprising storing information regarding parameters, hyperparameters, and network primitives for the S2S model.

6. The computer-implemented method of claim 4 , further comprising pre-processing data prior to training.

7. The computer-implemented method of claim 4 , wherein the network primitives include an encoder and a decoder, and at least one of the determined hyperparameters is related to a number of layers of the encoder and the decoder.

8. The computer-implemented method of claim 7 , wherein the network primitives further include one or more of an attention layer, a loss function, and an accuracy/quality metric.

9. The computer-implemented method of claim 4 , wherein the network primitives comprise one or more of convolutional neural network encoders, recurrent neural network encoders, convolutional neural network decoders, recurrent neural network decoders, dot product attention, scaled attention, cross entropy loss function, smooth cross entropy loss function, perplexity, and bilingual evaluation understudy (BLEU).

10. The computer-implemented method of claim 4 , wherein the S2S model is one of a translation model, summarization model, and a speech-to-text model.

11. The computer-implemented method of claim 4 , wherein determining any hyperparameters that apply to the at least one use case is performed by a search.

12. The computer-implemented method of claim 4 , wherein the built S2S algorithm is a part of a container.

13. The computer-implemented method of claim 12 , wherein training the S2S algorithm comprises:

initializing a virtual machine instance;

retrieving the training data;

initializing the container in the virtual machine instance; and

executing code stored in the container using the training data.

14. The computer-implemented method of claim 4 , further comprising performing an inference using the S2S model.

15. A system comprising:

a client device that is configured to send a request to build a sequence to sequence (S2S) model for a use case, wherein the request includes or identifies at least a training data set; and

a web services provider system including a S2S algorithm generator configured to receive, via an application programing interface (API), the request and, in response to the request,

determine, by an input data analyzer of the S2S algorithm generator using the training data set, and using previous model data for previously generated S2S models, at least one use case in the previously generated S2S models that is similar to a use case of the request, wherein a use case at least defines network primitives;

determine, by a parameter analyzer of the S2S algorithm generator using the previous model data and the at least one use case in the previously generated S2S models, any parameters that apply to the at least one use case in the previously generated S2S models;

determine, by a hyperparameter recommendation engine of the S2S algorithm generator using the previous model data and the at least one use case in the previously generated S2S models, any hyperparameters that apply to the at least one use case in the previously generated S2S models;

receiving, by an algorithm builder of the S2S algorithm generator, the parameters that apply to the at least one use case in the previously generated S2S models and the hyperparameters that apply to the at least one use case in the previously generated S2S models;

generate, by the algorithm builder using the parameters that apply to the at least one use case in the previously generated S2S models and the hyperparameters that apply to the at least one use case in the previously generated S2S models, parts of a S2S algorithm; and

train another S2S algorithm built from the parts of the S2S algorithm using the training data set to generate the S2S model.

16. The system of claim 15 , wherein the web services provider system is to store information regarding parameters, hyperparameters, and network primitives for the S2S model.

17. The system of claim 15 , wherein the web services provider system is to pre-process data prior to training.

18. The system of claim 15 , wherein the network primitives include an encoder and a decoder, and at least one of the determined hyperparameters is related to a number of layers of the encoder and the decoder.

19. The system of claim 18 , wherein the network primitives further include one car more of an attention layer, a loss function, and an accuracy/quality metric.

20. The system of claim 15 , wherein the network primitives comprise one or more of convolutional neural network encoders, recurrent neural network encoders, convolutional neural network decoders, recurrent neural network decoders, dot product attention, scaled attention, cross entropy loss function, smooth cross entropy loss function, perplexity, and bilingual evaluation understudy (BLEU).