In composing models, can you compose models that have somewhat different labels in training?

Bryce Philbert Salvador 20

For example in your first model you have x,y,z as your training labels but in the second model you will have x,y,z,a,b as the training labels. Is that possible? If yes will it affect the composed model accuracy?

Ramr-msft 17,731 Reputation points

2023-10-18T16:24:06.51+00:00

@Bryce Philbert Salvador Thanks for the question, Yes, it is possible to have different training labels for different models. In your example, the first model is trained on labels x, y, z while the second model is trained on labels x, y, z, a, b.

As for the effect on the composed model’s accuracy, it can vary. The additional labels (a, b) could potentially improve the accuracy of the second model if they provide meaningful information for the prediction task. However, if these additional labels introduce noise or irrelevant information, they could also potentially decrease the model’s performance.

It’s also important to note that when composing models (i.e., using one model’s output as input to another), the compatibility of their inputs and outputs needs to be ensured. If the first model expects only x, y, z as input but receives x, y, z, a, b from the second model, it may not function properly unless it’s designed to handle or ignore the extra inputs.

In conclusion, while it’s possible to train different models on different sets of labels, careful consideration should be given to how these models are used together and how their performance might be affected. It’s always a good idea to validate your models using a separate test set to ensure they’re performing as expected.
Bryce Philbert Salvador 20 Reputation points

2023-10-18T19:03:03.16+00:00

Thank you for your prompt response, Also is it recommended to put different formats of the same document in one model but they have the same types of data to be obtained or do I just train those documents in another model all together but compose them later in one model to be used in production? Thank you.
Bryce Philbert Salvador 20 Reputation points

2023-10-30T20:33:27.7833333+00:00

I see, thank you for the answer!

1 answer

Ramr-msft 17,731 Reputation points

2023-10-30T04:49:44.1233333+00:00

@Bryce Philbert Salvador Thanks for the details, It's generally recommended to include all variations of a document type in the same model, even if they have different formats, as long as they have the same types of data to be obtained. This is because custom neural models can generalize across different formats of a single document type, and including all variations in the same model can help improve the accuracy of the model.
Please sign in to rate this answer.

0 comments No comments
Sign in to comment

Share via

In composing models, can you compose models that have somewhat different labels in training?

1 answer