Reasonable number of image data to be trained on a custom model.

ryota.matsuda 140

I understand that this is a situation-specific number of data, but I need to know what the market rate is.

In the tutorial, I was able to train from 5 invoices. How many data sets do I need to train to accommodate various invoice types? For example, how much if I want to achieve one Model that can handle 10 different invoice types?

romungi-MSFT 42,791 Reputation points Microsoft Employee

2023-06-14T11:10:51.86+00:00

@ryota.matsuda If you have a scenario where your invoices are of different visual formats but the data to be extracted is the same, then you might need to use the custom neural model build type to train different models of the same type and then compose as a single model id with all the trained models of custom neural type.

A composed model can have up to 100 different models composed under one model id. So, you can have 100 different invoice types to recognize with one composed model id where each model under the composed model trained to recognize different invoice format. Please see the table on this page for limits of composed models. The minimum number of forms or images to train each model is 5, if you are able to achieve high accuracy with the 5 images then you need not add additional training images.
romungi-MSFT 42,791 Reputation points Microsoft Employee

2023-06-20T06:51:03.5933333+00:00

@ryota.matsuda Did you get a chance to check if the above response helped?
ryota.matsuda 140 Reputation points

2023-06-20T10:07:11.59+00:00

I understand that up to 100 different models can be used as one composite model.

Please tell me more.

Is there any possibility of overall accuracy loss due to synthesis?

For example, will the accuracy of Model 12, which is a composite of Model 1 and Model 2, be less accurate than Model 1, which specializes in Layout A, and Model 2, which specializes in Layout B, each held as an independent ID?
romungi-MSFT 42,791 Reputation points Microsoft Employee

2023-06-21T07:40:28.19+00:00

@ryota.matsuda Each model is held independently under a composed model id, when a document is submitted for analysis using a composed model, the service performs a classification to decide which custom model best represents the submitted document and returns the results accordingly. To check if the composed model is a list of added models, you can use GET MODEL API to check the list of models composed under it.

In your example, if you provide a document to a composed model that best fits Model 12 the result from analyze operation should represent the response as expected from Model 12. I hope this helps!!

Share via

Reasonable number of image data to be trained on a custom model.