How to Train Your Tensorflow

How to Train Your Tensorflow

Project Description

How to Train Your Tensorflow

Some of you may know our tool TrainYourTesseract. It’s a service by Anyline for creating a tesseract trained data file for your use case, which you can include on our Anyline SDK Module.

As you may have noticed, we switched most of our modules to a machine learning approach. That’s why we’re using Tensorflow. In order to provide the same service as we did with TrainYourTesseract, we are currently developing TrainYourTensorflow.

What can I do with TrainYourTensorflow?


This service aims to make your workflow easier. For example, you have a specific use-case for the Anyline OCR Module and want the SDK to be trained for your use-case (for example a specific font, that your application should scan). Of course, the general Anyline OCR Module has that covered in most cases. But as always, a small retraining of our OCR Module to your needs could improve your Anyline experience.

So, instead of doing lots of research on machine learning on your own, we let you automate this process. In TrainYourTensorflow you can easily upload a font-file and after a short amount of time, you get your personalized Tensorflow model via email. You can then implement it in your Anyline app as an asset.

How does it work?

We created a python server, which receives the font file uploaded by our frontend. Out of this file, millions of training images are generated with different mutations (Data Augmentation). For example:

Translations

Rotations

Distortions

Scaling

We’ve combined some of them and used them to retrain Anyline with the generated data (this is called transfer learning). The basic layers in our networks, which classify basic character shapes are already trained and ready for your use case (because they won’t change much in different fonts). The data augmentation for the pre-training and the training is also different. As in the pre-training, the lower convolutional layer is trained. This means that it’s important to have a lot of variance in the training set. Also, different fonts are used, to show the network that classes can differ. The higher layer consisting of a fully-connected layer, which is used for detecting specific details of a character, is retrained to your needs.

So, after training with your font, we create a model specified to fit into your workflow, and we send it to you!

When can I use it?

Yes, we know. We’re very excited too to release this new service so you can try it out and see how it all works, but there is still a lot to do to run a bullet-proof service that performs perfectly under this workload.

You can stay up-to-date on all TrainYourTensorflow news by signing up to the free mailing list below. We’ll inform you when the tool is ready for release and keep you informed about any updates or changes to the tool!

Stay Up to Date!

Be the first to find out when TrainYourTensorflow is released!

Get TrainYourTensorflow

QUESTIONS? LET US KNOW!

If you have questions, suggestions or feedback on this, please don’t hesitate to reach out to us via FacebookTwitter or simply via [email protected]! Cheers!

Recent Posts

Sign up for our Newsletter!

Be the first to receive the latest news, release notes and updates!

SUBSCRIBE