https://github.com/jsflo/androidnumberclassifier

Uses a model trained on with Tensorflow. Inference is done on the device through the Tensorflow Java Api
https://github.com/jsflo/androidnumberclassifier

android kotlin kotlin-android machine-learning tensorflow tensorflow-android

Last synced: about 1 month ago
JSON representation

Uses a model trained on with Tensorflow. Inference is done on the device through the Tensorflow Java Api

• Host: GitHub
• URL: https://github.com/jsflo/androidnumberclassifier
• Owner: JsFlo
• Created: 2017-09-17T01:38:12.000Z (over 7 years ago)
• Default Branch: master
• Last Pushed: 2018-03-28T00:07:46.000Z (over 6 years ago)
• Last Synced: 2024-04-20T17:21:01.099Z (8 months ago)
• Topics: android, kotlin, kotlin-android, machine-learning, tensorflow, tensorflow-android
• Language: Kotlin
• Size: 29.7 MB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Number Classifier

Basic demo showing how to include and run a custom

TensorFlow model.

|   |   |

Created a Classifier class to help speed up adding new models.

The classifier interface and abstract class can be extended

to use **any** method of classifying, not necessarily TensorFlow

or even **machine learning** classification.

## Note

The v1 model is based on the **no pooling** CNN.

The classification is sometimes off and I believe this has

to do with the way we pull the pixels from the Android

Bitmap.

* Color values: Model was trained with gray scale (1 dimension)

* Shape: Model was trained with a [1, 784] single array that gets reshaped

to a 28 x 28. In this app I do the **reshaping** on the device and just pass

a 28 x 28 image to the classifier. 

The closer I can get the user input (drawn on image view) to the 

input data that the model was trained on then the classification results should be better.

(Or try to see if the models are **too** **over-fit** to the input data)