DigitsClassifier

It is an attempt to classify digits drawn in an Android app from 0 to 9.

V1

1. Retrained mobilenet TF model for MNIST dataset to classify digits.
2. Accuracy was ~60% because of the difference in the digits drawn and the dataset.

V2

1. Trained the dataset created through the first app in a CNN model. Dateset is available here.
   
2. Saved weights in summary.h5 file, created the Graph.pb TF file using amir-abdi/keras_to_tensorflow. And then finally converted the Graph.pb file to a optimized_graph.lite file using tflite_convert(a TF utility).
   
3. The optimized_graph.lite file can be directly used in the app, which can now classify drawn digits with a ~60-70% accuracy.
   
4. The mobilenet model re-trained for the given dataset can classify drawn digits with accuracy upto ~90%.
   

Tensorboard Visualization of Model

Resources used

1. Keras Documentation
2. Tensorflow For Poets
3. Drawing View Library.
4. amir-abdi/keras_to_tensorflow
5. TFLite Convert