Digit recognizer trained on MNIST dataset using TensorFlow 1.0 and Keras in Python 3.6.0.
Data Source : http://yann.lecun.com/exdb/mnist/
References :
https://www.tensorflow.org/get_started/mnist/pros
https://github.com/aymericdamien/TensorFlow-Examples
https://keras.io/getting-started/sequential-model-guide/
https://github.com/fchollet/keras/tree/master/examples
mnist-tf-mlp.py, mnist-tf-cnn.py, mnist-tf-rnn.py
- Implementation of deep neural nets (multi-layer perceptrons, convoluntional neural net and recurrent neural net) are trained against 55000 handwritten digit images using TensorFlow 1.0.
- Evluation is made on the test set of 10000 handwritten digit images.
mnist-keras-mlp.py, mnist-keras-cnn.py, mnist-keras-rnn.py
- Implementation of deep neural nets (multi-layer perceptrons, convoluntional neural net and recurrent neural net) are trained against 60000 handwritten digit images using Keras.
- Evluation is made on the test set of 10000 handwritten digit images.
| Methods | Test accuracy (TensorFlow) | Test accuracy (Keras) |
|---|---|---|
| Multi-layer perceptrons | 94.4.% | 97.6% (epochs=10) |
| Convolutional neural network | 98.0.% | 98.8% (epochs=5) |
| Recurrent neural network (LSTM) | 97.6.% | 97.5% (epochs=5) |