EtienneMueller / Tensorista

Re-implementation of TensorFlow and Keras in Python and Numpy for use in Pythonista on iOS

Tensorista

Welcome to my fun little side project!

Some time ago I downloaded Pythonista to be able to tinker around with little python project when I am on the move. Unfortunately, TensorFlow is not supported and if you want to play around with neural networks, you have to implement them with Numpy by yourself. I did that in the beginning, but really liked the idea to be able to reuse the code I wrote on desktop. So, I started putting everything in the same structure as Keras so it is possible to simply exchange import tensorflow as tf with import tensorista as tf and keep the remaining code as is.

So far, it can do everything needed for a very basic neural network:

• Download MNIST and convert it to a numpy array
• Create a sequential model with dense layers
• Print a model summary
• Compile the model
• Train the model
• Evaluate the model

How to use in Pythonista

There a two ways: the easy one and the better one. The easy one would be to copy the tensorslow folder into site-packages from Pythonista. Then you can import everything like a normal module. But if you want to update, then you have to do it manually (although I might not work on this project that often). The nicer way is to use Working Copy as git client. As iOS is a bit strict you have to clone the repo the following way:

• press the plus button and chose "Setup synced directory"
• chose the local Pythonista folder
• then under "Repository" choose "Add Remote" and paste the URL to this repository
• Save, fetch and after a couple of seconds you should be able to see the folder in Pythonista

To Do

There is A LOT I want to add to this project! I mean, so far you can just create feedforward networks for MNIST. So here is what to come, as soon as I find the time:

• CIFAR10 / CIFAR100
• CNNs
• Dropout
• Adam and other optimizers
• Other losses and metrics
• ...

Structure of Keras (in bold the upcoming additions):

• tensorista
  • keras
    • dataset
      • boston_housing.py
      • cifar.py
      • cifar10.py
      • cifar100.py
      • fashion_mnist.py
      • imdb.py
      • mnist.py
      • reuters.py
    • engine
      • base_layer.py
        • Layer
      • training
        • Model
    • utils
      • generic_utils
        • Progbar
      • np_utils
      • to_categorical
      • normalize
    • backend
      • sigmoid
      • d_sigmoid
      • relu
      • d_relu
    • layers
      • Dense(Layer)
      • ReLU(Layer)
    • losses
      • SparseCategoricalCrossentropy
    • metrics
      • SparseCategoricalAccuracy
    • models
      • Functional
      • Sequential
    • optimizers
      • RMSprop
  • module
    • module
  • random