d-lowl / fantastic-bassoon

KotlinDL Softmax layer demo

fantastic-bassoon

This repo is a demonstration of Softmax layer implementation in KotlinDL.

Python

The script in ./python directory creates a simple network for MNIST classification utilising a Softmax layer, and exports it as a JSON file

Environment setup:

python -m venv env/
source env/bin/activate
pip install -r requirements.txt

Running:

python main.py

Copy the resulting keras_model.json from ./python to ./kotlin directory

Kotlin

The project in ./kotlin directory uses softmax-layer branch of d-lowl/KotlinDL repo, and has two functions to run

Training

space.dlowl.example.TrainKt imports MNIST dataset, builds a network, and trains it. Expected behaviour: no errors, accuracy is close to 1.

./gradlew runTrain

Importing

space.dlowl.example.ImportKt imports the previously exported model (i.e. ./kotlin/keras_model.json), compiles it, and prints the summary. Expected behaviour: no errors, the model summary is correctly printed

./gradlew runImport