A Neural Network (NN) implemented in Scala from scratch.
This is just the initial implementation of a vanilla NN that classifies images of ... cats!
I tried to do everything in a functional way, avoiding mutation and side effects. A cute idea to toy with is that forward prop is foldLeft and back prop is foldRight.
The initial code layout follows Deep Learning Coursera's course by Andrew Ng code organization into functions (Python). I use the same terminology when naming variables and same network structure with forward prop caches, etc. I also used the data for testing from the course.
This is only the start. There are many improvements that can be made.
To learn by implementing and to see design tradeoffs in a typesafe statically typed language - Scala.
If I find time I'll introduce better abstractions for all configurable algorithms and settings. I also have some ideas how to define NN in a more typesafe way than this or regular frameworks do.
Have java runtime JVM 8 installed. To build run from shell:
> ./sbt assembly
This will download dependencies, compile the code, run all the tests and create a self-contained JAR file. After the build is done run the JAR with:
> java -jar target/scala-2.12/scala-nn-assembly-0.1.jar
Default settings can be changed via -D Java arguments. For more info see reference.conf file.
Each image is of size (width_px, height_px, 3) - 3 channels (RGB).
Number of training examples: 209
Number of test examples: 50
Each image is of size: (64, 64, 3)
train_x_orig shape: (209, 64, 64, 3)
train_y shape: (1, 209)
test_x_orig shape: (50, 64, 64, 3)
test_y shape: (1, 50)
train_x's shape: (12288, 209)
test_x's shape: (12288, 50)
12,288 equals 64×64×3 which is the size of one reshaped image vector.