clj-dl4j

A clojure wrapper around DL4J, a well-known Java ML library built upon ND4J (the "Java NumPy") and JavaCPP.

Note

This library is at early stages of development and is so highly subject to future breaking changes. Futhermore, functionnalities are added carefully to avoid oversimplification or a rushed copied/pasted mess. Indeed, the library tries to follow DL4J beta releasing and so avoid to implement marginal features which are doomed to refactoring or worse.

Usage

Dependencies

This library being under AOT, you might skip the following libraries explicit inclusions which are not mandatory to test examples.

[jeremylcfr/clj-nd4j "0.1.0-SNAPSHOT"]
[jeremylcfr/clj-datavec "0.1.0-SNAPSHOT"]

Then, include :

[jeremylcfr/clj-dl4j "0.1.0-SNAPSHOT"]

You can theoritically use CUDA but I never tested it personnaly. For that, you just have to include :

[org.deeplearning4j/deeplearning4j-cuda-{{CUDA_VERSION}} "1.0.0-beta6"]

with {{CUDA_VERSION}} from 9.2 to 10.2 included. In that case, this library works the same (in theory again) with some optional parameters you can use. See DL4J doc.

Building a network - the XOR example

This basic example show how to build a neural network and to test it. You can test it directly by launching a REPL and typing the commands under ns alias "xor"

;; See https://github.com/deeplearning4j/dl4j-examples/blob/master/dl4j-examples/src/main/java/org/deeplearning4j/examples/feedforward/xor/XorExample.java

;; Define the XOR problem
;; vectorized training dataset
;; Here we provide the exhaustive binary definition
;; of XOR (exclusive OR)
;; First line represents the input (features), let's say X and Y
;; Last line represents the labels (~ results), here the [0 1] = false and [1 0] = true
;; (the problem and thus the network being symetrical, this could be reverted)
;; Let's say here that the first part of the label is the "percentage" of falseness
(def xor-training-data
  (dataset/->dataset
    [[0 0] [0 1] [1 0] [1 1]]
    [[0 1] [1 0] [1 0] [0 1]]))

;; You could have done that using DataVec which would have
;; vectorized a standard input like {x 0 , y 1 , xor 1}
;; from json/yaml or csv
(def xor-training-data
  (ddataset/read-datasets :jackson
    {:type-hint      :regular
     :dataset-type   :classification
     :use-nd4j?      false
     :schema         [:x :y :xor]
     :batch-size     4
     :label-idx      2
     :max-labels     2}
     "data/xor.json"))

;; Now let's define a simple neural network
;; with a single hidden layer of 4 units
(def xor-network-configuration
  {:mini-batch false
   :seed 123
   :weights-updater {:type :sgd
                     :learning-rate 0.1}
   :bias-init 0
   :layers [{:type                   :dense
             :n-in                   2
             :n-out                  4
             :activation-fn          :sigmoid
             :weight-distribution    {:type   :uniform
                                      :lower  0
                                      :upper  1}}
            {:type                   :output
             :n-in                   4
             :n-out                  2
             :activation-fn          :softmax
             :loss-fn                :negative-log-likehood
             :weight-distribution    {:type   :uniform
                                      :lower  0
                                      :upper  1}}]
   :training-listener {:type           :score-iteration
                       :nb-iterations  100}})

;; First, we can try to output a XOR from the untrained network...
(show!)

;; ...without surprise, unless you were lucky enough to initialize good weights,
;; the result is pretty bad
;; Try another time after a training (of 1000 iterations)
(train!)
(show!)

;; This should be better
;; Repeat it, and try it
(train! 100000)
(show!)

;; You can look the result on "real-life" looking data
(xor/show! [[0.15 0.95] [0.25 0.95] [0.35 0.95] [0.45 0.95]])

Other examples

You can find other examples under the corresponding folder. These examples are direct adaptations of DL4J official examples. The current examples are :

• MNIST anomaly detection : builds a network to decompose then recompose digits from samples, then plots the best and worst ones using Swing /!\ BEWARE /!\ : launching it for the first time will download MNIST data which is dumped under C://users/{you}/.deeplearning4j/data/MNIST under Windows and probably in the translated folder under Linux. Total size is ~50Mo

License

Copyright © 2018 Jérémy Le Corguillé

Apache License 2.0