A neural networks library for Clojure!
[org.clojars.mrdimosthenis/clj-synapses "1.0.3"]
(require '[clj-synapses.net :as net])
(def rand-network
(net/->net
[2 3 1]))
- Input layer: the first layer of the network has 2 nodes.
- Hidden layer: the second layer has 3 neurons.
- Output layer: the third layer has 1 neuron.
(net/->json
rand-network)
;;=> "[[{\"activationF\" : \"sigmoid\", \"weights\" : [-0.5,0.1,0.8]},
;; {\"activationF\" : \"sigmoid\", \"weights\" : [0.7,0.6,-0.1]},
;; {\"activationF\" : \"sigmoid\", \"weights\" : [-0.8,-0.1,-0.7]}],
;; [{\"activationF\" : \"sigmoid\", \"weights\" : [0.5,-0.3,-0.4,-0.5]}]]"
(def network
(net/json->
"[[{\"activationF\" : \"sigmoid\", \"weights\" : [-0.5,0.1,0.8]},
{\"activationF\" : \"sigmoid\", \"weights\" : [0.7,0.6,-0.1]},
{\"activationF\" : \"sigmoid\", \"weights\" : [-0.8,-0.1,-0.7]}],
[{\"activationF\" : \"sigmoid\", \"weights\" : [0.5,-0.3,-0.4,-0.5]}]]"))
(net/predict
network
[0.2 0.6])
;;=> [0.49131100324012494]
(net/fit
network
0.1
[0.2 0.6]
[0.9])
The fit
function returns a new neural network with the weights adjusted to a single observation.
In practice, for a neural network to be fully trained, it should be fitted with multiple observations, usually by reducing over a collection.
(reduce
(fn [acc [xs ys]]
(net/fit acc 0.1 xs ys))
network
[[[0.2 0.6] [0.9]]
[[0.1 0.8] [0.2]]
[[0.5 0.4] [0.6]]])
Every function is efficient because its implementation is based on lazy list and all information is obtained at a single pass.
For a neural network that has huge layers, the performance can be further improved by using the parallel counterparts
of net/predict
and net/fit
(net/par-predict
and net/par-fit
).
(net/->net
[2 3 1]
1000)
We can provide a seed
to create a non-random neural network. This way, we can use it for testing.
(require '[clj-synapses.fun :as fun])
(defn activation-f
[layer-index]
(condp = layer-index
0 fun/sigmoid
1 fun/identity
2 fun/leaky-re-lu
3 fun/tanh))
(defn weight-init-f
[layer-index]
(* (inc layer-index)
(- 1 (* 2.0 (rand)))))
(def custom-network
(net/->net
[4 6 8 5 3]
activation-f
weight-init-f))
- The
activation-f
function accepts the index of a layer and returns an activation function for its neurons. - The
weight-initf
function accepts the index of a layer and returns a weight for the synapses of its neurons.
If we don't provide these functions, the activation function of all neurons is sigmoid, and the weight distribution of the synapses is normal between -1.0 and 1.0.
(net/->svg
custom-network)
With its svg drawing, we can see what a neural network looks like. The color of each neuron depends on its activation function while the transparency of the synapses depends on their weight.
(require '[clj-synapses.stats :as stats])
(def exp-and-pred-vals
[[[0.0 0.0 1.0] [0.0 0.1 0.9]]
[[0.0 1.0 0.0] [0.8 0.2 0.0]]
[[1.0 0.0 0.0] [0.7 0.1 0.2]]
[[1.0 0.0 0.0] [0.3 0.3 0.4]]
[[0.0 0.0 1.0] [0.2 0.2 0.6]]])
- Root-mean-square error
(stats/rmse
exp-and-pred-vals)
;;=> 0.6957010852370435
- Classification accuracy score
(stats/score
exp-and-pred-vals)
;;=> 0.6
(require '[clj-synapses.codec :as codec])
- One hot encoding is a process that turns discrete attributes into a list of 0.0 and 1.0.
- Minmax normalization scales continuous attributes into values between 0.0 and 1.0.
(def setosa
{"petal_length" "1.5"
"petal_width" "0.1"
"sepal_length" "4.9"
"sepal_width" "3.1"
"species" "setosa"})
(def versicolor
{"petal_length" "3.8"
"petal_width" "1.1"
"sepal_length" "5.5"
"sepal_width" "2.4"
"species" "versicolor"})
(def virginica
{"petal_length" "6.0"
"petal_width" "2.2"
"sepal_length" "5.0"
"sepal_width" "1.5"
"species" "virginica"})
(def dataset
[setosa
versicolor
virginica])
You can use a codec
to encode and decode a data point.
(def preprocessor
(codec/->codec
[["petal_length" false]
["petal_width" false]
["sepal_length" false]
["sepal_width" false]
["species" true]]
dataset))
)
- The first parameter is a vector of pairs that define the name and the type (discrete or not) of each attribute.
- The second parameter is a collection that contains the data points.
(codec/->json
preprocessor)
;;=> "[{\"Case\" : \"SerializableContinuous\",
;; \"Fields\" : [{\"key\" : \"petal_length\",\"min\" : 1.5,\"max\" : 6.0}]},
;; {\"Case\" : \"SerializableContinuous\",
;; \"Fields\" : [{\"key\" : \"petal_width\",\"min\" : 0.1,\"max\" : 2.2}]},
;; {\"Case\" : \"SerializableContinuous\",
;; \"Fields\" : [{\"key\" : \"sepal_length\",\"min\" : 4.9,\"max\" : 5.5}]},
;; {\"Case\" : \"SerializableContinuous\",
;; \"Fields\" : [{\"key\" : \"sepal_width\",\"min\" : 1.5,\"max\" : 3.1}]},
;; {\"Case\" : \"SerializableDiscrete\",
;; \"Fields\" : [{\"key\" : \"species\",\"values\" : [\"virginica\",\"versicolor\",\"setosa\"]}]}]"
(codec/json->
"[{\"Case\" : \"SerializableContinuous\",
\"Fields\" : [{\"key\" : \"petal_length\",\"min\" : 1.5,\"max\" : 6.0}]},
{\"Case\" : \"SerializableContinuous\",
\"Fields\" : [{\"key\" : \"petal_width\",\"min\" : 0.1,\"max\" : 2.2}]},
{\"Case\" : \"SerializableContinuous\",
\"Fields\" : [{\"key\" : \"sepal_length\",\"min\" : 4.9,\"max\" : 5.5}]},
{\"Case\" : \"SerializableContinuous\",
\"Fields\" : [{\"key\" : \"sepal_width\",\"min\" : 1.5,\"max\" : 3.1}]},
{\"Case\" : \"SerializableDiscrete\",
\"Fields\" : [{\"key\" : \"species\",\"values\" : [\"virginica\",\"versicolor\",\"setosa\"]}]}]")
(def encoded-setosa
(codec/encode
preprocessor
setosa))
;; [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 1.0]
(codec/decode
preprocessor
encoded-setosa)
;;=> {"species" "setosa"
;; "sepal_width" "3.1"
;; "petal_width" "0.1",
;; "petal_length" "1.5"
;; "sepal_length" "4.9"}