Merge pull request #93 from SciML/esnfitted

Added ESNfitted function

Project.toml

@@ -1,7 +1,7 @@
 name = "ReservoirComputing"
 uuid = "7c2d2b1e-3dd4-11ea-355a-8f6a8116e294"
 authors = ["Francesco Martinuzzi"]
-version = "0.6.2"
+version = "0.6.3"
 
 [deps]
 Distances = "b4f34e82-e78d-54a5-968a-f98e89d6e8f7"

src/ReservoirComputing.jl

@@ -28,7 +28,7 @@ include("esn_reservoirs.jl")
 export init_reservoir_givendeg, init_reservoir_givensp, pseudoSVD, DLR, DLRB, SCR, CRJ
 
 include("echostatenetwork.jl")
-export ESN, ESNpredict, ESNpredict_h_steps
+export ESN, ESNpredict, ESNpredict_h_steps, ESNfitted
 
 include("dafesn.jl")
 export dafESN, dafESNpredict, dafESNpredict_h_steps

src/echostatenetwork.jl

@@ -225,3 +225,63 @@ function ESNpredict_h_steps(esn::AbstractLeakyESN,
     end
     return output
 end
+
+"""
+    ESNfitted(esn::AbstractLeakyESN, W_out::Matrix; autonomous=false)
+
+Return the prediction for the training data using the trained output layer. The autonomous trigger can be used to have have it return an autonomous prediction starting from the first point if true, or a point by point prediction if false.
+"""
+
+function ESNfitted(esn::AbstractLeakyESN, W_out::Matrix; autonomous=false)
+    train_len = size(esn.train_data, 2)
+    output = zeros(Float64, esn.in_size, train_len)
+    x = zeros(size(esn.states, 1))
+
+    if autonomous
+        out = esn.train_data[:,1]
+        return _fitted!(output, esn, x, train_len, W_out, out)
+    else
+        return _fitted!(output, esn, x, train_len, W_out, esn.train_data)
+    end
+end
+
+function _fitted!(output, esn, state, train_len, W_out, vector::Vector)
+    if esn.extended_states == false
+        for i=1:train_len
+            state = leaky_fixed_rnn(esn.activation, esn.alpha, esn.W, esn.W_in, state, vector)
+            x_new = nla(esn.nla_type, state)
+            vector = (W_out*x_new)
+            output[:, i] = vector
+        end
+    elseif esn.extended_states == true
+        for i=1:train_len
+            state = vcat(leaky_fixed_rnn(esn.activation, esn.alpha, esn.W, esn.W_in, state[1:esn.res_size], vector), vector)
+            x_new = nla(esn.nla_type, state)
+            vector = (W_out*x_new)
+            output[:, i] = vector
+        end
+    end
+    return output
+end
+
+function _fitted!(output, esn, state, train_len, W_out, vector::Matrix)
+    if esn.extended_states == false
+        for i=1:train_len
+            state = leaky_fixed_rnn(esn.activation, esn.alpha, esn.W, esn.W_in, state, vector[:,i])
+            x_new = nla(esn.nla_type, state)
+            out = (W_out*x_new)
+            output[:, i] = out
+        end
+    elseif esn.extended_states == true
+        for i=1:train_len
+            state = vcat(leaky_fixed_rnn(esn.activation, esn.alpha, esn.W, esn.W_in, state[1:esn.res_size], vector[:,i]), vector[:,i])
+            x_new = nla(esn.nla_type, state)
+            out = (W_out*x_new)
+            output[:, i] = out
+        end
+    end
+    return output
+end
+
+
+

test/extras/test_extended_states.jl

@@ -41,6 +41,13 @@ output = ESNpredict(esn, predict_len, W_out)
 output = ESNpredict_h_steps(esn, predict_len, h_steps, test, W_out)
 @test size(output) == (out_size, predict_len)
 
+#test esnfitted
+fit1 = ESNfitted(esn, W_out; autonomous=false)
+@test size(fit1) == size(train)
+
+fit2 = ESNfitted(esn, W_out; autonomous=true)
+@test size(fit1) == size(train)
+
 #test esgp
 mean = MeanZero()
 kernel = Lin(1.0)

test/runtests.jl

@@ -17,3 +17,4 @@ using SafeTestsets
 @time @safetestset "reca gol predict" begin include("training/test_recagol.jl") end
 @time @safetestset "RMM constructors" begin include("constructors/test_rmm_constructors.jl") end
 @time @safetestset "GRUESN constructors" begin include("constructors/test_gruesn_constructors.jl") end
+@time @safetestset "ESN fitted" begin include("training/test_esnfitted.jl") end

test/training/test_esnfitted.jl

@@ -0,0 +1,41 @@
+using ReservoirComputing
+using MLJLinearModels
+#model parameters
+const approx_res_size = 30
+const radius = 1.2
+const activation = tanh
+const degree = 6
+const sigma = 0.1
+const beta = 0.0
+const alpha = 1.0
+const nla_type = NLADefault()
+const in_size = 3
+const out_size = 3
+const extended_states = false
+const delta = 0.5
+
+
+const train_len = 50
+const predict_len = 12
+data = ones(Float64, in_size, 100)
+train = data[:, 1:1+train_len-1]
+test = data[:, train_len:train_len+predict_len-1]
+
+#constructor 1
+esn = ESN(approx_res_size,
+    train,
+    degree,
+    radius,
+    activation = activation,
+    sigma = sigma,
+    alpha = alpha,
+    nla_type = nla_type,
+    extended_states = extended_states) 
+
+W_out = ESNtrain(esn, beta)
+
+fit1 = ESNfitted(esn, W_out; autonomous=false)
+@test size(fit1) == size(train)
+
+fit2 = ESNfitted(esn, W_out; autonomous=true)
+@test size(fit1) == size(train)