Skip to content

Commit

Permalink
refactor ts_invariant_statistical_loss_multivariate
Browse files Browse the repository at this point in the history
  • Loading branch information
@josemanuel22
josemanuel22 committed Feb 1, 2024
1 parent cd28c38 commit 565cdd4
Show file tree
Hide file tree
Showing 2 changed files with 102 additions and 122 deletions.
143 changes: 78 additions & 65 deletions examples/time_series_predictions/benchmark_ts.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -77,7 +77,6 @@ Example:
@test_experiments "testing electricity-f" begin
csv_file_path = "examples/time_series_predictions/data/LD2011_2014.txt"


cols = [
"MT_005",
"MT_006",
Expand Down Expand Up @@ -117,9 +116,7 @@ Example:

#nn_model = Chain(RNN(5 => 32, relu), RNN(32 => 32, relu), Dense(32 => 1, identity))
rec = Chain(LSTM(1 => 16), LayerNorm(16))
gen = Chain(Dense(17, 32, elu),
Dropout(0.05),
Dense(32, 1, identity))
gen = Chain(Dense(17, 32, elu), Dropout(0.05), Dense(32, 1, identity))

start = 35040
num_training_data = 1000
Expand Down Expand Up @@ -158,7 +155,6 @@ Example:
v_std = std(aggregated_data_xtest)
aggregated_data_xtest = (aggregated_data_xtest .- v_mean) ./ v_std


loaderXtrain = Flux.DataLoader(
aggregated_data_xtrain;
batchsize=round(Int, num_training_data),
Expand Down Expand Up @@ -192,7 +188,6 @@ Example:
append!(ql5, ql5_)
end


mse = 0.0
mae = 0.0
count = 0
Expand Down Expand Up @@ -591,9 +586,7 @@ end
hparams = HyperParamsTS(; seed=1234, η=1e-3, epochs=2000, window_size=2000, K=10)

rec = Chain(LSTM(1 => 16), LayerNorm(16))
gen = Chain(Dense(17, 32, elu),
Dropout(0.05),
Dense(32, 1, identity))
gen = Chain(Dense(17, 32, elu), Dropout(0.05), Dense(32, 1, identity))

# start and finish of the training data
start = 35040
Expand Down Expand Up @@ -752,12 +745,12 @@ end

rec = Chain(
RNN(1 => 32, relu;),
RNN(32 => 32, relu;),
RNN(32 => 32, relu;),
RNN(32 => 64, relu;),
RNN(64 => 32, relu;),
RNN(32 => 32, relu;),
RNN(32 => 32, relu;),
)
gen = Chain(Dense(33, 64, relu;), Dense(64, 1, identity;))
gen = Chain(Dense(33, 128, relu;), Dense(128, 1, identity;))

ts = df1.FR22

Expand Down Expand Up @@ -1468,7 +1461,7 @@ end

df1 = CSV.File(csv1; delim=',', header=true, decimal='.')

hparams = HyperParamsTS(; seed=1234, η=1e-2, epochs=2000, window_size=2000, K=20)
hparams = HyperParamsTS(; seed=1234, η=1e-3, epochs=2000, window_size=2000, K=50)

rec = Chain(RNN(1 => 3, relu), LayerNorm(3), Dropout(0.1))
gen = Chain(Dense(4, 10, relu), Dropout(0.1), Dense(10, 1, identity))
Expand Down Expand Up @@ -1556,19 +1549,65 @@ end

losses = []
ql5 = []
@showprogress for _ in 1:10
mses = []
maes = []
@showprogress for _ in 1:5
loss, _ql5 = ts_invariant_statistical_loss(
rec, gen, loaderXtrain, loaderYtrain, hparams, loaderXtest; cond=0.3
)
append!(losses, loss)
append!(ql5, _ql5)

τ = 96
predictions, stds = ts_forecast(
rec, gen, collect(loaderXtrain)[1], collect(loaderXtest)[1], τ; n_average=1000
)

mse = 0.0
mae = 0.0
for ts in (1:(length(names(df)) - 1))
prediction = Vector{Float32}()
stds = Vector{Float32}()
xtrain = collect(loaderXtrain)[ts]
Flux.reset!(rec)
s = []
for j in ((length(xtrain) - 96):1:(length(xtrain) - 1))
s = rec([xtrain[j + 1]])
end

τ = 336
xtest = collect(loaderXtest)[ts]
noise_model = Normal(0.0f0, 1.0f0)
n_average = 1000
for j in (0:(τ):(length(xtest) - τ))
#s = rec(xtest[(j + 1):(j + τ)]')
#s = rec([xtest[j + 1]])
for i in 1:(τ)
xₖ = rand(noise_model, n_average)
y = hcat([gen(vcat(x, s)) for x in xₖ]...)
ȳ = mean(y)
σ = std(y)
s = rec([ȳ])
append!(prediction, ȳ)
append!(stds, σ)
end
end
ideal = collect(loaderXtest)[ts]
QLρ([x[1] for x in ideal][1:τ], prediction[1:τ]; ρ=0.5)
println(MSE([x[1] for x in ideal][1:τ], prediction[1:τ]))
println(MAE([x[1] for x in ideal][1:τ], prediction[1:τ]))
mse += MSE([x[1] for x in ideal][1:τ], prediction[1:τ])
mae += MAE([x[1] for x in ideal][1:τ], prediction[1:τ])
end
append!(mses, mse / 7.0)
append!(maes, mae / 7.0)
end

window_size = 50
ma_result = moving_average([x[1] for x in ql5], window_size)
plot(ma_result)

τ = 30
τ = 96
predictions, stds = ts_forecast(
rec, gen, collect(loaderXtrain)[1], collect(loaderXtest)[1], τ; n_average=1000
)
Expand All @@ -1585,7 +1624,7 @@ end
s = rec([xtrain[j + 1]])
end

τ = 720
τ = 336
xtest = collect(loaderXtest)[ts]
noise_model = Normal(0.0f0, 1.0f0)
n_average = 1000
Expand Down Expand Up @@ -2076,16 +2115,16 @@ end
end

@test_experiments "ETDataset" begin
csv1 = "/Users/jmfrutos/github/ETDataset/ETT-small/ETTm1.csv"
csv1 = "/Users/jmfrutos/github/ETDataset/ETT-small/ETTh2.csv"

#column_names = [:col1, :col2, :col3, :col4, :col5, :col6, :col7, :col8]

df1 = CSV.File(csv1; delim=',', header=true, decimal='.')

hparams = HyperParamsTS(; seed=1234, η=1e-4, epochs=2000, window_size=2000, K=10)
hparams = HyperParamsTS(; seed=1234, η=1e-3, epochs=2000, window_size=1000, K=40)

rec = Chain(RNN(7 => 5, relu), LayerNorm(5))
gen = Chain(Dense(6, 10, relu), Dropout(0.05), Dense(10, 7, identity))
rec = Chain(RNN(7 => 10, relu), LayerNorm(10))
gen = Chain(Dense(11, 15, relu), Dropout(0.05), Dense(15, 7, identity))

df = DataFrame(df1)
df = select(df, Not(:date))
Expand Down Expand Up @@ -2149,8 +2188,8 @@ end
v_std = std(ztrain)
ztrain = (ztrain .- v_mean) ./ v_std

mean_vals = mean(matriz, dims=1)
std_vals = std(matriz, dims=1)
mean_vals = mean(matriz; dims=1)
std_vals = std(matriz; dims=1)
matriz = (matriz .- mean_vals) ./ std_vals

dataX = [matriz[i, :] for i in 1:size(matriz, 1)]
Expand All @@ -2160,22 +2199,23 @@ end
batch_size = 4000

# Create the DataLoader
loaderXtrain = DataLoader(dataX, batchsize=batch_size, shuffle=false, partial = false)
loaderYtrain = DataLoader(dataY, batchsize=batch_size, shuffle=false, partial = false)
loaderXtrain = DataLoader(dataX; batchsize=batch_size, shuffle=false, partial=false)
loaderYtrain = DataLoader(dataY; batchsize=batch_size, shuffle=false, partial=false)

maes = []
mses = []
losses = []
@showprogress for i in 1:1000
loss = ts_invariant_statistical_loss_multivariate(rec, gen, loaderXtrain, loaderYtrain, hparams)
loss = ts_invariant_statistical_loss_multivariate(
rec, gen, loaderXtrain, loaderYtrain, hparams
)
append!(losses, loss)


mse = 0.0
mae = 0.0
count = 0
τ = 96
for ts in (1: length(collect(loaderXtrain)[1][1]))
τ = 336
for ts in (1:length(collect(loaderXtrain)[1][1]))
#τ = 96
s = 0
Flux.reset!(rec)
Expand All @@ -2187,7 +2227,7 @@ end
#ŷ = mean([gen(vcat(x, s)) for x in xₖ])
#push!(prediction, ŷ)
end
for i in (96:(96+τ))
for i in (96:(96 + τ))
xₖ = rand(Normal(0.0f0, 1.0f0), 100)
y = mean([gen(vcat(x, s)) for x in xₖ])
#y = mean([gen(vcat(x, s)) for x in xₖ])
Expand All @@ -2201,8 +2241,8 @@ end
#l = moving_average(primeraColumnaY, 10)
#println(MSE(primeraColumna[96:96+τ], primeraColumnaY))
#println(MAE(primeraColumna[96:96+τ], primeraColumnaY))
mse += MSE(primeraColumna[96:96+τ], primeraColumnaY)
mae += MAE(primeraColumna[96:96+τ], primeraColumnaY)
mse += MSE(primeraColumna[96:(96 + τ)], primeraColumnaY)
mae += MAE(primeraColumna[96:(96 + τ)], primeraColumnaY)
end
append!(mses, mse / count)
append!(maes, mae / count)
Expand Down Expand Up @@ -2260,8 +2300,8 @@ end
mae = 0.0
count = 0
τ = 720
s=0
for ts in (1: length(collect(loaderXtrain)[1][1]))
s = 0
for ts in (1:length(collect(loaderXtrain)[1][1]))
#τ = 96
Flux.reset!(rec)
xtrain = collect(loaderXtrain)[1]
Expand All @@ -2272,7 +2312,7 @@ end
ŷ = mean([gen(vcat(x, s)) for x in xₖ])
push!(prediction, ŷ)
end
for i in (96:(96+τ))
for i in (96:(96 + τ))
xₖ = rand(Normal(0.0f0, 1.0f0), 1000)
ŷ = mean([gen(vcat(x, s)) for x in xₖ])
#println(ŷ)
Expand All @@ -2283,10 +2323,10 @@ end
primeraColumna = [fila[ts] for fila in collect(loaderXtrain)[1]]
count += 1
l = moving_average(primeraColumnaY, 10)
println(MSE(primeraColumna[96:96+τ], primeraColumnaY))
println(MAE(primeraColumna[96:96+τ], primeraColumnaY))
mse += MSE(primeraColumna[96:96+τ], primeraColumnaY)
mae += MAE(primeraColumna[96:96+τ], primeraColumnaY)
println(MSE(primeraColumna[96:(96 + τ)], primeraColumnaY))
println(MAE(primeraColumna[96:(96 + τ)], primeraColumnaY))
mse += MSE(primeraColumna[96:(96 + τ)], primeraColumnaY)
mae += MAE(primeraColumna[96:(96 + τ)], primeraColumnaY)
end
mse / count
mae / count
Expand All @@ -2296,7 +2336,6 @@ end

MSE(primeraColumna[1:145], l)


mse = 0.0
mae = 0.0
for ts in (1:(length(names(df))))
Expand Down Expand Up @@ -2333,30 +2372,4 @@ end
mse += MSE([x[1] for x in ideal][1:τ], prediction[1:τ])
mae += MAE([x[1] for x in ideal][1:τ], prediction[1:τ])
end

end

function ts_invariant_statistical_loss_multivariate(rec, gen, Xₜ, Xₜ₊₁, hparams)
losses = []
optim_rec = Flux.setup(Flux.Adam(hparams.η), rec)
optim_gen = Flux.setup(Flux.Adam(hparams.η), gen)
Flux.reset!(rec)
@showprogress for (batch_Xₜ, batch_Xₜ₊₁) in zip(Xₜ, Xₜ₊₁)
loss, grads = Flux.withgradient(rec, gen) do rec, gen
aₖ = zeros(hparams.K + 1)
for j in (1:(hparams.window_size):length(batch_Xₜ))
s = rec(batch_Xₜ[j])
xₖ = rand(hparams.noise_model, hparams.K)
yₖ = hcat([gen(vcat(x, s)) for x in xₖ]...)
for i in 1:length(batch_Xₜ₊₁[j])
aₖ += generate_aₖ(yₖ[i:i, :], batch_Xₜ₊₁[j][i])
end
end
scalar_diff(aₖ ./ sum(aₖ))
end
Flux.update!(optim_rec, rec, grads[1])
Flux.update!(optim_gen, gen, grads[2])
push!(losses, loss)
end
return losses
end
Loading

0 comments on commit 565cdd4

Please sign in to comment.