Problem with Training and trying to plot loss function in Flux

New to Julia
1

Been testing some basic code and trying to plot its loss function but get some errors. Need your help to move on. Here is the code:

using Flux
actual(x) = 2x + 3
x_train, x_test = hcat(0:5…), hcat(6:10…)
y_train, y_test = actual.(x_train), actual.(x_test)

model = Dense(1, 1)
predict = model
loss(x, y) = Flux.Losses.mse(predict(x), y)
opt = Descent()
data = [(x_train, y_train)]
parameters = Flux.params(predict)
Flux.train!(loss, parameters, data, opt)’

I get error on this last block of code

epochs = Int64
loss_on_train = Float32
loss_on_test = Float32

for epoch in 1:2000
Flux.train!(loss, parameters,data, opt)
# we record our training loss
push!(epochs, epoch)
push!(loss_on_test, loss(x_test, y_test).data)
push!(loss_on_train, loss(x_train, y_train).data)
end

using Plots
plot(epochs, loss_on_train, lab=“Training”, c=:black, lw=2);
plot!(epochs, loss_on_test, lab=“Testing”, c=:teal, ls=:dot);
yaxis!(“Loss”, :log);
xaxis!(“Training epoch”)

Here is the error:
type Float32 has no field data

2

Hi @ak47! I’m willing to bet the error comes from these two lines: you don’t need to take the .data field of the loss, since it is already a number (Float32 in this case).

3

Thank you Gdalle. The error goes away but the plot still does not work…

4

Push records only the final loss and that seems the reason why it doesn’t plot. How do I modify this ?

5

I’m not sure what’s wrong, it works with the following code for me (which is the same as yours):

using Flux, Plots

actual(x) = 2x + 3
x_train, x_test = hcat(0:5...), hcat(6:10...)
y_train, y_test = actual.(x_train), actual.(x_test)

model = Dense(1, 1)
predict = model
loss(x, y) = Flux.Losses.mse(predict(x), y)
opt = Descent()
data = [(x_train, y_train)]
parameters = Flux.params(predict)
Flux.train!(loss, parameters, data, opt)

epochs = Int64[]
loss_on_train = Float32[]
loss_on_test = Float32[]

for epoch in 1:2000
    Flux.train!(loss, parameters,data, opt)
    # we record our training loss
    push!(epochs, epoch)
    push!(loss_on_test, loss(x_test, y_test))
    push!(loss_on_train, loss(x_train, y_train))
end

plot(epochs, loss_on_train, lab="Training", c=:black, lw=2);
plot!(epochs, loss_on_test, lab="Testing", c=:teal, ls=:dot);
yaxis!("Loss", :log);
xaxis!("Training epoch")

To make life easier for other people, it’s best to use backticks on Discourse when displaying code (like I just did). Maybe the problem you had came from a copy-pasting issue?

1 Like
6

Thanks for your time.