Question about how to train a neural network using Flux.jl

Specific Domains Machine Learning
,
1

I am trying to build a neural network that has about 200 Boolean features as inputs and I would like to train it against a regression output (so the y is continuous). I tried to setup the network like so

using Flux
model = Flux.Chain(
  Dense(14*16+4, 64, relu),
    Dense(64, 16, relu),
    Dense(16, 1, relu));

so I got 3 dense layers all activated by relu and now to test my understanding I just ran the x through the model

x = rand(Bool, 14*16+4)
model(x)

so far so good but the below gives error

y = 100
loss(x,y) = (model(x) - y)^2

Flux.train!(loss, [(x,y)])
# Flux.train!(loss, [(x,y)], ADAM(...)) # this didn't work either

and the error is

MethodError: no method matching train!(::typeof(loss), ::Array{Tuple{Array{Bool,2},Array{Float64,1}},1})
Closest candidates are:
train!(::Any, ::Any, !Matched::Any; cb) at /home/jrun/.julia/packages/Flux/UHjNa/src/optimise/train.jl:56

Stacktrace:
[1] top-level scope at In[106]:1

This is after having browsed through the model zoo and I can’t seem to figure out how to do this simple neural network. I am doing a simple prototype of reinforcement learning, so being able to update the model is also important so would appreciate pointers to Flux.train! and Flux.update! and how to make them work

2

loss(x,y) needs to return a scalar:

using Flux
model = Flux.Chain(
    Dense(14*16+4, 64, relu),
    Dense(64, 16, relu),
    Dense(16, 1, relu));
x = rand(Bool, 14*16+4)
y = 100
loss(x,y) = sum((model(x) .- y).^2)
opt = ADAM(params(model))
Flux.@epochs 100 Flux.train!(loss, [(x,y)], opt)
2 Likes