TastyPancakes/Flux.jl
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

cross entropy loss, loss checks

Browse Source
This commit is contained in:
Mike J Innes 2016-10-30 14:12:03 +00:00
parent 3b70ea6a42
commit b443425c6d
1 changed files with 9 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
src/backend/tensorflow/recurrent.jl
View File

@ -46,20 +46,25 @@ end
(m::SeqModel)(x::Seq) = first(m(batchone(x))) (m::SeqModel)(x::Seq) = first(m(batchone(x)))
function Flux.train!(m::SeqModel, train; epoch = 1, η = 0.1, function Flux.train!(m::SeqModel, Xs, Ys; epoch = 1, η = 0.1,
loss = (y, y′) -> reduce_sum((y - y′).^2)/2, loss = (y, ŷ) -> -reduce_sum(y .* log2(ŷ)),
opt = TensorFlow.train.GradientDescentOptimizer(η)) opt = TensorFlow.train.GradientDescentOptimizer(η))
batchlen, seqlen = length(first(Xs)), length(first(Xs)[1])
state = batchone.(m.m.model.state) state = batchone.(m.m.model.state)
Y = placeholder(Float32) Y = placeholder(Float32)
Loss = loss(m.m.output[end], Y) Loss = loss(Y, m.m.output[end])/batchlen/seqlen
minimize_op = TensorFlow.train.minimize(opt, Loss) minimize_op = TensorFlow.train.minimize(opt, Loss)
for e in 1:epoch for e in 1:epoch
info("Epoch $e\n") info("Epoch $e\n")
@progress for (x, y) in train @progress for (i, (x, y)) in enumerate(zip(Xs,Ys))
out = run(m.m.session, vcat(m.m.output..., Loss, minimize_op), out = run(m.m.session, vcat(m.m.output..., Loss, minimize_op),
merge(Dict(m.m.inputs[end]=>batchone(x), Y=>batchone(y)), merge(Dict(m.m.inputs[end]=>batchone(x), Y=>batchone(y)),
Dict(zip(m.m.inputs[1:end-1], state)))) Dict(zip(m.m.inputs[1:end-1], state))))
state = out[1:length(state)] state = out[1:length(state)]
loss = out[end-1]
isnan(loss) && error("Loss is NaN")
isinf(loss) && error("Loss is Inf")
(i-1) % 10 == 0 && @show loss
end end
end end
end end