Merge branch 'master' into tf-train

This commit is contained in:
Mike J Innes 2017-08-17 23:45:55 +01:00
commit cd9521a762
9 changed files with 141 additions and 5 deletions

View File

@ -23,7 +23,7 @@ include("core.jl")
import .FluxCore: back!, update!, graph
include("utils.jl")
include("ops.jl")
include("params.jl")
include("compiler/code.jl")

View File

@ -31,6 +31,15 @@ graph(::typeof(svd), x) = svd(x)
graph(::typeof(size), x, dim) = TensorFlow.size(x,convert(Tensor{Int32}, dim))
graph(::typeof(size), x) = TensorFlow.size(x)
graph(::typeof(chol), args...) = TensorFlow.transpose(TensorFlow.cholesky(args...))
graph(::typeof(reshape), x, dims) = TensorFlow.reshape(x,convert(Tensor{Int32},dims))
graph(::typeof(Flux.tile), args...) = TensorFlow.tile(args...)
graph(::typeof(fill), x, dims) = Ops.fill(convert(Tensor{Int32}, dims), Tensor(x))
graph(::typeof(Flux.cast), args...) = TensorFlow.cast(args...)
graph(::typeof(solve), A, b) = TensorFlow.matrix_solve(A, b)
graph(::typeof(triangular_solve), A, b) = TensorFlow.matrix_triangular_solve(A, b; lower=false)
graph(::typeof(randu), x) = Ops.random_uniform(convert(Tensor{Int32},x);dtype=Float32)
graph(::typeof(randn), x) = TensorFlow.random_normal(convert(Tensor{Int32},x);dtype=Float32)
graph(::typeof(Flux.expand_dims), x, dim) = TensorFlow.expand_dims(x,convert(Tensor{Int32},dim))
for op in (*, .*, .+, .^, log, exp, ceil, floor, sqrt, abs, cos,
sin, tan, atan, asin, acos, tanh, lgamma, erf, erfc, real, imag, conj,

View File

@ -1,4 +1,4 @@
using Flux: mapt, collectt, shapecheckt
using Flux: Param, mapt, collectt, shapecheckt
struct Exec
session ::Session

View File

@ -1,7 +1,7 @@
module TF
using ..Flux, DataFlow, TensorFlow, Juno
import Flux: accuracy, convertel, Param
import Flux: accuracy, convertel
export tf

18
src/ops.jl Normal file
View File

@ -0,0 +1,18 @@
export reshape, tile, fill, cast, solve, triangular_solve, randu, randn,
expand_dims
import Base: reshape, fill, randn
reshape(x::AbstractArray, dims::AbstractArray) = reshape(x,tuple(dims...))
tile(x::AbstractArray, mult::AbstractArray) = repeat(x,outer=tuple(mult...))
fill{T}(x::T, dims::AbstractArray) = fill(x,tuple(dims...))
cast{T}(x::AbstractArray, ::Type{T}) = convert(Array{T},x)
solve(A::AbstractArray, b::AbstractArray) = A\b
triangular_solve(A::AbstractArray, b::AbstractArray) = A\b
randu(x::AbstractArray) = rand(tuple(x...))
randn(x::AbstractArray) = randn(tuple(x...))
function expand_dims(x,dim)
s = [size(x)...]
reshape(x,tuple(vcat(s[1:dim-1],1,s[dim:end])...))
end

View File

@ -25,18 +25,59 @@ macro cb(ex, t, f)
end)
end
"""
Returns a function that when invoked, will only be triggered at most once
during `timeout` seconds. Normally, the throttled function will run
as much as it can, without ever going more than once per `wait` duration;
but if you'd like to disable the execution on the leading edge, pass
`leading=false`. To enable execution on the trailing edge, ditto.
"""
function throttle(f, timeout; leading=true, trailing=false)
cooldown = true
later = nothing
function throttled(args...; kwargs...)
yield()
if cooldown
if leading
f(args...; kwargs...)
else
later = () -> f(args...; kwargs...)
end
cooldown = false
@schedule try
while (sleep(timeout); later != nothing)
later()
later = nothing
end
finally
cooldown = true
end
elseif trailing
later = () -> f(args...; kwargs...)
end
nothing
end
end
function train!(m, train; cb = [],
epoch = 1, η = 0.1, loss = mse)
callback = throttle(()->foreach(f -> f(), cb), 5)
@progress for e in 1:epoch
info("Epoch $e")
@cb for (x, y) in train
for (x, y) in train
x, y = mapt(tobatch, (x, y))
= m(x)
any(isnan, ) && error("NaN")
Δ = back!(loss, 1, , y)
back!(m, Δ, x)
update!(m, η)
end 5 foreach(f -> f(), cb)
callback()
end
end
return m
end

View File

@ -47,6 +47,24 @@ end
A = randn(6,5)
A = A'*A
@test tf(@net x -> chol(x))(A) chol(A)
A = randn(Float32,(6,3))
@test transpose(tf(@net (x,y) -> reshape(x,y))(transpose(A),[2,9])) reshape(A,(9,2)) # Note: TF is row major and julia is not
A = randn(Float32,(4,3,1))
@test tf(@net (x,y) -> Flux.tile(x,y))(A,[1,1,3]) repeat(A,outer=(1,1,3))
@test tf(@net (x,y) -> fill(x,y))(3.2,[3,2]) convert(Array{Float32},3.2*ones(3,2))
@test typeof(tf(@net x -> Flux.cast(x,Int32))(A)) == Array{Int32,3}
A = randn(Float32,(5,5))
b = randn(Float32,(5,1))
@test tf(@net (x,y) -> solve(x,y))(A,b) A\b
_,A,_ = lu(A)
@test tf(@net (x,y) -> triangular_solve(x,y))(A,b) A\b
@test size(tf(@net x -> randu(x))([2,3])) == (2,3)
@test size(tf(@net x -> randn(x))([2,3])) == (2,3)
m = tf(@net (x,y) -> Flux.expand_dims(x,y))
A = randn(Float32,(3,2))
@test m(A,1) Flux.expand_dims(A,1)
@test m(A,2) Flux.expand_dims(A,2)
@test m(A,3) Flux.expand_dims(A,3)
end
end

View File

@ -18,6 +18,7 @@ include("backend/common.jl")
include("basic.jl")
include("recurrent.jl")
include("optimizer.jl")
include("throttle.jl")
@tfonly include("backend/tensorflow.jl")
@mxonly include("backend/mxnet.jl")

49
test/throttle.jl Normal file
View File

@ -0,0 +1,49 @@
using Flux.throttle
@testset "throttle" begin
@testset "default behaviour" begin
a = []
f = throttle(()->push!(a, now()), 1, leading=true, trailing=false)
f()
f()
f()
sleep(1.01)
@test length(a) == 1
end
@testset "leading behaviour" begin
a = []
f = throttle(()->push!(a, now()), 1, leading=true, trailing=false)
f()
@test length(a) == 1
f()
@test length(a) == 1
sleep(1.01)
f()
@test length(a) == 2
end
@testset "trailing behaviour" begin
a = []
f = throttle(()->push!(a, now()), 1, leading=false, trailing=true)
f()
@test length(a) == 0
f()
@test length(a) == 0
sleep(1.01)
@test length(a) == 1
end
@testset "arguments" begin
a = []
f = throttle((x)->push!(a, x), 1, leading=true, trailing=true)
f(1)
@test a == [1]
f(2)
@test a == [1]
f(3)
@test a == [1]
sleep(1.01)
@test a == [1, 3]
end
end