2D Conv transpose support #311

tejank10 · 2018-07-09T09:19:28Z

This PR adds support for 2D Conv Transpose for Flux, along with #54 in NNlib.jl

tejank10 · 2018-07-13T17:18:37Z

gradtests are failing for ∇conv_data. Gradtest wrt data is passing. I found that the culprit is the gradient wrt the weight which is differing drastically.
EDIT: results of gradtests
gradtest(∇conv_data, rand(10, 3, 2), randn(2, 2, 3))

sum.(ngradient(f, xs...)) = (-15.724534273147583, 45.65564203262329)
sum.(data.(gradient(f, xs...))) = (-15.72453422801545, -41.27033625893009)

gradtest(∇conv_data, rand(10, 10, 3, 2), randn(2, 2, 2, 3))

sum.(ngradient(f, xs...)) = (-728.0231256484985, -326.6875114440918)
sum.(data.(gradient(f, xs...))) = (-728.0231184076845, 926.8081786572867)

gradtest(∇conv_data, rand(10, 10, 10, 3, 2), randn(2, 2, 2, 2, 3))

sum.(ngradient(f, xs...)) = (2724.2830810546875, 3647.9795532226562)
sum.(data.(gradient(f, xs...))) = (2724.2841834190576, -3632.449182033016)

tejank10 · 2018-09-08T20:11:05Z

NNlib counterpart of this PR needs to be merged for the checks to pass

Aniket1998 · 2018-10-06T19:34:37Z

What's the status of this PR? Working on an InfoGAN model for the Flux model zoo, unable to write a version that would work on SVHN without conv transpose

yuehhua · 2018-10-19T05:11:39Z

It's there anything I can help? I'm waiting for this.

…transpose

MikeInnes · 2018-11-28T00:10:15Z

@tejank10 this PR looks generally good to me, mind just updating it?

vchuravy · 2018-12-02T21:46:58Z

test/tracker.jl

@@ -1,7 +1,11 @@
 using Flux
 using Flux.Tracker, Test, NNlib
+<<<<<<< HEAD


Merge artificat

vchuravy · 2018-12-02T21:51:27Z

src/layers/conv.jl

+              stride = 1, pad = 0, dilation = 1) where {T,N} =
+  ConvTranspose(σ, w, b, expand.(sub2(Val(N)), (stride, pad, dilation))...)
+
+ConvTranspose(k::NTuple{N,Integer}, ch::Pair{<:Integer,<:Integer}, σ = identity; init = initn,


julia> ConvTranspose((2, 2), 3=>3) ERROR: UndefVarError: initn not defined Stacktrace: [1] ConvTranspose(::Tuple{Int64,Int64}, ::Pair{Int64,Int64}, ::Function) at /home/vchuravy/.julia/packages/Flux/hguaX/src/layers/conv.jl:84 (repeats 2 times)

Also:

julia> ConvTranspose((2, 2), 3=>64; init=Flux.glorot_uniform)(rand(4, 4, 3, 10)) ERROR: MethodError: no method matching ∇conv_data(::Array{Float64,4}, ::Array{Float32,4}; stride=(1, 1), pad=(0, 0), dilation=(1, 1)) Closest candidates are: ∇conv_data(::AbstractArray, ::TrackedArray; kw...) at /home/vchuravy/.julia/packages/Flux/hguaX/src/tracker/lib/array.jl:390 ∇conv_data(::TrackedArray, ::AbstractArray; kw...) at /home/vchuravy/.julia/packages/Flux/hguaX/src/tracker/lib/array.jl:391 ∇conv_data(::A<:AbstractArray, ::A<:AbstractArray; size, pad, stride, dilation, flipkernel) where A<:AbstractArray at /home/vchuravy/.julia/packages/NNlib/nf8OC/src/conv.jl:74 ...

Thanks for pointing these out, I've fixed them in the latest commit.

vchuravy · 2018-12-08T22:25:46Z

Thanks! I tried to use it again and all seems to work fine on the CPU, but in the GPU case I get:

julia> layer = ConvTranspose((2, 2), 1=>64, relu, stride=2) |> gpu                                               [18/1889]
ConvTranspose((2, 2)                                                                                                      
                                                                                                                          
julia> layer(train[1])                                                                                                    
ERROR: conversion to pointer not defined for CuArray{Float32,4}                                                           
Stacktrace:                                                                                                               
 [1] error(::String) at ./error.jl:33                                                                                     
 [2] unsafe_convert(::Type{Ptr{Float32}}, ::CuArray{Float32,4}) at ./pointer.jl:67                                        
 [3] pointer at ./abstractarray.jl:861 [inlined]                                                                          
 [4] #conv2d_grad_x!#44(::Tuple{Int64,Int64}, ::Tuple{Int64,Int64}, ::Tuple{Int64,Int64}, ::Int64, ::Int64, ::Function, ::CuArray{Float32,4}, ::CuArray{Float32,4}, ::CuArray{Float32,4}) at /home/vchuravy/.julia/packages/NNlib/3j02R/src/impl/co$v.jl:

tejank10 · 2018-12-09T13:04:33Z

I just made a PR in CuArrays (#223) to reflect the changes in conv for CuArrays. With that, and latest conv_transpose branch of NNlib this should be working.

NTimmons · 2018-12-10T17:53:22Z

This looks like it never got merged. Is there no ConvTranspose in Flux?

tejank10 · 2018-12-12T16:55:43Z

@MikeInnes if #54 is fine then let's get it, and this PR merged?

staticfloat · 2019-02-02T02:03:51Z

src/layers/conv.jl

@@ -77,6 +125,7 @@ struct DepthwiseConv{N,F,A,V}
  bias::V
  stride::NTuple{N,Int}
  pad::NTuple{N,Int}
+  dilation::NTuple{N,Int}


This seems like perhaps an unrelated change?

staticfloat · 2019-02-03T05:00:56Z

In my limited autoencoder testing, this seems to be working, but I'm worried about the failing gradient tests. @tejank10 do you know why they're failing?

tejank10 · 2019-02-03T19:12:48Z

Trying tests with convAPI branch of CuArrays in this PR should get them passing.

MikeInnes · 2019-02-04T12:08:00Z

I just merged that PR. Worth testing this again, figuring out if we need a CuArrays tag etc.

tejank10 · 2019-02-05T13:01:58Z

The gradtests are failing because it is using NNlib v0.4.3. Trying the tests with the NNlib-master seems to be working.

MikeInnes · 2019-02-05T13:04:16Z

You should be able to add NNlib master to the manifest and get it tested that way.

MikeInnes · 2019-02-06T14:14:23Z

Awesome stuff @tejank10, thanks!

sidhantnagpal · 2019-02-06T14:17:18Z

Thanks for working on this. It will be useful in implementing GANs using Flux.

MikeInnes · 2019-02-06T14:20:12Z

Also thanks a lot to @staticfloat for the review!

AzamatB · 2019-02-06T16:13:02Z

Does this support only 2D images, as the name suggests, or 1D and 3D as well?

tejank10 · 2019-02-06T16:44:49Z

It should support 1D and 3D as well.

tejank10 changed the title ~~2D Conv transpose support~~ [WIP] 2D Conv transpose support Jul 9, 2018

avik-pal mentioned this pull request Jul 17, 2018

Image upsampling with fractional stride #319

Closed

make docs on 1.0

d5d9441

MikeInnes force-pushed the master branch 2 times, most recently from 65799d1 to 193c4de Compare September 5, 2018 15:53

1.0 fix for conv transpose

e86365e

tejank10 force-pushed the conv_transpose branch from cdaa0b4 to e86365e Compare September 8, 2018 19:45

tejank10 changed the title ~~[WIP] 2D Conv transpose support~~ 2D Conv transpose support Sep 11, 2018

AzamatB mentioned this pull request Sep 11, 2018

Add max/mean unpooling operations #395

Open

MikeInnes mentioned this pull request Oct 17, 2018

transposed convolution layer #440

Closed

tejank10 added 6 commits October 19, 2018 02:08

Merge branch 'master' of https://github.com/FluxML/Flux.jl into conv_…

e9bf86d

…transpose

merge with upstream

f540a0d

conv_data grad api change

0dc4ec4

in accordance with conv_data api

fca9347

conv_filter api changes

387df8c

in accordance with conv_filter api

a657c28

tejank10 added 8 commits November 28, 2018 10:55

1.0 fix for conv transpose

a71ee38

conv_data grad api change

9c3e34b

in accordance with conv_data api

ca8ad63

conv_filter api changes

10f3a8e

in accordance with conv_filter api

bc9bda9

resolved conflicts

89f2709

merge conflict resolved

95e490a

clean code

519c3db

vchuravy reviewed Dec 2, 2018

View reviewed changes

test/tracker.jl Outdated

@@ -1,7 +1,11 @@

using Flux

using Flux.Tracker, Test, NNlib

<<<<<<< HEAD

Copy link

vchuravy Dec 2, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Merge artificat

vchuravy reviewed Dec 2, 2018

View reviewed changes

fixes for layer and test

1648414

printing ConvTranspose layer

ed835f2

staticfloat reviewed Feb 2, 2019

View reviewed changes

tejank10 added 2 commits February 2, 2019 10:20

Merge branch 'master' into conv_transpose

e54df2d

fixed DepthwiseConv dilation

84eabcd

staticfloat closed this Feb 5, 2019

staticfloat reopened this Feb 5, 2019

Update NNlib to master in Manofest

cc4438c

staticfloat approved these changes Feb 5, 2019

View reviewed changes

MikeInnes merged commit e8b2ec6 into FluxML:master Feb 6, 2019

sidhantnagpal mentioned this pull request Feb 6, 2019

Project Proposal For Contribution to GSoC'19 FluxML/model-zoo#99

Closed

bcdarwin mentioned this pull request Dec 27, 2019

ConvTranspose produces nonsense results for 3D (aka 5D) inputs #978

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

2D Conv transpose support #311

2D Conv transpose support #311

tejank10 commented Jul 9, 2018 •

edited

Loading

tejank10 commented Jul 13, 2018 •

edited

Loading

tejank10 commented Sep 8, 2018

Aniket1998 commented Oct 6, 2018

yuehhua commented Oct 19, 2018

MikeInnes commented Nov 28, 2018

vchuravy Dec 2, 2018

vchuravy Dec 2, 2018

vchuravy Dec 2, 2018

tejank10 Dec 4, 2018

vchuravy commented Dec 8, 2018

tejank10 commented Dec 9, 2018

NTimmons commented Dec 10, 2018

tejank10 commented Dec 12, 2018

staticfloat Feb 2, 2019

staticfloat commented Feb 3, 2019

tejank10 commented Feb 3, 2019 •

edited

Loading

MikeInnes commented Feb 4, 2019

tejank10 commented Feb 5, 2019

MikeInnes commented Feb 5, 2019

MikeInnes commented Feb 6, 2019

sidhantnagpal commented Feb 6, 2019

MikeInnes commented Feb 6, 2019

AzamatB commented Feb 6, 2019 •

edited

Loading

tejank10 commented Feb 6, 2019

2D Conv transpose support #311

2D Conv transpose support #311

Conversation

tejank10 commented Jul 9, 2018 • edited Loading

tejank10 commented Jul 13, 2018 • edited Loading

tejank10 commented Sep 8, 2018

Aniket1998 commented Oct 6, 2018

yuehhua commented Oct 19, 2018

MikeInnes commented Nov 28, 2018

vchuravy Dec 2, 2018

Choose a reason for hiding this comment

vchuravy Dec 2, 2018

Choose a reason for hiding this comment

vchuravy Dec 2, 2018

Choose a reason for hiding this comment

tejank10 Dec 4, 2018

Choose a reason for hiding this comment

vchuravy commented Dec 8, 2018

tejank10 commented Dec 9, 2018

NTimmons commented Dec 10, 2018

tejank10 commented Dec 12, 2018

staticfloat Feb 2, 2019

Choose a reason for hiding this comment

staticfloat commented Feb 3, 2019

tejank10 commented Feb 3, 2019 • edited Loading

MikeInnes commented Feb 4, 2019

tejank10 commented Feb 5, 2019

MikeInnes commented Feb 5, 2019

MikeInnes commented Feb 6, 2019

sidhantnagpal commented Feb 6, 2019

MikeInnes commented Feb 6, 2019

AzamatB commented Feb 6, 2019 • edited Loading

tejank10 commented Feb 6, 2019

tejank10 commented Jul 9, 2018 •

edited

Loading

tejank10 commented Jul 13, 2018 •

edited

Loading

tejank10 commented Feb 3, 2019 •

edited

Loading

AzamatB commented Feb 6, 2019 •

edited

Loading