Work in progress in migrating the TensorFlow.jl into the Julia 1.0.0 #419

adamryczkowski · 2018-08-19T14:50:35Z

Don't merge it blindly, because this patch still contains at least one bug:

julia> import TensorFlow
[ Info: Precompiling TensorFlow [1d978283-2c37-5f34-9a8e-e9c0ece82495]
┌ Warning: Loading a new version of TensorFlow.jl for the first time. This initial load can take around 5 minutes as code is precompiled; subsequent usage will only take a few seconds.
└ @ TensorFlow ~/.julia/packages/TensorFlow/qrmc9/src/TensorFlow.jl:6
ERROR: LoadError: UndefVarError: broadcast_c not defined
Stacktrace:
 [1] include at ./boot.jl:317 [inlined]
 [2] include_relative(::Module, ::String) at ./loading.jl:1038
 [3] include(::Module, ::String) at ./sysimg.jl:29
 [4] top-level scope at none:2
 [5] eval at ./boot.jl:319 [inlined]
 [6] eval(::Expr) at ./client.jl:389
 [7] top-level scope at ./none:3
in expression starting at /home/adam/.julia/packages/TakingBroadcastSeriously/WQSQT/src/TakingBroadcastSeriously.jl:3
ERROR: LoadError: LoadError: Failed to precompile TakingBroadcastSeriously [fa778a35-eff5-5ebc-91c5-11ea73160f07] to /home/adam/.julia/compiled/v1.1/TakingBroadcastSeriously/RVPKc.ji.
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] macro expansion at ./logging.jl:313 [inlined]
 [3] compilecache(::Base.PkgId, ::String) at ./loading.jl:1184
 [4] _require(::Base.PkgId) at ./logging.jl:311
 [5] require(::Base.PkgId) at ./loading.jl:852
 [6] require(::Module, ::Symbol) at ./logging.jl:311
 [7] include at ./boot.jl:317 [inlined]
 [8] include_relative(::Module, ::String) at ./loading.jl:1038
 [9] include at ./sysimg.jl:29 [inlined]
 [10] include(::String) at /home/adam/.julia/packages/TensorFlow/qrmc9/src/TensorFlow.jl:2
 [11] top-level scope at none:0
 [12] include at ./boot.jl:317 [inlined]
 [13] include_relative(::Module, ::String) at ./loading.jl:1038
 [14] include(::Module, ::String) at ./sysimg.jl:29
 [15] top-level scope at none:2
 [16] eval at ./boot.jl:319 [inlined]
 [17] eval(::Expr) at ./client.jl:389
 [18] top-level scope at ./none:3
in expression starting at /home/adam/.julia/packages/TensorFlow/qrmc9/src/ops.jl:5
in expression starting at /home/adam/.julia/packages/TensorFlow/qrmc9/src/TensorFlow.jl:189
ERROR: Failed to precompile TensorFlow [1d978283-2c37-5f34-9a8e-e9c0ece82495] to /home/adam/.julia/compiled/v1.1/TensorFlow/IhIhf.ji.
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] macro expansion at ./logging.jl:313 [inlined]
 [3] compilecache(::Base.PkgId, ::String) at ./loading.jl:1184
 [4] macro expansion at ./logging.jl:311 [inlined]
 [5] _require(::Base.PkgId) at ./loading.jl:941
 [6] require(::Base.PkgId) at ./loading.jl:852
 [7] require(::Module, ::Symbol) at ./logging.jl:311

I believe that the proper solution is to factor out the TakingBroadcastSeriously dependency, but that task is still to difficult for me to do.

malmaud · 2018-08-20T01:31:11Z

Thanks a ton for doing this work! I think you're right that we should get the broadcasting stuff using the new Julia 1.0 techniques as a first step. I'm cautiously optimistic I'll have time to get to that this or next week, although I'm happy for someone to beat me to it.

oxinabox · 2018-08-20T03:33:02Z

We should represent unknown sizes with missings, rather than nullables, as they match some of the logic needed. But thatbcan come later

oxinabox · 2018-08-22T12:15:36Z

I have pushed some commits to this
it is still not working it does take broadcasting seriously. (I think).

So someone else can continue with it from there.

Current error is that it uses JLD.
Cross-ref: #414
Once that is merged, a quick rebase will sort the current error.
Which is all

ERROR: LoadError: LoadError: Failed to precompile JLD

Now, back to my thesis

oxinabox · 2018-08-22T12:23:14Z

can femtocleaner be pointed in this direction?
Since this branch is on @adamryczkowski 's account,
they might have to add Femtoclearner
and configure it to point at this repo.
Following:
https://github.com/JuliaComputing/FemtoCleaner.jl#invoking-femtocleaner

oxinabox · 2018-08-22T13:20:52Z

@adamryczkowski can you rebase this? Since #414 is merged.

Fix deprecations

adamryczkowski · 2018-08-27T08:06:12Z

I see. logistic.jl used to work before your commit, now it throwed the error. Now it is fixed. Thank you.

…tp://www.imagemagick.org Copyright: Copyright (C) 1999-2014 ImageMagick Studio LLC Features: DPC Modules OpenMP Delegates: bzlib cairo djvu fftw fontconfig freetype jbig jng jpeg lcms lqr ltdl lzma openexr pangocairo png rsvg tiff wmf x xml zlib Usage: convert [options ...] file [ [options ...] file ...] [options ...] file Image Settings: -adjoin join images into a single multi-image file -affine matrix affine transform matrix -alpha option activate, deactivate, reset, or set the alpha channel -antialias remove pixel-aliasing -authenticate password decipher image with this password -attenuate value lessen (or intensify) when adding noise to an image -background color background color -bias value add bias when convolving an image -black-point-compensation use black point compensation -blue-primary point chromaticity blue primary point -bordercolor color border color -caption string assign a caption to an image -channel type apply option to select image channels -clip-mask filename associate a clip mask with the image -colors value preferred number of colors in the image -colorspace type alternate image colorspace -comment string annotate image with comment -compose operator set image composite operator -compress type type of pixel compression when writing the image -define format:option define one or more image format options -delay value display the next image after pausing -density geometry horizontal and vertical density of the image -depth value image depth -direction type render text right-to-left or left-to-right -display server get image or font from this X server -dispose method layer disposal method -dither method apply error diffusion to image -encoding type text encoding type -endian type endianness (MSB or LSB) of the image -family name render text with this font family -fill color color to use when filling a graphic primitive -filter type use this filter when resizing an image -font name render text with this font -format "string" output formatted image characteristics -fuzz distance colors within this distance are considered equal -gravity type horizontal and vertical text placement -green-primary point chromaticity green primary point -intensity method method to generate intensity value from pixel -intent type type of rendering intent when managing the image color -interlace type type of image interlacing scheme -interline-spacing value set the space between two text lines -interpolate method pixel color interpolation method -interword-spacing value set the space between two words -kerning value set the space between two letters -label string assign a label to an image -limit type value pixel cache resource limit -loop iterations add Netscape loop extension to your GIF animation -mask filename associate a mask with the image -matte store matte channel if the image has one -mattecolor color frame color -moments report image moments -monitor monitor progress -orient type image orientation -page geometry size and location of an image canvas (setting) -ping efficiently determine image attributes -pointsize value font point size -precision value maximum number of significant digits to print -preview type image preview type -quality value JPEG/MIFF/PNG compression level -quiet suppress all warning messages -red-primary point chromaticity red primary point -regard-warnings pay attention to warning messages -remap filename transform image colors to match this set of colors -respect-parentheses settings remain in effect until parenthesis boundary -sampling-factor geometry horizontal and vertical sampling factor -scene value image scene number -seed value seed a new sequence of pseudo-random numbers -size geometry width and height of image -stretch type render text with this font stretch -stroke color graphic primitive stroke color -strokewidth value graphic primitive stroke width -style type render text with this font style -support factor resize support: > 1.0 is blurry, < 1.0 is sharp -synchronize synchronize image to storage device -taint declare the image as modified -texture filename name of texture to tile onto the image background -tile-offset geometry tile offset -treedepth value color tree depth -transparent-color color transparent color -undercolor color annotation bounding box color -units type the units of image resolution -verbose print detailed information about the image -view FlashPix viewing transforms -virtual-pixel method virtual pixel access method -weight type render text with this font weight -white-point point chromaticity white point Image Operators: -adaptive-blur geometry adaptively blur pixels; decrease effect near edges -adaptive-resize geometry adaptively resize image using 'mesh' interpolation -adaptive-sharpen geometry adaptively sharpen pixels; increase effect near edges -alpha option on, activate, off, deactivate, set, opaque, copy transparent, extract, background, or shape -annotate geometry text annotate the image with text -auto-gamma automagically adjust gamma level of image -auto-level automagically adjust color levels of image -auto-orient automagically orient (rotate) image -bench iterations measure performance -black-threshold value force all pixels below the threshold into black -blue-shift factor simulate a scene at nighttime in the moonlight -blur geometry reduce image noise and reduce detail levels -border geometry surround image with a border of color -bordercolor color border color -brightness-contrast geometry improve brightness / contrast of the image -canny geometry detect edges in the image -cdl filename color correct with a color decision list -charcoal radius simulate a charcoal drawing -chop geometry remove pixels from the image interior -clamp keep pixel values in range (0-QuantumRange) -clip clip along the first path from the 8BIM profile -clip-path id clip along a named path from the 8BIM profile -colorize value colorize the image with the fill color -color-matrix matrix apply color correction to the image -connected-components connectivity connected-components uniquely labeled -contrast enhance or reduce the image contrast -contrast-stretch geometry improve contrast by `stretching' the intensity range -convolve coefficients apply a convolution kernel to the image -cycle amount cycle the image colormap -decipher filename convert cipher pixels to plain pixels -deskew threshold straighten an image -despeckle reduce the speckles within an image -distort method args distort images according to given method ad args -draw string annotate the image with a graphic primitive -edge radius apply a filter to detect edges in the image -encipher filename convert plain pixels to cipher pixels -emboss radius emboss an image -enhance apply a digital filter to enhance a noisy image -equalize perform histogram equalization to an image -evaluate operator value evaluate an arithmetic, relational, or logical expression -extent geometry set the image size -extract geometry extract area from image -features distance analyze image features (e.g. contrast, correlation) -fft implements the discrete Fourier transform (DFT) -flip flip image vertically -floodfill geometry color floodfill the image with color -flop flop image horizontally -frame geometry surround image with an ornamental border -function name parameters apply function over image values -gamma value level of gamma correction -gaussian-blur geometry reduce image noise and reduce detail levels -geometry geometry preferred size or location of the image -grayscale method convert image to grayscale -hough-lines geometry identify lines in the image -identify identify the format and characteristics of the image -ift implements the inverse discrete Fourier transform (DFT) -implode amount implode image pixels about the center -interpolative-resize geometry resize image using 'point sampled' interpolation -kuwahara geometry edge preserving noise reduction filter -lat geometry local adaptive thresholding -level value adjust the level of image contrast -level-colors color,color level image with the given colors -linear-stretch geometry improve contrast by `stretching with saturation' -liquid-rescale geometry rescale image with seam-carving -magnify double the size of the image with pixel art scaling -mean-shift geometry delineate arbitrarily shaped clusters in the image -median geometry apply a median filter to the image -mode geometry make each pixel the 'predominant color' of the neighborhood -modulate value vary the brightness, saturation, and hue -monochrome transform image to black and white -morphology method kernel apply a morphology method to the image -motion-blur geometry simulate motion blur -negate replace every pixel with its complementary color -noise geometry add or reduce noise in an image -normalize transform image to span the full range of colors -opaque color change this color to the fill color -ordered-dither NxN add a noise pattern to the image with specific amplitudes -paint radius simulate an oil painting -perceptible epsilon pixel value less than |epsilon| become epsilon or -epsilon -polaroid angle simulate a Polaroid picture -posterize levels reduce the image to a limited number of color levels -profile filename add, delete, or apply an image profile -quantize colorspace reduce colors in this colorspace -radial-blur angle radial blur the image (deprecated use -rotational-blur -raise value lighten/darken image edges to create a 3-D effect -random-threshold low,high random threshold the image -region geometry apply options to a portion of the image -render render vector graphics -repage geometry size and location of an image canvas -resample geometry change the resolution of an image -resize geometry resize the image -roll geometry roll an image vertically or horizontally -rotate degrees apply Paeth rotation to the image -rotational-blur angle rotational blur the image -sample geometry scale image with pixel sampling -scale geometry scale the image -segment values segment an image -selective-blur geometry selectively blur pixels within a contrast threshold -sepia-tone threshold simulate a sepia-toned photo -set property value set an image property -shade degrees shade the image using a distant light source -shadow geometry simulate an image shadow -sharpen geometry sharpen the image -shave geometry shave pixels from the image edges -shear geometry slide one edge of the image along the X or Y axis -sigmoidal-contrast geometry increase the contrast without saturating highlights or shadows -sketch geometry simulate a pencil sketch -solarize threshold negate all pixels above the threshold level -sparse-color method args fill in a image based on a few color points -splice geometry splice the background color into the image -spread radius displace image pixels by a random amount -statistic type geometry replace each pixel with corresponding statistic from the neighborhood -strip strip image of all profiles and comments -swirl degrees swirl image pixels about the center -threshold value threshold the image -thumbnail geometry create a thumbnail of the image -tile filename tile image when filling a graphic primitive -tint value tint the image with the fill color -transform affine transform image -transparent color make this color transparent within the image -transpose flip image vertically and rotate 90 degrees -transverse flop image horizontally and rotate 270 degrees -trim trim image edges -type type image type -unique-colors discard all but one of any pixel color -unsharp geometry sharpen the image -vignette geometry soften the edges of the image in vignette style -wave geometry alter an image along a sine wave -white-threshold value force all pixels above the threshold into white Image Sequence Operators: -append append an image sequence -clut apply a color lookup table to the image -coalesce merge a sequence of images -combine combine a sequence of images -compare mathematically and visually annotate the difference between an image and its reconstruction -complex operator perform complex mathematics on an image sequence -composite composite image -crop geometry cut out a rectangular region of the image -deconstruct break down an image sequence into constituent parts -evaluate-sequence operator evaluate an arithmetic, relational, or logical expression -flatten flatten a sequence of images -fx expression apply mathematical expression to an image channel(s) -hald-clut apply a Hald color lookup table to the image -layers method optimize, merge, or compare image layers -morph value morph an image sequence -mosaic create a mosaic from an image sequence -poly terms build a polynomial from the image sequence and the corresponding terms (coefficients and degree pairs). -print string interpret string and print to console -process arguments process the image with a custom image filter -separate separate an image channel into a grayscale image -smush geometry smush an image sequence together -write filename write images to this file Image Stack Operators: -clone indexes clone an image -delete indexes delete the image from the image sequence -duplicate count,indexes duplicate an image one or more times -insert index insert last image into the image sequence -reverse reverse image sequence -swap indexes swap two images in the image sequence Miscellaneous Options: -debug events display copious debugging information -distribute-cache port distributed pixel cache spanning one or more servers -help print program options -list type print a list of supported option arguments -log format format of debugging information -version print version information By default, the image format of `file' is determined by its magic number. To specify a particular image format, precede the filename with an image format name and a colon (i.e. ps:image) or specify the image type as the filename suffix (i.e. image.ps). Specify 'file' as '-' for standard input or output. an object of type String to an object of type Array{UInt8,1}

adamryczkowski · 2018-08-27T14:48:49Z

I think I stumped upon another non-trivial issue. To reproduce:

julia> import Pkg
julia> Pkg.add(Pkg.PackageSpec(url="https://github.com/adamryczkowski/TensorFlow.jl"))
julia> import TensorFlow
julia> filename = joinpath(mktempdir(), "tfrecords")

julia> TensorFlow.io.RecordWriter(filename) do writer
    write(writer, UInt8[1,2])
    write(writer, "hi")
end

julia> reader = TensorFlow.io.RecordIterator(filename)
julia> records = collect(reader)
ERROR: MethodError: Cannot `convert` an object of type String to an object of type Array{UInt8,1}
Closest candidates are:
  convert(::Type{Array{UInt8,1}}, ::PyCall.PyObject) at /home/adam/.julia/packages/PyCall/akNFy/src/conversions.jl:118
  convert(::Type{T<:Array}, ::AbstractArray) where T<:Array at array.jl:489
  convert(::Type{T<:AbstractArray}, ::T<:AbstractArray) where T<:AbstractArray at abstractarray.jl:14
  ...
Stacktrace:
 [1] convert(::Type{Nullables.Nullable{Array{UInt8,1}}}, ::Nullables.Nullable{String}) at /home/adam/.julia/packages/Nullables/3JHB8/src/nullable.jl:57
 [2] TensorFlow.io.TFRecord.RecordIteratorState(::Nullables.Nullable{String}) at /home/adam/tmp/TensorFlow.jl/src/io/tfrecord.jl:57
 [3] macro expansion at /home/adam/tmp/TensorFlow.jl/src/TensorFlow.jl:179 [inlined]
 [4] _next(::TensorFlow.io.TFRecord.RecordIterator) at /home/adam/tmp/TensorFlow.jl/src/io/tfrecord.jl:76
 [5] iterate at /home/adam/tmp/TensorFlow.jl/src/io/tfrecord.jl:94 [inlined]
 [6] _collect(::UnitRange{Int64}, ::TensorFlow.io.TFRecord.RecordIterator, ::Base.HasEltype, ::Base.SizeUnknown) at ./array.jl:568
 [7] collect(::TensorFlow.io.TFRecord.RecordIterator) at ./array.jl:557
 [8] top-level scope at none:0

One obvious solution is to overload convert so the following code succeeds:

convert(Vector{UInt8}, "Hello world")

But obviously, this change will interfere with the language, and that is why this conversion is not built-in. I have read the discussion JuliaLang/julia#24388 but still I am not sure, how to solve this problem cleanly.

oxinabox · 2018-08-27T16:10:41Z

In this case, how to fix that doesn't matter.
Because now julia's iteration protocol and python's are in sync -- no need to buffer on item ahead anymore
Which means we can just delete the RecordIteratorState type.

Which looks something like:

struct RecordIterator
    pyo::Distributed.Future
end


"""
    `RecordIterator(path)`
Returns a Julia iterator that returns the records in the TF Record file
at `path` as `Vector{UInt8}` objects.
"""
function RecordIterator(path::AbstractString)
    pyo = @tf.py_proc py_tf[][:python_io][:tf_record_iterator]($path)
    RecordIterator(pyo)
end

function _next(iter::RecordIterator)
    try
        @static if PyCall.pyversion >= v"3.0.0"
            fetch(@tf.py_proc $(iter.pyo)[:__next__]())
        else
            #Python 2
            fetch(@tf.py_proc $(iter.pyo)[:next]())
        end
        
    catch err
        if isa(err, Distributed.RemoteException) && isa(err.captured.ex, PyCall.PyError)
            # Only catch it, if it could be an  StopIteration exception thrown in python
            # which signifies the end of iteration being reached normally
            nothing # signal to stop julia iteration
        else
            rethrow(err)
        end
    end
end

function Base.iterate(iter::RecordIterator, state=iter)
	record = _next(iter)
	if record isa nothing
		nothing # end iteration
	else
		(record, iter)
	end
end

I don't have a working copy of libtensorflow.so right now so can't test.
But that is the gist of it.

May also need to change the eltype of the iterator itself, or remove the declaration of the eltype.

adamryczkowski · 2018-08-27T19:43:14Z

Many thanks!

adamryczkowski · 2018-08-28T08:23:36Z

I do not understand, how your code would fix the need to have an implicit conversion from String to Array{UInt8,1}.

The error remains exactly the same. But I found a very simple fix. Just convert whatever Python returned into a byte array.

Now I found an interesting problem (and already fixed it) - I do not have Julia 0.6.x to test it, but I have a reason to believe, that polygamma and zeta could not possibly have worked due to a bug. These functions were not exported (i.e. defined) by the TensorFlow, so either the math.jl test failed (as it did in Julia 1.0.0) or Julia cheated and silently used the native version of those functions.

I have inserted

for jl_func_name in [
    :polygamma,
    :zeta]
    @eval @define_binary SpecialFunctions.$jl_func_name Ops.$jl_func_name
end

to ops/math.jl.

I still have 1-2 hours of time to work on the porting. I will tell you, @oxinabox when I am done for today.

oxinabox · 2018-08-28T09:50:15Z

src/ops/imported_ops.jl

@@ -2,6 +2,7 @@

 module Ops
 import TensorFlow
+import SpecialFunctions


You shouldn't directly edit this file BTW.
It is automatically generated,

But this change does nothing anyway, so it can happily be removed.

oxinabox · 2018-08-28T10:26:34Z

Now I found an interesting problem (and already fixed it) - I do not have Julia 0.6.x to test it, but I have a reason to believe, that polygamma and zeta could not possibly have worked due to a bug. These functions were not exported (i.e. defined) by the TensorFlow, so either the math.jl test failed (as it did in Julia 1.0.0) or Julia cheated and silently used the native version of those functions.

Yeah, I see what you mean.
I don't know why it works in the tests in julia 0.6.
But it does (I just tested), and it even gives the right answer. (I added a println).
But the code to calculate that doesn't even exist in the julia 0.6 source.
So something funky is going on.

Very mysterious but ultimately not a concern

I do not understand, how your code would fix the need to have an implicit conversion from String to Array{UInt8,1}.
The error remains exactly the same. But I found a very simple fix. Just convert whatever Python returned into a byte array.

huh, I must have misread which line the error was on, I thought it was in the constructor for the RecordIteratorState
But yes, in general one does need to convert what python returns.

adamryczkowski · 2018-08-28T13:04:23Z

@malmaud I found a bug in the existing TensorFlow code which apparently slipped the unit tests. You might want to take a look, because there may be more potential bugs that were not the detected by the unit tests as well. The bug is described in my previous comment.

adamryczkowski · 2018-08-28T13:36:18Z

Suppose I have

a = TensorFlow.constant(10)

How can I get the number 10 out of a without calling TensorFlow (i.e. without calling run(TensorFlow.Session(), a)?

Is it possible?

I believe I need it to solve a following bug

julia> using TensorFlow
julia> import LinearAlgebra

julia> a = TensorFlow.constant(LinearAlgebra.Diagonal(ones(5,5)))
julia> idxs = TensorFlow.constant(map(Int64, [1,1,2,3,1]))
julia> n = TensorFlow.constant(Int32(3))
julia> d = unsorted_segment_sum(a, idxs, n);
julia> TensorFlow.get_shape(d)
ERROR: MethodError: no method matching Array(::TensorFlow.RawTensor)
Closest candidates are:
  Array(::LinearAlgebra.SymTridiagonal) at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.0/LinearAlgebra/src/tridiag.jl:111
  Array(::LinearAlgebra.Tridiagonal) at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.0/LinearAlgebra/src/tridiag.jl:518
  Array(::LinearAlgebra.AbstractTriangular) at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.0/LinearAlgebra/src/triangular.jl:106
  ...
Stacktrace:
 [1] load_const(::Operation) at /home/adam/tmp/TensorFlow.jl/src/shape_inference.jl:260
 [2] load_const at /home/adam/tmp/TensorFlow.jl/src/shape_inference.jl:331 [inlined]
 [3] (::getfield(TensorFlow.ShapeInference, Symbol("##42#43")))(::Operation) at /home/adam/tmp/TensorFlow.jl/src/shape_inference.jl:425
 [4] _get_shape(::Tensor{Float64}) at /home/adam/tmp/TensorFlow.jl/src/shape_inference.jl:107
 [5] get_shape(::Tensor{Float64}) at /home/adam/tmp/TensorFlow.jl/src/shape_inference.jl:78
 [6] top-level scope at none:0

Or maybe the solution is different. That is all I can do for today.

oxinabox · 2018-08-28T14:39:55Z

The normal way to do that is to use load_const,
which is the function where your error occurs.
There are lower level ways to do it like

julia> x = constant(10)
<Tensor Const_2:1 shape=() dtype=Int64>

julia> x.op["value"].tensor.int64_val
1-element Array{Int64,1}:
 10

In anycase
That error to me looks like the use of a conversion-constructor that has plagued use constantly in this PR.

In particular,
I expect that

value = Array(tf.load_proto(get_def(op).attr["value"]))

Should be changed to

raw_value = tf.load_proto(get_def(op).attr["value"])
value  = convert(Array, raw_value)

Not tested, but that is my guess..

malmaud · 2018-08-28T15:07:10Z

wrt how we'll merge this, my opinion is just a merge commit. Then everyone's contributions are clearly recorded but the main branch of the history still doesn't have a plethora of intermediate commits.

malmaud · 2018-08-29T00:10:06Z

Tests pass, so I'm going to merge this now. Thanks everyone for putting this together!

adamryczkowski · 2018-08-29T11:29:35Z

I can confirm, that all tests pass in all my systems as well. Just make sure, you have proper Python installed with the TensorFlow library accessible from it.

adamryczkowski added 14 commits August 19, 2018 12:10

Fixes is_windows depreciation

c324df3

Fixes depreciations in the mv command

c661209

Replace info, warn & error with the macros

70d50d7

Replace info, warn & error with the macros

ac2ad35

Replace Void into Cvoid to avoid deprecation

c2e9b5e

Add Nullable dependency

cb0bfec

Add Nullable dependency

03e058b

Another deprecation fix

d404c71

Another deprecation fix with ...

6a6f45f

Complex64, Coplex128 -> ComplexF64, ComplexF128

7b7c132

Complex64, Coplex128 -> ComplexF32, ComplexF64

447fef1

Complex64, Coplex128 -> ComplexF32, ComplexF64

3022797

Iterator for OperationIterator converted to 1.0 syntax

06b9a38

Iterator for OperationIterator converted to 1.0 syntax

42c4be7

adamryczkowski requested review from malmaud and oxinabox as code owners August 19, 2018 14:50

adamryczkowski mentioned this pull request Aug 19, 2018

build fails on Julia 1.0 with "is_windows not defined" #418

Closed

femtocleaner bot and others added 2 commits August 21, 2018 16:11

Fix deprecations

df56552

fix up broadcasting

038cf86

oxinabox self-assigned this Aug 22, 2018

oxinabox added 2 commits August 22, 2018 19:52

fix math

416dae4

put using Nullables in the places needed

b8270fc

oxinabox removed their assignment Aug 22, 2018

use 0.7 travis ci

70316ef

Merge pull request #1 from adamryczkowski/fbot/deps

42c2447

Fix deprecations

oxinabox mentioned this pull request Aug 28, 2018

WIP: Just enough of 0.7 to load #376

Closed

Removes RecordIteratorState type

b7a8506

Cumulative batch of deprecation fixes

fc9add4

oxinabox reviewed Aug 28, 2018

View reviewed changes

malmaud added 12 commits August 28, 2018 11:21

Fix build on OS X.

3bf08ba

Explicitly call 'convert' for Arrays.

8c5c576

Fix issues in math.jl.

1ecc98d

More broadcasting fixes.

9bd1c60

Fix meta.

8e0c538

Fix squeeze to dropdims.

835de9a

Fixes to nn.jl

5539580

Shape inference.

ac7981b

Fix transformations.

1b79036

Various fixes.

06ddcf4

Disable allowed failures

b30079a

Update travis

111178c

malmaud merged commit ce253c7 into malmaud:master Aug 29, 2018

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Work in progress in migrating the TensorFlow.jl into the Julia 1.0.0 #419

Work in progress in migrating the TensorFlow.jl into the Julia 1.0.0 #419

adamryczkowski commented Aug 19, 2018

malmaud commented Aug 20, 2018

oxinabox commented Aug 20, 2018

oxinabox commented Aug 22, 2018 •

edited

Loading

oxinabox commented Aug 22, 2018

oxinabox commented Aug 22, 2018

adamryczkowski commented Aug 27, 2018

adamryczkowski commented Aug 27, 2018

oxinabox commented Aug 27, 2018

adamryczkowski commented Aug 27, 2018

adamryczkowski commented Aug 28, 2018

oxinabox Aug 28, 2018

oxinabox commented Aug 28, 2018 •

edited

Loading

adamryczkowski commented Aug 28, 2018

adamryczkowski commented Aug 28, 2018

oxinabox commented Aug 28, 2018

malmaud commented Aug 28, 2018

malmaud commented Aug 29, 2018

adamryczkowski commented Aug 29, 2018 •

edited

Loading

Work in progress in migrating the TensorFlow.jl into the Julia 1.0.0 #419

Work in progress in migrating the TensorFlow.jl into the Julia 1.0.0 #419

Conversation

adamryczkowski commented Aug 19, 2018

malmaud commented Aug 20, 2018

oxinabox commented Aug 20, 2018

oxinabox commented Aug 22, 2018 • edited Loading

oxinabox commented Aug 22, 2018

oxinabox commented Aug 22, 2018

adamryczkowski commented Aug 27, 2018

adamryczkowski commented Aug 27, 2018

oxinabox commented Aug 27, 2018

adamryczkowski commented Aug 27, 2018

adamryczkowski commented Aug 28, 2018

oxinabox Aug 28, 2018

Choose a reason for hiding this comment

oxinabox commented Aug 28, 2018 • edited Loading

adamryczkowski commented Aug 28, 2018

adamryczkowski commented Aug 28, 2018

oxinabox commented Aug 28, 2018

malmaud commented Aug 28, 2018

malmaud commented Aug 29, 2018

adamryczkowski commented Aug 29, 2018 • edited Loading

oxinabox commented Aug 22, 2018 •

edited

Loading

oxinabox commented Aug 28, 2018 •

edited

Loading

adamryczkowski commented Aug 29, 2018 •

edited

Loading