Example in README does not work correctly? #21
Comments
|
Oh weird. Taking a look now |
|
Same problem with Click to show output:```julia julia> using Metalhead, Metalhead.Flux, Torchjulia> vgg = VGG19() julia> tvgg = Flux.fmap(Torch.to_tensor, vgg) julia> ip = rand(Float32, 224, 224, 3, 1) [:, :, 2, 1] = [:, :, 3, 1] = julia> tip = tensor(ip, dev = 0) [:, :, 2, 1] = [:, :, 3, 1] = julia> vgg(ip) julia> tvgg(tip) julia> ip == convert(Array, tip) julia> versioninfo(; verbose = true) Memory: 187.5776710510254 GB (172788.265625 MB free) julia> import Pkg julia> Pkg.status() julia> Pkg.status(; mode = Pkg.PKGMODE_MANIFEST) |
|
Interesting, could you run specifically with the |
Could you explain how to do that? |
|
Note the |
|
Hmm, not able to reproduce. julia> tresnet2(tip)
1000×5 Tensor{Float32,2}:
0.000490387 2.8881e-18 … 9.43812e-11 2.06312e-6
2.59334e-14 1.0 1.0 0.0783882
0.684608 5.45478e-13 1.51907e-9 5.58125e-5
0.314902 6.37972e-18 1.99655e-8 0.000203497
2.20994e-15 1.11982e-15 4.3811e-22 0.92135
5.03003e-17 0.0756458 … 0.000474865 3.06938e-8 |
|
I tried both of these: tresnet = Flux.fmap(Torch.to_tensor, resnet)tresnet = Flux.fmap(Torch.to_tensor, resnet.layers)Unfortunately, in both of these cases, |
|
Also, the dimension of Click to show output:julia> typeof(ip)
Array{Float32,4}
julia> ndims(ip)
4
julia> size(ip)
(224, 224, 3, 1)
julia> typeof(tip)
Tensor{Float32,4}
julia> ndims(tip)
4
julia> size(tip)
(224, 224, 3, 1)
julia> typeof(resnet(ip))
Array{Float32,2}
julia> ndims(resnet(ip))
2
julia> size(resnet(ip))
(1000, 1)
julia> typeof(tresnet(tip))
Tensor{Float32,2}
julia> ndims(tresnet(tip))
2
julia> size(tresnet(tip))
(1000, 1) |
|
Does converting it back to an |
|
Ah, yes, that should be fixed on master. |
julia> tresnet(tip)
1000×1 Tensor{Float32,2}:
1.0
1.0
1.0
1.0
1.0
1.0
1.0
⋮
1.0
1.0
1.0
1.0
1.0
1.0
1.0
julia> convert(Array, tresnet(tip))
1000×1 Array{Float32,2}:
1.0
1.0
1.0
1.0
1.0
1.0
1.0
⋮
1.0
1.0
1.0
1.0
1.0
1.0
1.0 |
Let me do |
|
So... on julia> using Metalhead, Metalhead.Flux, Torch
julia> resnet = ResNet();
julia> tresnet = Flux.fmap(Torch.to_tensor, resnet.layers);
julia> ip = rand(Float32, 224, 224, 3, 1);
julia> tip = tensor(ip, dev = 0);
julia> resnet(ip)
1000×1 Array{Float32,2}:
0.00087634224
0.000810427
0.0006268254
0.0007971713
0.001212693
0.0013046135
0.0007224348
⋮
0.00033712923
0.0006899684
0.00080008985
0.0007252848
0.0007919692
0.0016419729
0.001108708
julia> tresnet(tip)
ERROR: MethodError: no method matching conv(::Tensor{Float32,4}, ::Tensor{Float32,4}, ::Tensor{Float32,1}, ::DenseConvDims{2,(1, 1),64,64,(1, 1),(0, 0, 0, 0),(1, 1),false}; stride=1, pad=0, dilation=1)
Closest candidates are:
conv(::Tensor{xT,N}, ::Tensor{T,N}, ::Tensor{T,N} where N, ::DenseConvDims{M,K,C_in,C_out,S,P,D,F}; stride, pad, dilation) where {T, N, xT, M, K, C_in, C_out, S, P, D, F} at /users/daluthge/.julia/packages/Torch/W539X/src/nnlib.jl:10
conv(::Tensor, ::Tensor, ::DenseConvDims; stride, pad, dilation) at /users/daluthge/.julia/packages/Torch/W539X/src/nnlib.jl:15
conv(::Any, ::AbstractArray{T,N}; stride, pad, dilation, flipped) where {T, N} at /users/daluthge/.julia/packages/NNlib/FAI3o/src/conv.jl:174
...
Stacktrace:
[1] conv(::Tensor{Float32,4}, ::Tensor{Float32,4}, ::DenseConvDims{2,(1, 1),64,64,(1, 1),(0, 0, 0, 0),(1, 1),false}; stride::Int64, pad::Int64, dilation::Int64) at /users/daluthge/.julia/packages/Torch/W539X/src/nnlib.jl:16
[2] conv(::Tensor{Float32,4}, ::Tensor{Float32,4}, ::DenseConvDims{2,(1, 1),64,64,(1, 1),(0, 0, 0, 0),(1, 1),false}) at /users/daluthge/.julia/packages/Torch/W539X/src/nnlib.jl:15
[3] (::Conv{2,2,typeof(identity),Tensor{Float32,4},Tensor{Float32,1}})(::Tensor{Float32,4}) at /users/daluthge/.julia/packages/Flux/Fj3bt/src/layers/conv.jl:61
[4] (::Metalhead.ResidualBlock)(::Tensor{Float32,4}) at /users/daluthge/.julia/packages/Metalhead/RZn9O/src/resnet.jl:26
[5] applychain(::Tuple{Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,MeanPool{2,4},Metalhead.var"#103#104",Dense{typeof(identity),Tensor{Float32,2},Tensor{Float32,1}},typeof(softmax)}, ::Tensor{Float32,4}) at /users/daluthge/.julia/packages/Flux/Fj3bt/src/layers/basic.jl:36 (repeats 3 times)
[6] (::Chain{Tuple{Conv{2,2,typeof(identity),Tensor{Float32,4},Tensor{Float32,1}},MaxPool{2,2},Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,Metalhead.ResidualBlock,MeanPool{2,4},Metalhead.var"#103#104",Dense{typeof(identity),Tensor{Float32,2},Tensor{Float32,1}},typeof(softmax)}})(::Tensor{Float32,4}) at /users/daluthge/.julia/packages/Flux/Fj3bt/src/layers/basic.jl:38
[7] top-level scope at REPL[6]:1 |
|
Oh oops, let me check that out. Sorry about that! |
|
Seems to have broken in #19 |
|
I tried on julia> tresnet(tip)
1000×1 Tensor{Float32,2}:
1.0
1.0
1.0
1.0
1.0
1.0
1.0
⋮
1.0
1.0
1.0
1.0
1.0
1.0
1.0 |
|
Let me try and reproduce on a different machine. |
|
I'm able to reproduce this issue as well, MWE follows: using NNlib
using Torch
ip = permutedims(Float32[1 2 3])
tip = Torch.to_tensor(ip) # 0 => GPU:0 in Torch
println("Default")
@show op = softmax(ip)
println("Torch")
# This line fails because dims=1
# @show top = softmax(tip) |> collect
@show top = softmax(tip, dims=0) |> collect
@assert op ≈ topIn short, torch dimensions start at 0, so the default dim of 1 was incorrectly calculating softmax along the batch dimension instead of the label dim. This also applies to mean, sum and any other method using dimension indices. @dhairyagandhi96 should this indexing match julia's 1-based dims? If so, I can put in a PR for the aforementioned methods. |
|
It should, I'd appreciate a PR It had seemed handled, in the tests I had run earlier. Do add this to the test suite. |
|
@DilumAluthge could you confirm that omitting the |
Sorry, I'm very new to everything related to Flux, Metalhead, etc. In the README example, how would I modify the example to omit the softmax layer? |
Instead of just patching a certain set of methods, would it be better to have a robust way of automatically converting between Julia's 1-based indexing and Python's/C++'s 0-based indexing whenever you convert a Julia (Flux, Metalhead, etc.) model to a PyTorch model? |
So, I had a crack at doing just this while trying to make sum and mean work. Turns out it's a much deeper rabbit hole than expected! To elaborate, the current On a more general note, maybe it would help to deduplicate some of the work here with ThArrays.jl? For example, I noticed that they already have a stride-aware tensor constructor (see https://github.com/TuringLang/ThArrays.jl/blob/05bd5ceb2d358dee7be1fe3242d796805313b7e5/csrc/torch_capi_tensor.cpp#L40), a pretty complete device API, etc. |
|
We can check the API with For the What do we currently lack from our device API? Some choices to turn off |
you just need to call the forward pass with |
|
Using |
I guess the question is whether or not possible to avoid lines like Line 25 in e658fa2 from_blob, I was able to get pretty close to Base semantics locally:
# Performs error checking and converts singleton dimensions to tuples
function _normalize_dims(t, dims)
inds = collect(reduced_indices(t, dims))
eachindex(inds)[length.(inds) .== 1] .- 1
end
function Statistics.mean(t::Tensor{T,N}; dims = :) where {T,N}
ptr = Ref(Ptr{Cvoid}())
if dims isa Colon
atg_mean(ptr, t.ptr, options[T])
Tensor{T,0}(ptr[], on(t))
else
# To match Julia's behaviour, we always keep dimensions when dims != :
atg_mean1(ptr, t.ptr, dims, length(dims), true, options[T])
Tensor{T,N-length(dims)}(ptr[], on(t))
end
end
function Statistics.sum(t::Tensor{T,N}; dims = :) where {T,N}
ptr = Ref(Ptr{Cvoid}())
if dims isa Colon
atg_sum(ptr, t.ptr, options[T])
Tensor{T,0}(ptr[], on(t))
else
@show dims = _normalize_dims(t, dims)
atg_sum1(ptr, t.ptr, dims, length(dims), true, options[T])
Tensor{T,N-length(dims)}(ptr[], on(t))
end
end |
Absolutely! Reversing the dimensions every time seems error prone. Ideally we'd have a more automatic solution that can interface with the binary with minimal handling on that side since changes there would be harder to debug. |
|
Maybe PR the dims change? |
|
Will see if I can find some time later tomorrow, will probably hack the from_blob/tensor constructor changes first so that size reversal can be completely excised from the implementation (no more |
|
I have a naive |
|
Add a look at the c++ -> c wrapper and I think |
|
Yeah, that would be fine. |
|
should be fixed on master |
I'm getting a weird result when I run the example in the README. Specifically, the output of
tresnet(tip)produces all ones, while the output ofresnet(ip)seems to produce the correct output.Click to show output:
The text was updated successfully, but these errors were encountered: