Change irim block - add invertible UNET #57
Conversation
Codecov ReportBase: 88.11% // Head: 87.94% // Decreases project coverage by
Additional details and impacted files@@ Coverage Diff @@
## master #57 +/- ##
==========================================
- Coverage 88.11% 87.94% -0.17%
==========================================
Files 31 32 +1
Lines 2330 2390 +60
==========================================
+ Hits 2053 2102 +49
- Misses 277 288 +11
Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here. ☔ View full report at Codecov. |
| C::Conv1x1 | ||
| RB::Union{ResidualBlock, FluxBlock} | ||
| C::AbstractArray{Conv1x1, 1} | ||
| RB::AbstractArray{ResidualBlock, 1} | ||
| end |
There was a problem hiding this comment.
Why are we removing FluxBlock from allowed types?
| @@ -75,10 +75,18 @@ end | |||
| # Constructors | |||
|
|
|||
| # Constructor | |||
| function ResidualBlock(n_in, n_hidden; k1=3, k2=3, p1=1, p2=1, s1=1, s2=1, fan=false, ndims=2) | |||
| function ResidualBlock(n_in, n_hidden; d=nothing, k1=3, k2=3, p1=1, p2=1, s1=1, s2=1, fan=false, ndims=2) | |||
There was a problem hiding this comment.
Not clear what d is. Maybe add a docstring.
|
|
||
| # Check if downsampling factor d is defined | ||
| if !isnothing(d) | ||
| k1 = d |
There was a problem hiding this comment.
why k1 = s1 = d? Please add a reference for this choice.
| @@ -66,7 +66,7 @@ end | |||
| @Flux.functor NetworkLoop | |||
|
|
|||
| # 2D Constructor | |||
| function NetworkLoop(n_in, n_hidden, maxiter, Ψ; k1=4, k2=3, p1=0, p2=1, s1=4, s2=1, type="additive", ndims=2) | |||
| function NetworkLoop(n_in, n_hidden, maxiter, Ψ; n_hiddens=nothing, ds=nothing, k1=4, k2=3, p1=0, p2=1, s1=4, s2=1, type="additive", ndims=2) | |||
There was a problem hiding this comment.
Similar variable names n_hidden and and n_hiddens. Maybe add a docstring explaining these different inputs and think of a more clear variable name.
| @@ -0,0 +1,133 @@ | |||
| # Invertible network layer from Putzky and Welling (2019): https://arxiv.org/abs/1911.10914 | |||
| # Author: Philipp Witte, [email protected] | |||
There was a problem hiding this comment.
Are these information correct?
| @@ -64,6 +64,7 @@ include("layers/invertible_layer_hint.jl") | |||
| # Invertible network architectures | |||
| include("networks/invertible_network_hint_multiscale.jl") | |||
| include("networks/invertible_network_irim.jl") # i-RIM: Putzky and Welling (2019) | |||
| include("networks/invertible_network_unet.jl") # single loop i-RIM: Putzky and Welling (2019) | |||
There was a problem hiding this comment.
| C = Conv1x1(n_in) | ||
| RB = ResidualBlock(n_in÷2, n_hidden; k1=k1, k2=k2, p1=p1, p2=p2, s1=s1, s2=s2, ndims=ndims) | ||
| if length(n_hiddens) != length(ds) | ||
| throw("Number of downsampling factors in ds must be the same defined hidden channels in n_hidden") |
There was a problem hiding this comment.
n_hiddens and ds must have equal length but the provided lengths are .... and ....
| else | ||
| ΔX, Δθ_C2 = L.C.inverse((ΔX_, X_); set_grad=set_grad)[1:2] | ||
| # Initialize layer parameters | ||
| !set_grad && (p1 = Array{Parameter, 1}(undef, 0)) |
There was a problem hiding this comment.
why do we need to initialize?
| end | ||
|
|
||
| set_grad ? (return ΔX, X) : (return ΔX, cat(Δθ_C1+Δθ_C2, Δθ_RB; dims=1), X) | ||
| set_grad ? (return ΔY, Y) : (ΔY, cat(p1, p2; dims=1), Y) |
There was a problem hiding this comment.
Would be good to keep the naming convention the same, i.e., this should return ΔX.
| end | ||
|
|
||
| @test isapprox(err3[end] / (err3[1]/2^(maxiter-1)), 1f0; atol=1f0) | ||
| @test isapprox(err4[end] / (err4[1]/4^(maxiter-1)), 1f0; atol=1f0) |
There was a problem hiding this comment.
Missing jacobian tests.
add very simple invertible unet. No gradient as irim, just an input for clear comparison with traditional unets.
take away indexing
-Changing irim block to generate multiple RBs with different dilations and different hidden channels. This is the proper welling implementation.
-This should break some examples. If we are okay with this new block I will go and change all examples to run properly. It is as easy as changing NetworkIRIM(n_in, n_hidden ....)->NetworkIRIM(n_in, [n_hidden], [4];) thus defining a single unet layer with conv dilation 4 which is the current IRIM implementation.
-Add new network invertible unet this is basically a single loop unrolled iteration of irim. Name comes from welling code.
-Directly takes in a precomputed gradient.
-meant for inverse problems where your operator is too expensive to use online.
-Doesnt have aggressive memory savings such as inplace conv1x1 yet but should work well with moderately sized 3D. Will be testing this.