Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[Feature] Logit focal loss #2138

Draft
wants to merge 4 commits into
base: master
Choose a base branch
from
Draft

[Feature] Logit focal loss #2138

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
dabd4f5
implement logit_focal_loss
josephsdavid Dec 20, 2022
6a91e0d
doctest maybe?
josephsdavid Dec 20, 2022
a86eb7b
fix typo
josephsdavid Dec 20, 2022
7698243
Update src/losses/functions.jl
josephsdavid Dec 21, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
48 changes: 44 additions & 4 deletions src/losses/functions.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -603,14 +603,54 @@ function focal_loss(ŷ, y; dims=1, agg=mean, γ=2, ϵ=epseltype(ŷ))
agg(sum(@. -y * (1 - ŷ)^γ * log(ŷ); dims=dims))
end

"""
logit_focal_loss(ŷ, y; dims=1, agg=mean, γ=2, ϵ=eps(ŷ))

Return the [focal_loss](https://arxiv.org/pdf/1708.02002.pdf)
which can be used in classification tasks with highly imbalanced classes.
It down-weights well-classified examples and focuses on hard examples.
The input, 'ŷ', is expected to be normalized (i.e. [softmax](@ref Softmax) output).

The modulating factor, `γ`, controls the down-weighting strength.
For `γ == 0`, the loss is mathematically equivalent to [`Losses.crossentropy`](@ref).

# Example
```jldoctest
julia> y = [1 0 0 0 1
0 1 0 1 0
0 0 1 0 0]
3×5 Matrix{Int64}:
1 0 0 0 1
0 1 0 1 0
0 0 1 0 0

julia> ŷ = reshape(-7:7, 3, 5) .* 1f0
3×5 Matrix{Float32}:
0.0900306 0.0900306 0.0900306 0.0900306 0.0900306
0.244728 0.244728 0.244728 0.244728 0.244728
0.665241 0.665241 0.665241 0.665241 0.665241

julia> Flux.logit_focal_loss(ŷ, y) ≈ 1.1277571935622628
true
Copy link
Member

Choose a reason for hiding this comment

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

This example output doesn't match what's written.

More importantly, the example is an opportunity to show exactly how this relates to focal_loss, i.e. where the softmax goes. And perhaps (if you can think of a compact & clear way) the relation to crossentropy (or rather logitcrossentropy?) too.

Copy link
Author

Choose a reason for hiding this comment

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

Yeah i still need to work through these tests, did not realize about the docstring tests until after already putting tests elsewhere :) Can do !

```

See also: [`Losses.focal_loss`](@ref)

"""
function logit_focal_loss(ŷ, y; γ=2.0f0, agg=mean, dims=1, ϵ=epseltype(ŷ))
Copy link
Member

@mcabbott mcabbott Dec 20, 2022 •
edited
Loading

Choose a reason for hiding this comment

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

Some have crept in & need fixing, but there should not be greek-letter keywords. These can be gamma and eps?

Also, as written, γ=1.5 will cause Float32 input to be promoted to Float64. Can you avoid this somehow? Perhaps there should be a line like γ = gamma isa Integer ? gamma : convert(eltype(logpt), gamma). (Integer powers are faster.)

Copy link
Author

Choose a reason for hiding this comment

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

Oh nice idea, can do!

_check_sizes(ŷ, y)
logpt = logsoftmax(ŷ; dims)
agg(sum(@. -y * (1 - exp(logpt + ϵ))^γ * (logpt + ϵ); dims))
end

"""
siamese_contrastive_loss(ŷ, y; margin = 1, agg = mean)

Return the [contrastive loss](http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf)
which can be useful for training Siamese Networks. It is given by
agg(@. (1 - y) * ŷ^2 + y * max(0, margin - ŷ)^2)

agg(@. (1 - y) * ŷ^2 + y * max(0, margin - ŷ)^2)

Specify `margin` to set the baseline for distance at which pairs are dissimilar.

# Example
Expand Down
17 changes: 15 additions & 2 deletions test/losses.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,7 @@ const ALL_LOSSES = [Flux.Losses.mse, Flux.Losses.mae, Flux.Losses.msle,
Flux.Losses.dice_coeff_loss,
Flux.Losses.poisson_loss,
Flux.Losses.hinge_loss, Flux.Losses.squared_hinge_loss,
Flux.Losses.binary_focal_loss, Flux.Losses.focal_loss, Flux.Losses.siamese_contrastive_loss]
Flux.Losses.binary_focal_loss, Flux.Losses.focal_loss, Flux.Losses.siamese_contrastive_loss, Flux.Losses.logit_focal_loss]


@testset "xlogx & xlogy" begin
Expand Down Expand Up @@ -210,7 +210,20 @@ end
@test Flux.focal_loss(ŷ1, y1) ≈ 0.45990566879720157
@test Flux.focal_loss(ŷ, y; γ=0.0) ≈ Flux.crossentropy(ŷ, y)
end


@testset "logit_focal_loss" begin
rng = Random.seed!(Random.default_rng(), 5)
y = rand(rng, Float32, 6, 40, 2)
yhat = rand(rng, Float32, 6, 40, 2)

@test logit_focal_loss(yhat, y; γ=0) ≈
Flux.Losses.logitcrossentropy(yhat, y)


@test logit_focal_loss(yhat, y; γ=2) ==
Flux.Losses.focal_loss(Flux.softmax(yhat; dims=1), y; γ=2)
end

@testset "siamese_contrastive_loss" begin
y = [1 0
0 0
Expand Down