Gabriel Previato

Weights and positive weights for labels in Flux losses

5

### Motivation and description Hello folks, I searched in Flux issues and documentations but couldn't find anything about having (or a reason for not having) weights for labels in the...

Using AbstractFloat instead of Float64 in warning check for slow conv

5

### PR Checklist - [X] Tests are added - [X] Documentation, if applicable https://github.com/FluxML/NNlib.jl/pull/383 silence some warnings that are not necessarily useful (as discussed in the PR), but keep the...

Fix rounding problems in _mean function

4

As stated in https://github.com/JuliaGPU/CUDA.jl/issues/1773, with the current [_mean ](https://github.com/JuliaGPU/GPUArrays.jl/blob/b2f396dabdab907866b885cd7abb647214cddace/src/host/statistics.jl#L29) function, for bigger arrays you get some rounding problems. ```julia julia> A = CUDA.ones((640, 640, 32, 1)) 640×640×32×1 CuArray{Float32, 4, CUDA.Mem.DeviceBuffer}:...

Mean overflows when using smaller types (e.g. Float16)

9

Since the [current mean](https://github.com/JuliaStats/Statistics.jl/blob/e9ac70b760dcf87b77affe6c068548a3325d6e2b/src/Statistics.jl#L169) implementation is calculated by summing all elements and then dividing it by the total number of elements, when working with smaller types (such as Float16) it's...

On this year there were some upgrades in EfficientNet architecture, named as [EfficientNetV2](https://arxiv.org/abs/2104.00298). This PR still need some work, and I'm doing it "on the fly" while I'm travelling, so...