FluxML/Flux.jl v0.16.12

on GitHub

Flux v0.16.12

Diff since v0.16.11

Merged pull requests:

• fix stack overflow when passing empty struct (#2670) (@hardikxk)
• BatchNorm: error on a single value per channel in training (fixes #1992) (#2716) (@CarloLucibello)
• Non-zero padding modes, :same, and even-filter same-padding fix (#2717) (@CarloLucibello)
• Fix Dense with a zero-sized batch dimension (#2407) (#2718) (@CarloLucibello)
• Add :named show option to @layer macro (#2719) (@hardikxk)
• Bump JuliaRegistries/TagBot from 1.25.10 to 1.25.11 (#2720) (@dependabot[bot])
• add missing pad_mode argument (#2724) (@Alexander-Barth)
• return send/receive buffers in __reduce! and __reduceall! for non-GPU aware MPI implementations (#2725) (@Alexander-Barth)

Closed issues:

• ConvRNN Cell and ConvLSTM Cell (#91)
• Add tests for FP16 support (#1574)
• A better Flux overview for non-Flux users (#1576)
• WebDataset.jl, a linearly scalable data loader based on iterable datasets (#2037)
• Allow BatchNorm training on CUDA with track_stats=false (#1606)
• Remove onehotbatch? (#1659)
• Allow user to change the channel axis for BatchNorm function and the likes (#1664)
• Segmentation faults in Flux Losses with RGB images (#1676)
• Benchmark optimizer latency on GPU (#1699)
• Make it possible to run subset of tests (#1700)
• Model's gradient into loss function (#1753)
• Figure out consistent code style (#1765)
• Unable to differentiate broadcasted function on sparse array (#1768)
• Add an "Extending Flux" page to the docs (#1811)
• Overloading Flux.batch for custom eltypes (#1821)
• Use Tuple instead of Vector{Any} in the definition of Optimiser (#1823)
• Roadmap (#1829)
• More robust testing for AlphaDropout (#1851)
• Moving tutorials and ecosystem section from homepage to the docs? (#1858)
• Initialising weights outside of layer declarations (#1879)
• Follow Up to the Addition of Siamese Loss function PR (#1894)
• Conv with circular padding (#1917)
• @functor is a bad name and nobody knows to use it (#1946)
• Flux.normalize vs Flux.normalise (#1952)
• Flux.params for primitive types (#1991)
• Batchnorm's variance becomes NaN when minibatch contains just one sample (#1992)
• Enable optimisations with Chain (#2004)
• Dropout layer not working with CUDA (#2019)
• load flat parameters without mutation or restructure (#2026)
• count_params function? (#2043)
• Derivative of output neurons wrt to input neurons in loss (#2050)
• reducing invalidations (#2080)
• [docs] Highlight update! API more to attract DL researchers (#2104)
• Documentation overhaul roadmap (#2105)
• Runtime dispatch in train! (#2113)
• ForwardDiff + destructure is different from Zygote, on a model with BatchNorm (#2122)
• [BUG] conv_im2col! Scalar Indexing in CUDA (conv_im2col) (#2135)
• BatchNorm is not twice-differentiable (#2154)
• Feature Request: Load pre-trained weights (.pth files etc.) into a flux model (#2164)
• On some recent Flux experiences (#2171)
• wrong gradient for BatchNorm on gpu in testmode (#2179)
• 2020-09-15-deep-learning-flux tutorial crashes (#2183)
• Random Fourier Features (#2207)
• Error when running Flux.train! under VS Code debugger (#2244)
• Group equivariant CNN (G-CNN) (#2246)
• Add docs note about saving/loading models with anonymous functions (#2263)
• Mixed precision training (#2291)
• Create how-to guide for custom layers (#2304)
• Intel Arc GPU support. (#2345)
• Convolutional network slower than tensorflow on CPU (#2350)
• Hard error using dice loss (#2383)
• Zero-sized arrays cannot be applied to Dense layers. (#2407)
• SamePad() for even sized filters. (#2431)
• Enable github Discussions (#2451)
• is Flux.huber_loss type-unstable ? (#2459)
• ConvTranspose with padding on cpu throws exception (#2465)
• DifferentiationInterface testing (#2469)
• any way to do causal pad in Flux ? (#2505)
• Weights shape not validated against kernel, channels (#2506)
• missing codecov token in buildkite (#2517)
• use DocumenterVitepress.jl for documentation (#2519)
• display fieldnames in show of a container (#2543)
• infinite recursion when applying f32 on empty layer (#2571)
• xLSTM: Extended Long Short-Term Memory Layer (#2582)
• Some models saved under Flux v0.14 do not load on v0.15+ (#2584)
• Documentation's example using Enzyme does not work on GPU (#2621)
• DepthwiseConv significantly slower than Conv during training (#2626)
• Failure with distributed data parallel training (DDP) with non-CUDA/ROCM aware MPI (#2722)
• pad_mode argument for conv layers missing when tranferring to AMDGPU (#2723)