Changelog
All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
[0.2.1] - 2025-02-28
Added
FlashAttention 2 / SDPA support in
SpatialSelfAttentionviaattn_backendparameter ("auto","sdpa","manual"). Automatically dispatches to the fastest kernel on PyTorch >= 2.0.attn_backendparameter propagated toSpatialTransformerBlockandSpatialViT.AttnBackendtype alias exported fromidx_flow.
[0.2.0] - 2025-02-28
Changed
Split
layers.pyinto separate modules by responsibility:conv,mlp,norm,regularization,attention,vit,pooling,functional. All public names remain importable fromidx_flowdirectly.Split
test_layers.pyinto per-module test files with shared fixtures inconftest.py.Removed
layers.pyandtest_vit_layers.py.
[0.1.0] - 2025-01-15
Initial release.
Added
Layers
SpatialConv– spatial convolution for downsamplingSpatialTransposeConv– transpose convolution for upsamplingSpatialUpsampling– distance-based interpolation upsamplingSpatialPooling– mean, max, and sum poolingSpatialMLP– MLP with spatial gathering, dropout, batch norm, residual connectionsGlobalMLP– channel-wise MLP for pointwise transformationsSpatialBatchNorm,SpatialLayerNorm,SpatialInstanceNorm,SpatialGroupNormSpatialDropout,ChannelDropoutSpatialSelfAttention– multi-head self-attentionSpatialPatchEmbedding,SpatialTransformerBlock,SpatialViTSqueeze,Unsqueeze
Utilities
compute_connection_indiceshp_distanceget_weightsget_healpix_resolution_infoget_initializer,get_activation
Initialization options: Xavier, Kaiming, orthogonal, normal, uniform, zeros.
Activations: ReLU, SELU, Leaky ReLU, GELU, ELU, Tanh, Sigmoid, Swish, Mish.