From lgatr v1 to v2

The lgatr v2 release renames parts of the interface, mostly to make names consistent between the LGATr and LGATrSlim architectures. The concrete changes are:

• ConditionalLGATr and ConditionalLGATrSlim __init__ arguments:

  • condition_mv_channels → mv_channels_cond

  • condition_v_channels → v_channels_cond

  • condition_s_channels → s_channels_cond

  and forward arguments:

  • multivectors_condition → multivectors_cond

  • vectors_condition → vectors_cond

  • scalars_condition → scalars_cond

• MLPConfig:

  • activation → nonlinearity

  • increase_hidden_channels → mlp_ratio

  • num_hidden_layers → num_layers_mlp (num_layers_mlp counts all layers instead of only the hidden ones, so explicit values have to be increased by one if the parameter is specified)

• SelfAttentionConfig and CrossAttentionConfig:

  • increase_hidden_channels → attn_ratio.

• The global gatr_config object is replaced by PrimitivesConfig, which is passed to each network as primitives. This way several networks with different settings can coexist. The flags are renamed:

  • use_fully_connected_subgroup → subgroup

  • use_bivector → bivector

  • use_geometric_product → geometric_product

• The compile_mode and compile_dynamic arguments are replaced by a single compile_kwargs dict that is passed on to torch.compile(). Note that dynamic is not enabled by default anymore, use compile_kwargs={"dynamic": True} to recover the v1 behavior.