Configure TabICL2 model architecture
Usage
tab_icl2_config(
embed_dim = 128L,
col_n_block = 3L,
row_n_block = 3L,
icl_n_block = 12L,
col_n_head = 8L,
row_n_head = 8L,
icl_n_head = 8L,
feature_group_size = 3L,
col_n_cls = 128L,
row_n_cls = 4L,
num_quantiles = 999L,
...
)Arguments
- embed_dim
A positive integer. Embedding dimension shared across all transformer stages.
- col_n_block
A positive integer. Number of column-wise transformer blocks.
- row_n_block
A positive integer. Number of row-wise transformer blocks.
- icl_n_block
A positive integer. Number of in-context-learning transformer blocks.
- col_n_head
A positive integer. Number of attention heads for column blocks.
- row_n_head
A positive integer. Number of attention heads for row blocks.
- icl_n_head
A positive integer. Number of attention heads for ICL blocks.
- feature_group_size
A positive integer. Size of feature groups for cyclic repeated grouping.
- col_n_cls
A positive integer. Number of CLS tokens per column, concatenated to produce the row representation fed into the ICL stage.
- row_n_cls
A positive integer. Number of inducing vectors used in the induced self-attention column blocks.
- num_quantiles
A positive integer (>= 5). Number of quantiles predicted by the regression head.
- ...
Additional named arguments stored in the returned config object. Use this for architecture settings not covered by the named parameters above (e.g., arguments added in newer versions of the model).