ytm._core.config.BaseTMConfig#

class ytm._core.config.BaseTMConfig#

Bases: object

Settings shared by all TM models.

See also

ytm._core.initializer: Available TA initializers.

n_clauses: int#: Number of clauses per class or clause bank, at least 1. When coalesced, all classes share one clause bank, so total clauses = n_clauses. When not coalesced, total clauses = n_clauses * n_classes.

s: float#: Specificity, at least 1.0.

dim: int | tuple[int, ...]#: Input shape: D, (H, W) or (H, W, C). Missing trailing sizes are set to 1.

n_classes: int#: Number of outputs (classes or labels).

feat_mins: int | Sequence[int] | ndarray = 0#

Inclusive integer range of each feature, together with feat_maxs. For binary features, use 0 and 1 (the defaults). For thermometer encoded features, a feature uses feat_maxs - feat_mins thermometer bits.

A single int sets the same min or max for all features.
Without convolution, an array with one entry per feature sets min and max per feature.
With convolution, pass an int, or a tuple of length C for min and max per channel.

feat_maxs: int | Sequence[int] | ndarray = 1#: Upper end of the inclusive integer range of each feature. See feat_mins.

patch_dim: tuple[int, int] | None = None#: Convolution window (h, w). None means the whole image (no convolution). An entry <= 0 or larger than the image uses the full size in that axis.

stride: tuple[int, int] = (1, 1)#: Window step in y and x.

track_patch_weights: bool = True#: Count, per clause and patch, which patches the clause picks. Needed by wic on convolutional models.

coalesced: bool = True#: One clause bank shared by all classes with per-class weights. False gives each class its own bank.

negative_clauses: bool = True#: Start half the clauses with negative polarity (voting against a class).

negated_literals: bool = True#: Append the negation of each literal. Doubles the literal count.

position_literals: bool = True#: Add thermometer-coded patch position literals. Only used when there is more than one patch.

max_includes: int | None = None#: Literal budget of a clause. None, <= 0 or more than the literal count means no limit.

n_states: int = 256#: TA states per literal, at least 2.

include_state: int | None = None#: Lowest state that counts as include. None or -1 means n_states // 2.

ta_init: Literal['random', 'middle', 'random_include'] | str | int | TAInit = 'middle'#

Initial TA states.

"middle": just below include.
"random": uniform over all states.
"random_include": uniform over the two states around the include boundary.
"random:N": uniform in a band of N states on each side of the boundary.
An int sets every TA to that state.
A TAInit (TA initializer) defines how to initialize the states or clauses.

weighted: bool = True#: Learn clause weights. False keeps the initial weights.

max_weight: float = 1073741824.0#: Largest weight magnitude.

allow_polarity_change: bool = True#: Let a weight cross zero and flip the clause’s polarity.

skip_t1a_fb: bool = False#: Turn off Type Ia feedback.

skip_t1b_fb: bool = False#: Turn off Type Ib feedback.

skip_t2_fb: bool = False#: Turn off Type II feedback.

boost_tp_inc: bool = True#: Type Ia increment steps for true literals always happen, instead of with probability 1 - 1/s.

boost_tp_dec: bool = False#: Type Ia decrement steps always happen, instead of with probability 1/s.

seed: int | None = None#: RNG seed. None picks a random seed.

prepare_X(X: ndarray) → ndarray#: Check X against the feature bounds and shift it to [0, therm_bits], shaped (N, *self._dim).

prepare_Y(Y: ndarray) → ndarray#: Check Y, and convert any type of provided Y to multihot/onehot.