ytm._core.backends.buffers.FitBuffers

class ytm._core.backends.buffers.FitBuffers

Bases: AutoPtrs

Arrays the fit kernels read and write, kept on the device for a whole epoch.

Work arrays (clause outputs, votes, feedback lists) are made once in __post_init__. start_epoch takes the epoch’s inputs: per-sample fields (per_sample_fields) stay on the host, the rest are moved to the device. load_chunk(start, end) moves one slice of the per-sample fields to the device, so an epoch can run in chunks that fit in memory. end_epoch drops the epoch’s inputs.

With make_ptrs, every numpy array set as an attribute also gets a ctypes pointer p_<name> for the cpu kernels. It is None when the array is None.

Parameters:

xpmodule

Array module, numpy or cupy.

cfgBaseTMConfig

Model config, for sizes and dtypes.

make_ptrsbool, default=True

Make the p_<name> pointers. Only the cpu backend needs them.

Attributes:

epoch_fieldsdict

The keyword arguments of the last start_epoch, on the host.

per_sample_fieldstuple of str

Fields with one row per sample, needed for loading chunk by chunk.

Nint

Total number of samples in the epoch.

chunk_start, chunk_endint

Sample range [chunk_start, chunk_end) currently loaded.

Xarray of shape (chunk, *dim)

Inputs of the loaded chunk, cast for the kernels.

Yarray of shape (chunk, n_classes)

Targets of the loaded chunk.

rng_keysarray of shape (chunk,)

RNG key of each sample in the loaded chunk.

clause_drop_maskarray of shape (total_clauses,)

Clauses dropped for the whole epoch (Drop Clause).

lit_drop_maskarray of shape (n_literals,)

Literals to skip.

feat_drop_bitsarray of shape (n_bin_feat_chunks,), uint64

All binary features only: bit fid set when feature fid is skipped. Shape (1,) otherwise.

x_bitsarray of shape (n_patches, n_bin_feat_chunks), uint64

All binary features only: bit packed patches of the current sample. Shape (1,) otherwise.

class_activearray of shape (n_classes,), uint8

1 when the class gives feedback on the current sample.

clause_outputarray of shape (total_clauses,)

1 when the clause fired on the current sample.

selected_pidsarray of shape (total_clauses,)

Selected patch ids for the active clauses. Shape (1,) with one patch.

votesarray of shape (n_classes,)

Weighted vote sum for the current sample.

t1b_n, t1a_t2_narray of shape (1,), uint32

Number of valid entries in the T1b list and in the T1a/T2 list.

t1b_cid, t1a_t2_cidarray of shape (total_clauses,), uint32

Clauses with T1b feedback, and clauses with T1a or T2 feedback, on the current sample.

t1b_cntarray of shape (total_clauses,)

T1b count over classes, parallel to t1b_cid.

t1a_cnt, t2_cntarray of shape (total_clauses,)

T1a and T2 counts over classes, parallel to t1a_t2_cid.

end_epoch() → None

Drop the epoch’s inputs and reset the chunk range.

load_chunk(start: int, end: int) → None

Move samples [start, end) of every per-sample field to the device.

start_epoch(**fields: Any) → None

Take the epoch’s inputs as keyword arguments, X included.

Per-sample fields are kept on the host until load_chunk. The others are set now, numpy arrays moved to the device.