dataset Namespace Reference

Classes
class	ChessMoEFactorizedDataset
class	ChessExpertFactorizedDataset

Functions
	decode_feature_row (row, feature_layout, planes_per_type, packed_offsets, packed_branch_planes, num_bypass_planes)

Variables
list	PLANES_PER_GROUP_CONST = [4, 4, 5, 5, 5, 4, 4, 4, 5, 5, 5, 4]
list	PACKED_OFFSETS_CONST
	parser = argparse.ArgumentParser()
	help
	type
	int
	default
	args = parser.parse_args()
	ds = ChessMoEFactorizedDataset(args.data_dir, max_samples=args.max, n_globals=args.n_globals)
	sample = ds[0]

Detailed Description

@file dataset.py
@brief Dataset utilities for training model.py on packed factorized features.

Loads packed factorized records produced by preprocessing in --factorized mode:
  - features.bin: PackedFactorizedInput
  - labels.bin: WDLOutput (3 floats: win, draw, loss)
  - expert_weights.bin: 6 floats [base0-3, survivor, killer]

PackedFactorizedInput C++ layout (alignas(64)):
    uint8_t branches[54][64]   # only used branch planes (4/5 per group)
  uint8_t bypass[12][64]
  float   global[32]

Function Documentation

decode_feature_row()

decode_feature_row	(		row,
			feature_layout,
			planes_per_type,
			packed_offsets,
			packed_branch_planes,
			num_bypass_planes )

Here is the caller graph for this function:

Variable Documentation

args

dataset.args = parser.parse_args()

default

dataset.default

ds

dataset.ds = ChessMoEFactorizedDataset(args.data_dir, max_samples=args.max, n_globals=args.n_globals)

help

dataset.help

int

dataset.int

PACKED_OFFSETS_CONST

list dataset.PACKED_OFFSETS_CONST

Initial value:

=  [
    sum(PLANES_PER_GROUP_CONST[:i]) for i in range(len(PLANES_PER_GROUP_CONST))
]

parser

dataset.parser = argparse.ArgumentParser()

PLANES_PER_GROUP_CONST

list dataset.PLANES_PER_GROUP_CONST = [4, 4, 5, 5, 5, 4, 4, 4, 5, 5, 5, 4]

sample

dataset.sample = ds[0]

type

dataset.type