Chess input representation v2.8

[QueensGambit]

This PR adds 5 material count planes to the input representation of 2.7 (#122).

Feature	Planes	Description
P1 piece	6	(pieces are ordered: PAWN, KNIGHT, BISHOP, ROOK, QUEEN, KING)
P2 piece	6	(pieces are ordered: PAWN, KNIGHT, BISHOP, ROOK, QUEEN, KING)
En-passant square	1	(Binary map indicating the square where en-passant capture is possible)
P1 castling	2	(One if castling is possible, else zero)
P2 castling	2	(One if castling is possible, else zero)
Last Moves	2	Last move played
is960	1	indicate if 960 is active
P1 pieces	1	A grouped mask of all P1 pieces
P2 pieces	1	A grouped mask of all P2 pieces
Chessboard	1	A chess board pattern
P1 Material Diff	5	(pieces are ordered: PAWN, KNIGHT, BISHOP, ROOK, QUEEN), normalized with 8, + means positive, - means negative
Opposite Color Bishops	1	Indicates if they are only two bishops and the bishops are opposite color
Checkers	1	Indicates all pieces giving check
Checking Moves	2	Indicates all checking moves (from sq, to sq)
P1 Mobility	1	Gives the number of legal moves for the active player
P1 Material Count	5	(pieces are ordered: PAWN, KNIGHT, BISHOP, ROOK, QUEEN), normalized with 8

Total Planes: 38

Constants

NORMALIZE_MOBILITY = 64
NORMALIZE_PIECE_NUMBER = 8
# These constant describe the starting channel for the corresponding info
CHANNEL_PIECES = 0
CHANNEL_EN_PASSANT = 12
CHANNEL_CASTLING = 13
CHANNEL_LAST_MOVES = 17
CHANNEL_IS_960 = 19
CHANNEL_PIECE_MASK = 20
CHANNEL_CHECKERBOARD = 22
CHANNEL_MATERIAL = 23
CHANNEL_OPP_BISHOPS = 28
CHANNEL_CHECKERS = 29
CHANNEL_CHECK_MOVES = 30
CHANNEL_MOBILITY = 32
CHANNEL_MATERIAL_COUNT = 33

Adding material count features appears to both help the value and policy loss.

Model Info

kernels = [3] * 7

se_types = [None] * len(kernels)
se_types[5] = "eca_se"

symbol = rise_mobile_v3_symbol(channels=256, channels_operating_init=224, channel_expansion=32, act_type='relu',
                               channels_value_head=8, value_fc_size=256,
                               channels_policy_head=NB_POLICY_MAP_CHANNELS,
                               grad_scale_value=tc.val_loss_factor, grad_scale_policy=tc.policy_loss_factor, 
                               dropout_rate=tc.dropout_rate, select_policy_from_plane=True,
                               kernels=kernels, se_types=se_types, use_avg_features=False)

Performance

Representation 2.7

print(val_metric_values_best)
{'value_loss': 0.5191286853872813, 'policy_loss': 1.9544962583444057, 'value_acc_sign': 0.6256210535194651, 'policy_acc': 0.4121719751602564, 'loss': 1.9401425826148344}

metrics = metrics_gluon
evaluate_metrics(metrics, test_data, net, nb_batches=None, sparse_policy_label=True, ctx=ctx,
                 apply_select_policy_from_plane=tc.select_policy_from_plane)

{'loss': 1.948354888682973,
 'value_loss': 0.5401958049762816,
 'policy_loss': 1.9625787178113232,
 'value_acc_sign': 0.6418024494760046,
 'policy_acc': 0.4118828864816599}

Representation 2.8

print(val_metric_values_best)
{'value_loss': 0.5170979887629167, 'policy_loss': 1.9267710447311401, 'value_acc_sign': 0.632907013273931, 'policy_acc': 0.42357772435897434, 'loss': 1.912674314171458}

metrics = metrics_gluon
evaluate_metrics(metrics, test_data, net, nb_batches=None, sparse_policy_label=True, ctx=ctx,
                 apply_select_policy_from_plane=tc.select_policy_from_plane)
{'loss': 1.9227169821895014,
 'value_loss': 0.5378539902823312,
 'policy_loss': 1.936705497259271,
 'value_acc_sign': 0.6454561619801322,
 'policy_acc': 0.42199873111356534}