|
import torch |
|
import torch.nn as nn |
|
import torch.nn.functional as F |
|
import math |
|
|
|
class PositionalEncoding(nn.Module): |
|
def __init__(self, d_model, max_len=5000): |
|
super(PositionalEncoding, self).__init__() |
|
pe = torch.zeros(max_len, d_model) |
|
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1) |
|
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)) |
|
pe[:, 0::2] = torch.sin(position * div_term) |
|
pe[:, 1::2] = torch.cos(position * div_term) |
|
pe = pe.unsqueeze(0).transpose(0, 1) |
|
self.register_buffer('pe', pe) |
|
|
|
def forward(self, x): |
|
x = x + self.pe[:x.size(0), :] |
|
return x |
|
|
|
class StochasticDepth(nn.Module): |
|
def __init__(self, p=0.8): |
|
super().__init__() |
|
self.p = p |
|
|
|
def forward(self, x, residual): |
|
if self.training: |
|
if torch.rand(1).item() < self.p: |
|
return x + residual |
|
else: |
|
return x |
|
else: |
|
return x + self.p * residual |
|
|
|
class AdvancedTransformerLayer(nn.Module): |
|
def __init__(self, d_model, nhead, dropout=0.1, stoch_depth_p=0.8): |
|
super().__init__() |
|
dim_feedforward = 4 * d_model |
|
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout) |
|
self.ff = nn.Sequential( |
|
nn.Linear(d_model, dim_feedforward), |
|
nn.ReLU(), |
|
nn.Linear(dim_feedforward, d_model) |
|
) |
|
self.norm1 = nn.LayerNorm(d_model) |
|
self.norm2 = nn.LayerNorm(d_model) |
|
self.dropout = nn.Dropout(dropout) |
|
self.stoch_depth = StochasticDepth(stoch_depth_p) |
|
|
|
def forward(self, x, src_mask=None, src_key_padding_mask=None): |
|
|
|
norm_x = self.norm1(x) |
|
|
|
|
|
if src_key_padding_mask is not None: |
|
src_key_padding_mask = src_key_padding_mask.float().masked_fill( |
|
src_key_padding_mask, float('-inf')).masked_fill(~src_key_padding_mask, float(0.0)) |
|
|
|
attn_output, _ = self.self_attn(norm_x, norm_x, norm_x, |
|
attn_mask=src_mask, |
|
key_padding_mask=src_key_padding_mask) |
|
x = self.stoch_depth(x, self.dropout(attn_output)) |
|
|
|
norm_x = self.norm2(x) |
|
ff_output = self.ff(norm_x) |
|
x = self.stoch_depth(x, self.dropout(ff_output)) |
|
return x |
|
|
|
class ChessTransformer(nn.Module): |
|
def __init__(self, num_layers=64, d_model=1024, nhead=8, dropout=0.1, stoch_depth_p=0.9, num_tokens=2066, pad_token_id=2064): |
|
super().__init__() |
|
self.embedding = nn.Embedding(num_tokens, d_model) |
|
self.pos_encoder = PositionalEncoding(d_model) |
|
self.layers = nn.ModuleList([ |
|
AdvancedTransformerLayer(d_model, nhead, dropout, stoch_depth_p) |
|
for _ in range(num_layers) |
|
]) |
|
self.norm = nn.LayerNorm(d_model) |
|
self.output = nn.Linear(d_model, num_tokens) |
|
self.d_model = d_model |
|
self.padding_idx = pad_token_id |
|
|
|
def generate_square_subsequent_mask(self, sz): |
|
mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1) |
|
mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0)) |
|
return mask |
|
|
|
def pad_sequences(self, sequences): |
|
padding_value = self.padding_idx |
|
max_len = max(len(seq) for seq in sequences) |
|
padded_seqs = [seq + [padding_value] * (max_len - len(seq)) for seq in sequences] |
|
return torch.LongTensor(padded_seqs) |
|
|
|
def forward(self, x): |
|
|
|
batch_size, seq_len = x.size() |
|
|
|
|
|
padding_mask = (x == self.padding_idx) |
|
|
|
|
|
causal_mask = self.generate_square_subsequent_mask(seq_len).to(x.device) |
|
|
|
|
|
x = self.embedding(x).transpose(0, 1) * math.sqrt(self.d_model) |
|
x = self.pos_encoder(x) |
|
|
|
|
|
for layer in self.layers: |
|
x = layer(x, src_mask=causal_mask, src_key_padding_mask=padding_mask) |
|
|
|
x = self.norm(x) |
|
output = self.output(x.transpose(0, 1)) |
|
|
|
return output |
|
|
|
def winning_moves_loss(output, ground_truth, win_labels, pad_token_id=2064, start_token_id=2065): |
|
""" |
|
Compute the loss only for the winning moves of white and black. |
|
""" |
|
output = output.cuda() |
|
ground_truth = ground_truth.cuda() |
|
win_labels = win_labels.cuda() |
|
|
|
batch_size, seq_len, num_tokens = output.shape |
|
|
|
|
|
ground_truth_shifted = ground_truth[:, 1:].contiguous() |
|
output_shifted = output[:, :-1, :].contiguous() |
|
|
|
|
|
output_flat = output_shifted.view(-1, num_tokens) |
|
ground_truth_flat = ground_truth_shifted.view(-1) |
|
|
|
|
|
output_log_softmax = F.log_softmax(output_flat, dim=-1) |
|
|
|
|
|
win_labels_expanded = win_labels.unsqueeze(1).repeat(1, seq_len - 1).view(-1) |
|
|
|
|
|
move_indices = torch.arange(seq_len - 1, device=output.device).unsqueeze(0).repeat(batch_size, 1).view(-1) |
|
white_win_mask = (win_labels_expanded == 1) & (move_indices % 2 == 0) |
|
black_win_mask = (win_labels_expanded == 0) & (move_indices % 2 == 1) |
|
|
|
|
|
selected_moves_mask = (white_win_mask | black_win_mask) & (ground_truth_flat != pad_token_id) & (ground_truth_flat != start_token_id) |
|
|
|
|
|
loss = F.nll_loss(output_log_softmax, ground_truth_flat, reduction='none') |
|
|
|
loss = loss * selected_moves_mask.float() |
|
|
|
|
|
selected_moves_count = selected_moves_mask.float().sum() |
|
if selected_moves_count > 0: |
|
loss = loss.sum() / selected_moves_count |
|
else: |
|
loss = loss.sum() |
|
|
|
return loss |
|
|
|
def all_moves_loss(output, ground_truth, pad_token_id=2064, start_token_id=2065): |
|
""" |
|
Compute the loss for all valid moves in the sequence, excluding start and padding tokens. |
|
""" |
|
batch_size, seq_len, num_tokens = output.shape |
|
|
|
output = output.cuda() |
|
ground_truth = ground_truth.cuda() |
|
|
|
|
|
output_shifted = output[:, :-1, :].contiguous() |
|
ground_truth_shifted = ground_truth[:, 1:].contiguous() |
|
|
|
|
|
output_flat = output_shifted.view(-1, num_tokens) |
|
ground_truth_flat = ground_truth_shifted.view(-1) |
|
|
|
|
|
output_log_softmax = F.log_softmax(output_flat, dim=-1) |
|
|
|
|
|
valid_moves_mask = ((ground_truth_flat != pad_token_id) & |
|
(ground_truth_flat != start_token_id)) |
|
|
|
|
|
loss = F.nll_loss(output_log_softmax, ground_truth_flat, reduction='none') |
|
|
|
|
|
loss = loss * valid_moves_mask.float() |
|
|
|
|
|
valid_moves_count = valid_moves_mask.float().sum() |
|
if valid_moves_count > 0: |
|
loss = loss.sum() / valid_moves_count |
|
else: |
|
loss = loss.sum() |
|
|
|
return loss |
|
|
|
def weighted_chess_loss(output, ground_truth, win_labels, winning_weight=1.0, losing_weight=0.1, pad_token_id=2064, start_token_id=2065): |
|
""" |
|
Compute a weighted loss for all moves, with higher weight for winning moves. |
|
""" |
|
output = output.cuda() |
|
ground_truth = ground_truth.cuda() |
|
win_labels = win_labels.cuda() |
|
|
|
batch_size, seq_len, num_tokens = output.shape |
|
|
|
|
|
ground_truth_shifted = ground_truth[:, 1:].contiguous() |
|
output_shifted = output[:, :-1, :].contiguous() |
|
|
|
|
|
output_flat = output_shifted.view(-1, num_tokens) |
|
ground_truth_flat = ground_truth_shifted.view(-1) |
|
|
|
|
|
output_log_softmax = F.log_softmax(output_flat, dim=-1) |
|
|
|
|
|
win_labels_expanded = win_labels.unsqueeze(1).repeat(1, seq_len - 1).view(-1) |
|
|
|
|
|
move_indices = torch.arange(seq_len - 1, device=output.device).unsqueeze(0).repeat(batch_size, 1).view(-1) |
|
white_win_mask = (win_labels_expanded == 1) & (move_indices % 2 == 0) |
|
black_win_mask = (win_labels_expanded == 0) & (move_indices % 2 == 1) |
|
winning_moves_mask = white_win_mask | black_win_mask |
|
|
|
|
|
valid_moves_mask = (ground_truth_flat != pad_token_id) & (ground_truth_flat != start_token_id) |
|
|
|
|
|
loss = F.nll_loss(output_log_softmax, ground_truth_flat, reduction='none') |
|
|
|
|
|
weights = torch.where(winning_moves_mask & valid_moves_mask, winning_weight, losing_weight) |
|
|
|
|
|
weighted_loss = loss * weights * valid_moves_mask.float() |
|
|
|
|
|
valid_moves_count = valid_moves_mask.float().sum() |
|
if valid_moves_count > 0: |
|
avg_loss = weighted_loss.sum() / valid_moves_count |
|
else: |
|
avg_loss = weighted_loss.sum() |
|
|
|
return avg_loss |