File size: 8,830 Bytes
ee7f83c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 | import torch
import torch.nn as nn
import torch.nn.functional as F
class BahdanauAttention(nn.Module):
"""
Bahdanau (additive) attention from Bahdanau et al. (2015).
At each decoder step, the attention module compares the current decoder hidden
state against every encoder output position and produces a context vector —
a weighted sum of encoder outputs — that tells the decoder where to look in the
source sentence. This is especially useful for translation because word order
between Filipino and the regional languages isn't always 1-to-1.
The energy score for each source position i is computed as:
score(decoder_hidden, encoder_output_i)
= score_vector( tanh( W_enc * encoder_output_i
+ W_dec * decoder_hidden ) )
Scores are then passed through softmax to get normalized attention weights.
"""
def __init__(self, encoder_hidden_size: int, decoder_hidden_size: int) -> None:
super().__init__()
# Projects encoder outputs (bidirectional, so * 2) into the attention space
self.encoder_projection = nn.Linear(
encoder_hidden_size * 2, decoder_hidden_size
)
# Projects the decoder hidden state into the same attention space
self.decoder_projection = nn.Linear(decoder_hidden_size, decoder_hidden_size)
# Collapses the attention space down to a single scalar score per source position
self.score_vector = nn.Linear(decoder_hidden_size, 1, bias=False)
def forward(
self,
decoder_hidden: torch.Tensor,
encoder_outputs: torch.Tensor,
source_padding_mask: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
# decoder_hidden: (batch_size, decoder_hidden_size)
# encoder_outputs: (batch_size, source_len, encoder_hidden_size * 2)
# Expand the decoder hidden state so it can be broadcast across all source positions
decoder_hidden_expanded = decoder_hidden.unsqueeze(1)
# decoder_hidden_expanded: (batch_size, 1, decoder_hidden_size)
# Add the two projections and apply tanh — this is the "energy" function
energy = torch.tanh(
self.encoder_projection(encoder_outputs)
+ self.decoder_projection(decoder_hidden_expanded)
)
# energy: (batch_size, source_len, decoder_hidden_size)
# Reduce to one score per source position
scores = self.score_vector(energy).squeeze(2)
# scores: (batch_size, source_len)
# Padding positions should never get attention, so we set their scores to -inf
# before softmax so they come out as zero weight
if source_padding_mask is not None:
scores = scores.masked_fill(source_padding_mask == 0, float("-inf"))
# Softmax turns the raw scores into a proper probability distribution
attention_weights = F.softmax(scores, dim=1)
# attention_weights: (batch_size, source_len)
# Build the context vector as a weighted sum of the encoder outputs
# bmm performs batch matrix multiplication: (batch, 1, source_len) x (batch, source_len, hidden*2)
context_vector = torch.bmm(attention_weights.unsqueeze(1), encoder_outputs)
# context_vector: (batch_size, 1, encoder_hidden_size * 2)
return context_vector.squeeze(1), attention_weights
class Decoder(nn.Module):
"""
Autoregressive GRU/LSTM decoder with Bahdanau attention.
At each decoding step the decoder:
1. Embeds the previous output token (or <sos> at the first step)
2. Attends over all encoder outputs to build a context vector
3. Concatenates the embedding and context vector, then runs one RNN step
4. Projects the RNN output to a distribution over the target vocabulary
The final prediction combines the RNN output, context vector, and embedding
(a technique called "input feeding" / deep output) so the model has direct
access to what it just generated and what it attended to.
"""
def __init__(
self,
target_vocab_size: int,
embedding_dim: int,
encoder_hidden_size: int,
decoder_hidden_size: int,
dropout: float,
rnn_type: str = "gru",
) -> None:
super().__init__()
self.rnn_type = rnn_type.lower()
self.encoder_hidden_size = encoder_hidden_size
# Embedding for the target language (Waray / Kapampangan / Hiligaynon)
self.embedding = nn.Embedding(target_vocab_size, embedding_dim, padding_idx=0)
self.dropout = nn.Dropout(dropout)
self.attention = BahdanauAttention(encoder_hidden_size, decoder_hidden_size)
# The RNN input at each step is the concatenation of:
# - the embedded previous token (embedding_dim)
# - the attention context vector (encoder_hidden_size * 2)
rnn_input_dim = embedding_dim + encoder_hidden_size * 2
rnn_options = {"rnn": nn.RNN, "gru": nn.GRU, "lstm": nn.LSTM}
if self.rnn_type not in rnn_options:
raise ValueError(
f"rnn_type must be one of {list(rnn_options.keys())}, got '{rnn_type}'"
)
self.rnn = rnn_options[self.rnn_type](
input_size=rnn_input_dim,
hidden_size=decoder_hidden_size,
num_layers=1,
batch_first=True,
)
# The output projection combines three sources of information for a richer prediction:
# - the new RNN hidden state (decoder_hidden_size)
# - the attention context vector (encoder_hidden_size * 2)
# - the embedded input token (embedding_dim)
self.output_projection = nn.Linear(
decoder_hidden_size + encoder_hidden_size * 2 + embedding_dim,
target_vocab_size,
)
def forward_step(
self,
input_token: torch.Tensor,
decoder_hidden: torch.Tensor,
decoder_cell: torch.Tensor | None,
encoder_outputs: torch.Tensor,
source_padding_mask: torch.Tensor | None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None, torch.Tensor]:
"""
Performs a single decoder step and returns the vocabulary logits for that step.
Args:
input_token: (batch_size,) — token id fed as input at this step
decoder_hidden: (1, batch_size, decoder_hidden_size) — current hidden state
decoder_cell: (1, batch_size, decoder_hidden_size) — current cell (LSTM only)
encoder_outputs: (batch_size, source_len, encoder_hidden_size * 2)
source_padding_mask:(batch_size, source_len) — 1 for real tokens, 0 for padding
Returns:
vocab_logits: (batch_size, target_vocab_size) — unnormalized scores
decoder_hidden: (1, batch_size, decoder_hidden_size) — updated hidden state
decoder_cell: (1, batch_size, decoder_hidden_size) or None
attention_weights: (batch_size, source_len) — where the model looked
"""
embedded = self.dropout(self.embedding(input_token.unsqueeze(1)))
# embedded: (batch_size, 1, embedding_dim)
# Squeeze out the leading num_layers dimension (always 1 here) for the attention call
context_vector, attention_weights = self.attention(
decoder_hidden.squeeze(0), encoder_outputs, source_padding_mask
)
# context_vector: (batch_size, encoder_hidden_size * 2)
# attention_weights: (batch_size, source_len)
# Concatenate the embedding and context vector, then step the RNN
rnn_input = torch.cat([embedded, context_vector.unsqueeze(1)], dim=2)
# rnn_input: (batch_size, 1, embedding_dim + encoder_hidden_size * 2)
if self.rnn_type == "lstm":
rnn_output, (decoder_hidden, decoder_cell) = self.rnn(
rnn_input, (decoder_hidden, decoder_cell)
)
else:
rnn_output, decoder_hidden = self.rnn(rnn_input, decoder_hidden)
decoder_cell = None
# rnn_output: (batch_size, 1, decoder_hidden_size)
# Combine rnn output, context, and embedding to form the final prediction input
# This gives the model a direct view of what it attended to and its last token
prediction_input = torch.cat(
[rnn_output.squeeze(1), context_vector, embedded.squeeze(1)], dim=1
)
# prediction_input: (batch_size, decoder_hidden_size + encoder_hidden_size * 2 + embedding_dim)
vocab_logits = self.output_projection(prediction_input)
# vocab_logits: (batch_size, target_vocab_size)
return vocab_logits, decoder_hidden, decoder_cell, attention_weights
|