Skip to content

3. Natural Language Processing

3.1 moe

3.1.1 ExpertModel(input_dim, output_dim, hidden_dim)

Modelo experto individual para MoE

Initializes an expert model with a simple feed-forward network.

Parameters:

Name Type Description Default
input_dim int

Dimensionality of the input data.

 required
output_dim int

Dimensionality of the output data.

 required
hidden_dim int

Dimensionality of the hidden layer.

 required
Source code in src/layers/nlp/moe.py 
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
def __init__(self, input_dim: int, output_dim: int, hidden_dim: int) -> None:
    """
    Initializes an expert model with a simple feed-forward network.

    Args:
        input_dim: Dimensionality of the input data.
        output_dim: Dimensionality of the output data.
        hidden_dim: Dimensionality of the hidden layer.
    """

    super().__init__()

    self.input_dim = input_dim
    self.output_dim = output_dim
    self.hidden_dim = hidden_dim

    self.model = nn.Sequential(
        nn.Linear(in_features=self.input_dim, out_features=self.hidden_dim),
        nn.ReLU(),
        nn.Linear(in_features=self.hidden_dim, out_features=self.output_dim),
    )

3.1.1.1 forward(input_tensor)

Forward pass through the expert model.

Parameters:

Name Type Description Default
input_tensor Tensor

Input tensor to the model.

 required

Returns:

Type Description
Tensor

The model's output tensor.
Source code in src/layers/nlp/moe.py 
34
35
36
37
38
39
40
41
42
43
44
45
def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
    """
    Forward pass through the expert model.

    Args:
        input_tensor: Input tensor to the model.

    Returns:
        The model's output tensor.
    """

    return self.model(input_tensor)

3.1.2 Gating(input_dim, num_experts, dropout_rate=0.2)

Gating mechanism to select experts.

Initializes a gating network for expert selection.

Parameters:

Name Type Description Default
input_dim int

Dimensionality of the input data.

 required
num_experts int

Number of experts to select from.

 required
dropout_rate float

Rate of dropout for regularization.

 0.2
Source code in src/layers/nlp/moe.py 
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
def __init__(
    self, input_dim: int, num_experts: int, dropout_rate: float = 0.2
) -> None:
    """
    Initializes a gating network for expert selection.

    Args:
        input_dim: Dimensionality of the input data.
        num_experts: Number of experts to select from.
        dropout_rate: Rate of dropout for regularization.
    """

    super().__init__()

    self.input_dim = input_dim
    self.num_experts = num_experts
    self.dropout_rate = dropout_rate

    self.model = nn.Sequential(
        nn.Linear(in_features=self.input_dim, out_features=128),
        nn.Dropout(self.dropout_rate),
        nn.LeakyReLU(),
        nn.Linear(in_features=128, out_features=256),
        nn.LeakyReLU(),
        nn.Dropout(self.dropout_rate),
        nn.Linear(in_features=256, out_features=128),
        nn.LeakyReLU(),
        nn.Dropout(self.dropout_rate),
        nn.Linear(in_features=128, out_features=num_experts),
    )

3.1.2.1 forward(input_tensor)

Forward pass through the gating network.

Parameters:

Name Type Description Default
input_tensor Tensor

Input tensor to the network.

 required

Returns:

Type Description
Tensor

Softmax probabilities for expert selection.
Source code in src/layers/nlp/moe.py 
84
85
86
87
88
89
90
91
92
93
94
95
def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
    """
    Forward pass through the gating network.

    Args:
        input_tensor: Input tensor to the network.

    Returns:
        Softmax probabilities for expert selection.
    """

    return F.softmax(self.model(input_tensor), dim=-1)

3.1.3 MoE(trained_experts, input_dim, dropout_rate=0.2)

Mixture of Experts

Initializes a mixture of experts with gating.

Parameters:

Name Type Description Default
trained_experts list[ExpertModel]

List of trained expert models.

 required
input_dim int

Dimensionality of the input data.

 required
dropout_rate float

Rate of dropout in the gating network.

 0.2
Source code in src/layers/nlp/moe.py 
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
def __init__(
    self,
    trained_experts: list[ExpertModel],
    input_dim: int,
    dropout_rate: float = 0.2,
) -> None:
    """
    Initializes a mixture of experts with gating.

    Args:
        trained_experts: List of trained expert models.
        input_dim: Dimensionality of the input data.
        dropout_rate: Rate of dropout in the gating network.
    """

    super().__init__()

    self.experts = nn.ModuleList(trained_experts)
    self.num_experts = len(trained_experts)
    self.input_dim = input_dim
    self.dropout_rate = dropout_rate

    self.gating_layer = Gating(
        input_dim=self.input_dim,
        num_experts=self.num_experts,
        dropout_rate=self.dropout_rate,
    )

3.1.3.1 forward(input_tensor)

Forward pass through the mixture of experts.

Parameters:

Name Type Description Default
input_tensor Tensor

Input tensor to the model.

 required

Returns:

Type Description
Tensor

Weighted sum of expert outputs.
Source code in src/layers/nlp/moe.py 
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
def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
    """
    Forward pass through the mixture of experts.

    Args:
        input_tensor: Input tensor to the model.

    Returns:
        Weighted sum of expert outputs.
    """

    # Obtenemos los pesos del selector
    expert_weights = self.gating_layer(input_tensor)

    # Obtenemos la salida de todos los expertos
    _expert_outputs: list[torch.Tensor] = []
    for expert in self.experts:
        _expert_outputs.append(expert(input_tensor))

    # Stack de salidas [b, output_dim, num_experts]
    expert_outputs = torch.stack(_expert_outputs, dim=-1)

    # [b, num_experts] -> [b, 1, num_experts]
    expert_weights = expert_weights.unsqueeze(1)

    # Suma ponderada de la selección de expertos
    # [b, output_dim, num_experts] * [b, 1, num_experts]
    return torch.sum(expert_outputs * expert_weights, dim=-1)

3.2 transformer

3.2.1 DecoderBlock(d_model, d_ff, h, dropout_rate)

Decoder block with masked attention, cross-attention, and feed-forward layers.

Initializes decoder block.

Parameters:

Name Type Description Default
d_model int

Dimensionality of model embeddings.

 required
d_ff int

Dimensionality of feed-forward layer.

 required
h int

Number of attention heads.

 required
dropout_rate float

Rate of dropout for regularization.

 required
Source code in src/layers/nlp/transformer.py 
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
def __init__(self, d_model: int, d_ff: int, h: int, dropout_rate: float) -> None:
    """
    Initializes decoder block.

    Args:
        d_model: Dimensionality of model embeddings.
        d_ff: Dimensionality of feed-forward layer.
        h: Number of attention heads.
        dropout_rate: Rate of dropout for regularization.
    """

    super().__init__()

    # Parametros
    self.d_model = d_model
    self.d_ff = d_ff
    self.h = h
    self.dropout_rate = dropout_rate

    self.masked_multi_head_attention_layer = MultiHeadAttention(
        d_model=self.d_model, h=self.h, dropout_rate=self.dropout_rate
    )
    self.residual_layer_1 = ResidualConnection(
        features=d_model, dropout_rate=self.dropout_rate
    )
    self.multi_head_attention_layer = MultiHeadAttention(
        d_model=self.d_model, h=self.h, dropout_rate=self.dropout_rate
    )
    self.residual_layer_2 = ResidualConnection(
        features=d_model, dropout_rate=self.dropout_rate
    )
    self.feed_forward_layer = FeedForward(
        d_model=self.d_model, d_ff=self.d_ff, dropout_rate=self.dropout_rate
    )
    self.residual_layer_3 = ResidualConnection(
        features=d_model, dropout_rate=self.dropout_rate
    )

3.2.1.1 forward(decoder_input, encoder_output, src_mask=None, tgt_mask=None)

Forward pass through decoder block.

Parameters:

Name Type Description Default
decoder_input Tensor

Input tensor to the decoder block.

 required
encoder_output Tensor

Output tensor from the encoder.

 required
src_mask Tensor | None

Optional source mask tensor.

 None
tgt_mask Tensor | None

Optional target mask tensor.

 None

Returns:

Type Description
Tensor

Tensor after processing by the decoder block.
Source code in src/layers/nlp/transformer.py 
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
def forward(
    self,
    decoder_input: torch.Tensor,
    encoder_output: torch.Tensor,
    src_mask: torch.Tensor | None = None,  # Máscara para el encoder (padding)
    tgt_mask: torch.Tensor | None = None,  # Máscara causal para el decoder
) -> torch.Tensor:
    """
    Forward pass through decoder block.

    Args:
        decoder_input: Input tensor to the decoder block.
        encoder_output: Output tensor from the encoder.
        src_mask: Optional source mask tensor.
        tgt_mask: Optional target mask tensor.

    Returns:
        Tensor after processing by the decoder block.
    """

    # Utilizamos self-attention, por lo que k, q, v son del mismo vector de entrada
    decoder_input = self.residual_layer_1(
        decoder_input,
        lambda x: self.masked_multi_head_attention_layer(
            k=x, q=x, v=x, mask=tgt_mask
        ),
    )

    # Aquí tenemos que hacer cross-attention, usamos como K, V los encoder
    # y Q del decoder
    decoder_input = self.residual_layer_2(
        decoder_input,
        lambda x: self.multi_head_attention_layer(
            k=encoder_output, q=x, v=encoder_output, mask=src_mask
        ),
    )

    decoder_output = self.residual_layer_3(
        decoder_input, lambda x: self.feed_forward_layer(x)
    )

    return decoder_output

3.2.2 EncoderBlock(d_model, d_ff, h, dropout_rate)

Encoder block with attention and feed-forward layers.

Initializes encoder block.

Parameters:

Name Type Description Default
d_model int

Dimensionality of model embeddings.

 required
d_ff int

Dimensionality of feed-forward layer.

 required
h int

Number of attention heads.

 required
dropout_rate float

Rate of dropout for regularization.

 required
Source code in src/layers/nlp/transformer.py 
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
def __init__(self, d_model: int, d_ff: int, h: int, dropout_rate: float) -> None:
    """
    Initializes encoder block.

    Args:
        d_model: Dimensionality of model embeddings.
        d_ff: Dimensionality of feed-forward layer.
        h: Number of attention heads.
        dropout_rate: Rate of dropout for regularization.
    """

    super().__init__()

    # Parametros
    self.d_model = d_model
    self.d_ff = d_ff
    self.h = h
    self.dropout_rate = dropout_rate

    # Definicion de las capas
    self.multi_head_attention_layer = MultiHeadAttention(
        d_model=self.d_model, h=self.h, dropout_rate=self.dropout_rate
    )
    self.residual_layer_1 = ResidualConnection(
        features=d_model, dropout_rate=self.dropout_rate
    )
    self.feed_forward_layer = FeedForward(
        d_model=self.d_model, d_ff=self.d_ff, dropout_rate=self.dropout_rate
    )
    self.residual_layer_2 = ResidualConnection(
        features=d_model, dropout_rate=self.dropout_rate
    )

3.2.2.1 forward(input_tensor, mask=None)

Forward pass through encoder block.

Parameters:

Name Type Description Default
input_tensor Tensor

Input tensor to the encoder block.

 required
mask Tensor | None

Optional mask tensor.

 None

Returns:

Type Description
Tensor

Tensor after processing by the encoder block.
Source code in src/layers/nlp/transformer.py 
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
def forward(
    self, input_tensor: torch.Tensor, mask: torch.Tensor | None = None
) -> torch.Tensor:
    """
    Forward pass through encoder block.

    Args:
        input_tensor: Input tensor to the encoder block.
        mask: Optional mask tensor.

    Returns:
        Tensor after processing by the encoder block.
    """

    # Utilizamos self-attention, por lo que k, q, v son del mismo vector de entrada
    input_tensor = self.residual_layer_1(
        input_tensor,
        lambda x: self.multi_head_attention_layer(k=x, q=x, v=x, mask=mask),
    )

    # Segunda conexión residual con feed-forward
    input_tensor = self.residual_layer_2(
        input_tensor, lambda x: self.feed_forward_layer(x)
    )

    return input_tensor

3.2.3 FeedForward(d_model, d_ff, dropout_rate)

Feed-forward neural network layer.

Initializes feed-forward network.

Parameters:

Name Type Description Default
d_model int

Dimensionality of model embeddings.

 required
d_ff int

Dimensionality of feed-forward layer.

 required
dropout_rate float

Rate of dropout for regularization.

 required
Source code in src/layers/nlp/transformer.py 
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
def __init__(self, d_model: int, d_ff: int, dropout_rate: float) -> None:
    """
    Initializes feed-forward network.

    Args:
        d_model: Dimensionality of model embeddings.
        d_ff: Dimensionality of feed-forward layer.
        dropout_rate: Rate of dropout for regularization.
    """

    # Constructor de la clase
    super().__init__()

    # Definimos los parámetros de la clase
    self.d_model = d_model
    self.d_ff = d_ff

    # Creamos el modelo secuencial
    self.ffn = nn.Sequential(
        nn.Linear(in_features=self.d_model, out_features=self.d_ff),
        nn.ReLU(),
        nn.Dropout(dropout_rate),
        nn.Linear(in_features=self.d_ff, out_features=self.d_model),
    )

3.2.3.1 forward(input_tensor)

Forward pass through feed-forward network.

Parameters:

Name Type Description Default
input_tensor Tensor

Tensor of input embeddings.

 required

Returns:

Type Description
Tensor

Tensor processed by feed-forward network.
Source code in src/layers/nlp/transformer.py 
202
203
204
205
206
207
208
209
210
211
212
213
214
def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
    """
    Forward pass through feed-forward network.

    Args:
        input_tensor: Tensor of input embeddings.

    Returns:
        Tensor processed by feed-forward network.
    """

    # (B, sequence_length, d_model)
    return self.ffn(input_tensor)

3.2.4 InputEmbedding(d_model, vocab_size)

Embeds input tokens into vectors of dimension d_model.

Initializes input embedding layer.

Parameters:

Name Type Description Default
d_model int

Dimensionality of the embedding vectors.

 required
vocab_size int

Size of the vocabulary.

 required
Source code in src/layers/nlp/transformer.py 
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
def __init__(self, d_model: int, vocab_size: int) -> None:
    """
    Initializes input embedding layer.

    Args:
        d_model: Dimensionality of the embedding vectors.
        vocab_size: Size of the vocabulary.
    """

    # Constructor de la clase
    super().__init__()

    # Definimos los parámetros de la clase
    self.d_model = d_model
    self.vocab_size = vocab_size

    # Utilizamos la capa Embedding de PyTorch que funciona como
    # una tabal lookup that stores embeddings of a fixed dictionary and size.
    # Osea que es un diccionario que tiene por cada token, hasta un total de
    # vocab_size, un vector de tamaño d_model. En el paper: we use learned
    # embeddings to convert the input tokens and output tokens to vectors
    # of dimension dmodel
    self.embedding = nn.Embedding(num_embeddings=vocab_size, embedding_dim=d_model)

3.2.4.1 forward(input_tensor)

Forward pass through the embedding layer.

Parameters:

Name Type Description Default
input_tensor Tensor

Input tensor of token indices.

 required

Returns:

Type Description
Tensor

Tensor of embedded input scaled by sqrt(d_model).
Source code in src/layers/nlp/transformer.py 
37
38
39
40
41
42
43
44
45
46
47
48
49
50
def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
    """
    Forward pass through the embedding layer.

    Args:
        input_tensor: Input tensor of token indices.

    Returns:
        Tensor of embedded input scaled by sqrt(d_model).
    """

    # Paper: In the embedding layers, we multiply those weights by sqrt(d_model)
    # Input_tensor (B, ...) -> (B, ..., d_model)
    return self.embedding(input_tensor) * math.sqrt(self.d_model)

3.2.5 LayerNormalization(features, eps=1e-06)

Applies layer normalization to input embeddings.

Initializes layer normalization.

Parameters:

Name Type Description Default
features int

Number of features in the input.

 required
eps float

Small constant for numerical stability.

 1e-06
Source code in src/layers/nlp/transformer.py 
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
def __init__(self, features: int, eps: float = 1e-6) -> None:
    """
    Initializes layer normalization.

    Args:
        features: Number of features in the input.
        eps: Small constant for numerical stability.
    """

    # Constructor de la clase
    super().__init__()

    # Definimos los parámetros de la clase
    self.features = features
    self.eps = eps

    # Utilizamos un factor alpha para multiplicar el valor de la normalización
    self.alpha = nn.Parameter(torch.ones(self.features))
    # Utilizamos un factor del sesgo para sumar
    self.bias = nn.Parameter(torch.zeros(self.features))

3.2.5.1 forward(input_embedding)

Forward pass for layer normalization.

Parameters:

Name Type Description Default
input_embedding Tensor

Tensor of input embeddings.

 required

Returns:

Type Description
Tensor

Normalized tensor.
Source code in src/layers/nlp/transformer.py 
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
def forward(self, input_embedding: torch.Tensor) -> torch.Tensor:
    """
    Forward pass for layer normalization.

    Args:
        input_embedding: Tensor of input embeddings.

    Returns:
        Normalized tensor.
    """

    # (B, sequence_length, d_model)
    mean = torch.mean(input=input_embedding, dim=-1, keepdim=True)
    var = torch.var(input=input_embedding, dim=-1, keepdim=True, unbiased=False)
    return (
        self.alpha * ((input_embedding - mean) / (torch.sqrt(var + self.eps)))
        + self.bias
    )

3.2.6 MultiHeadAttention(d_model, h, dropout_rate)

Applies multi-head attention mechanism.

Initializes multi-head attention layer.

Parameters:

Name Type Description Default
d_model int

Dimensionality of model embeddings.

 required
h int

Number of attention heads.

 required
dropout_rate float

Rate of dropout for regularization.

 required
Source code in src/layers/nlp/transformer.py 
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
def __init__(self, d_model: int, h: int, dropout_rate: float) -> None:
    """
    Initializes multi-head attention layer.

    Args:
        d_model: Dimensionality of model embeddings.
        h: Number of attention heads.
        dropout_rate: Rate of dropout for regularization.
    """

    # Constructor de la clase
    super().__init__()

    # el tamalo de los embeddings debe ser proporcional al número de cabezas
    # para realizar la división, por lo que es el resto ha de ser 0
    if d_model % h != 0:
        raise ValueError("d_model ha de ser divisible entre h")

    self.d_model = d_model
    self.h = h
    self.dropout = nn.Dropout(dropout_rate)

    # Valore establecidos en el paper
    self.d_k = self.d_model // self.h
    self.d_v = self.d_model // self.h

    # Parámetros
    self.W_K = nn.Linear(
        in_features=self.d_model, out_features=self.d_model, bias=False
    )
    self.W_Q = nn.Linear(
        in_features=self.d_model, out_features=self.d_model, bias=False
    )
    self.W_V = nn.Linear(
        in_features=self.d_model, out_features=self.d_model, bias=False
    )
    self.W_OUTPUT_CONCAT = nn.Linear(
        in_features=self.d_model, out_features=self.d_model, bias=False
    )

3.2.6.1 attention(k, q, v, mask=None, dropout=None) staticmethod

Computes scaled dot-product attention.

Parameters:

Name Type Description Default
k Tensor

Key tensor.

 required
q Tensor

Query tensor.

 required
v Tensor

Value tensor.

 required
mask Tensor | None

Optional mask tensor.

 None
dropout Dropout | None

Optional dropout layer.

 None

Returns:

Type Description

Tuple of attention output and scores.
Source code in src/layers/nlp/transformer.py 
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
@staticmethod
def attention(
    k: torch.Tensor,
    q: torch.Tensor,
    v: torch.Tensor,
    mask: torch.Tensor | None = None,
    dropout: nn.Dropout | None = None,
):
    """
    Computes scaled dot-product attention.

    Args:
        k: Key tensor.
        q: Query tensor.
        v: Value tensor.
        mask: Optional mask tensor.
        dropout: Optional dropout layer.

    Returns:
        Tuple of attention output and scores.
    """

    # Primero realizamos el producto matricial con la transpuesta
    # q = (Batch, h, seq_len, d_k)
    # k.T = (Batch, h, d_k, seq_len)
    # matmul_q_k = (Batch, h, seq_len, seq_len)
    matmul_q_k = q @ k.transpose(-2, -1)

    # Luego realizamos el escalado
    d_k = k.shape[-1]
    matmul_q_k_scaled = matmul_q_k / math.sqrt(d_k)

    # El enmascarado es para el decoder, relleno de infinitos
    if mask is not None:
        matmul_q_k_scaled.masked_fill_(mask == 0, -1e9)

    # Obtenemos los scores/puntuación de la atención
    attention_scores = F.softmax(input=matmul_q_k_scaled, dim=-1)

    # Aplicamos dropout
    if dropout is not None:
        attention_scores = dropout(attention_scores)

    # Multiplicamos por el valor
    # attention_scores = (Batch, h, seq_len, seq_len)
    # v = (Batch, h, seq_len, d_k)
    # Output = (Batch, h, seq_len, d_k)
    return (attention_scores @ v), attention_scores

3.2.6.2 forward(k, q, v, mask=None)

Forward pass through multi-head attention.

Parameters:

Name Type Description Default
k Tensor

Key tensor.

 required
q Tensor

Query tensor.

 required
v Tensor

Value tensor.

 required
mask Tensor | None

Optional mask tensor.

 None

Returns:

Type Description
Tensor

Tensor after attention and concatenation.
Source code in src/layers/nlp/transformer.py 
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
def forward(
    self,
    k: torch.Tensor,
    q: torch.Tensor,
    v: torch.Tensor,
    mask: torch.Tensor | None = None,
) -> torch.Tensor:
    """
    Forward pass through multi-head attention.

    Args:
        k: Key tensor.
        q: Query tensor.
        v: Value tensor.
        mask: Optional mask tensor.

    Returns:
        Tensor after attention and concatenation.
    """

    # k -> (Batch, seq_len, d_model) igual para el resto
    key_prima = self.W_K(k)
    query_prima = self.W_Q(q)
    value_prima = self.W_V(v)

    # Cambiamos las dimensiones y hacemos el split de los embedding para cada head
    # Pasando de (Batch, seq_len, d_model) a (Batch, seq_len, h, d_k)
    # Para luego pasar de (Batch, seq_len, h, d_k) a (Batch, h, seq_len, d_k)
    key_prima = key_prima.view(
        key_prima.shape[0], key_prima.shape[1], self.h, self.d_k
    ).transpose(1, 2)
    query_prima = query_prima.view(
        query_prima.shape[0], query_prima.shape[1], self.h, self.d_k
    ).transpose(1, 2)
    value_prima = value_prima.view(
        value_prima.shape[0], value_prima.shape[1], self.h, self.d_k
    ).transpose(1, 2)

    # Obtenemos la matriz de atencion y la puntuación
    # attention = (Batch, h, seq_len, d_k)
    # attention_scores = (Batch, h, seq_len, seq_len)
    attention, attention_scores = MultiHeadAttention.attention(
        k=key_prima,
        q=query_prima,
        v=value_prima,
        mask=mask,
        dropout=self.dropout,
    )

    # Tenemos que concatenar la información de todas las cabezas
    # Queremos (Batch, seq_len, d_model)
    # self.d_k = self.d_model // self.h; d_model = d_k * h
    attention = attention.transpose(1, 2)  # (Batch, seq_len, h, d_k)
    b, seq_len, h, d_k = attention.size()
    # Al parecer, contiguous permite evitar errores de memoria
    attention_concat = attention.contiguous().view(
        b, seq_len, h * d_k
    )  # (Batch, seq_len, h * d_k) = (Batch, seq_len, d_model)

    return self.W_OUTPUT_CONCAT(attention_concat)

3.2.7 PositionalEncoding(d_model, sequence_length, dropout_rate)

Adds positional encoding to input embeddings.

Initializes positional encoding layer.

Parameters:

Name Type Description Default
d_model int

Dimensionality of the embedding vectors.

 required
sequence_length int

Maximum sequence length.

 required
dropout_rate float

Rate of dropout for regularization.

 required
Source code in src/layers/nlp/transformer.py 
 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
def __init__(self, d_model: int, sequence_length: int, dropout_rate: float) -> None:
    """
    Initializes positional encoding layer.

    Args:
        d_model: Dimensionality of the embedding vectors.
        sequence_length: Maximum sequence length.
        dropout_rate: Rate of dropout for regularization.
    """

    # Constructor de la clase
    super().__init__()

    # Definimos los parámetros de la clase
    self.d_model = d_model

    # Cuando le damos una secuencia de tokens, tenemos que saber
    # la longitud máxima de la secuencia
    self.sequence_length = sequence_length
    self.dropout = nn.Dropout(dropout_rate)

    # Creamos una matriz del positional embedding
    # (sequence_length, d_model)
    pe_matrix = torch.zeros(size=(self.sequence_length, self.d_model))

    # # Ahora rellenamos la matriz de posiciones
    # # La posición va hasta el máximo de la longitud de la secuencia
    # for pos in range(self.sequence_length):
    # 	for i in range(0, d_model, 2):
    # 		# Para las posiciones pares usamos el seno
    # 		pe_matrix[pos, i] = torch.sin(pos / (10000 ** ((2 * i) / d_model)))
    # 		# Para las posiciones impares usamos el coseno
    # 		pe_matrix[pos, i + 1] = torch.cos(
    # 			pos / (10000 ** ((2 * (i + 1)) / d_model))
    # 		)

    # Crear vector de posiciones
    position = torch.arange(0, self.sequence_length, dtype=torch.float).unsqueeze(1)

    # Crear vector de divisores
    div_term = torch.exp(
        torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)
    )

    # Aplicar sin y cos
    pe_matrix[:, 0::2] = torch.sin(position * div_term)
    pe_matrix[:, 1::2] = torch.cos(position * div_term)

    # Tenemos que convertirlo a (1, sequence_length, d_model) para
    # procesarlo por lotes
    pe_matrix = pe_matrix.unsqueeze(0)

    # Esta matriz no se aprende, es fija, la tenemos que guardar con el modelo
    self.register_buffer(name="pe_matrix", tensor=pe_matrix)

3.2.7.1 forward(input_embedding)

Forward pass to add positional encoding.

Parameters:

Name Type Description Default
input_embedding Tensor

Tensor of input embeddings.

 required

Returns:

Type Description
Tensor

Tensor of embeddings with added positional encoding.
Source code in src/layers/nlp/transformer.py 
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
def forward(self, input_embedding: torch.Tensor) -> torch.Tensor:
    """
    Forward pass to add positional encoding.

    Args:
        input_embedding: Tensor of input embeddings.

    Returns:
        Tensor of embeddings with added positional encoding.
    """

    # (B, ..., d_model) -> (B, sequence_length, d_model)
    # Seleccionamos
    x = input_embedding + (
        self.pe_matrix[:, : input_embedding.shape[1], :]  # type: ignore
    ).requires_grad_(False)
    return self.dropout(x)

3.2.8 ProjectionLayer(d_model, vocab_size)

Converts d_model dimensions back to vocab_size.

Initializes projection layer.

Parameters:

Name Type Description Default
d_model int

Dimensionality of model embeddings.

 required
vocab_size int

Size of the vocabulary.

 required
Source code in src/layers/nlp/transformer.py 
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
def __init__(self, d_model: int, vocab_size: int) -> None:
    """
    Initializes projection layer.

    Args:
        d_model: Dimensionality of model embeddings.
        vocab_size: Size of the vocabulary.
    """

    super().__init__()

    self.d_model = d_model
    self.vocab_size = vocab_size

    self.projection_layer = nn.Linear(in_features=d_model, out_features=vocab_size)

3.2.8.1 forward(input_tensor)

Forward pass through projection layer.

Parameters:

Name Type Description Default
input_tensor Tensor

Input tensor to the projection layer.

 required

Returns:

Type Description
Tensor

Tensor with projected dimensions.
Source code in src/layers/nlp/transformer.py 
571
572
573
574
575
576
577
578
579
580
581
582
def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
    """
    Forward pass through projection layer.

    Args:
        input_tensor: Input tensor to the projection layer.

    Returns:
        Tensor with projected dimensions.
    """

    return self.projection_layer(input_tensor)

3.2.9 ResidualConnection(features, dropout_rate)

Applies residual connection around a sublayer.

Initializes residual connection layer.

Parameters:

Name Type Description Default
features int

Number of features in the input.

 required
dropout_rate float

Rate of dropout for regularization.

 required
Source code in src/layers/nlp/transformer.py 
374
375
376
377
378
379
380
381
382
383
384
385
386
def __init__(self, features: int, dropout_rate: float) -> None:
    """
    Initializes residual connection layer.

    Args:
        features: Number of features in the input.
        dropout_rate: Rate of dropout for regularization.
    """

    super().__init__()

    self.dropout = nn.Dropout(dropout_rate)
    self.layer_norm = LayerNormalization(features=features)

3.2.9.1 forward(input_tensor, sublayer)

Forward pass using residual connection.

Parameters:

Name Type Description Default
input_tensor Tensor

Input tensor to the residual layer.

 required
sublayer Module

Sublayer to apply within the residual connection.

 required

Returns:

Type Description
Tensor

Tensor with residual connection applied.
Source code in src/layers/nlp/transformer.py 
388
389
390
391
392
393
394
395
396
397
398
399
400
def forward(self, input_tensor: torch.Tensor, sublayer: nn.Module) -> torch.Tensor:
    """
    Forward pass using residual connection.

    Args:
        input_tensor: Input tensor to the residual layer.
        sublayer: Sublayer to apply within the residual connection.

    Returns:
        Tensor with residual connection applied.
    """

    return input_tensor + self.dropout(sublayer(self.layer_norm(input_tensor)))

3.2.10 Transformer(src_vocab_size, tgt_vocab_size, src_seq_len, tgt_seq_len, num_encoders, num_decoders, d_model, d_ff, h, dropout_rate)

Transformer model with encoder and decoder blocks.

Initializes transformer model.

Parameters:

Name Type Description Default
src_vocab_size int

Size of source vocabulary.

 required
tgt_vocab_size int

Size of target vocabulary.

 required
src_seq_len int

Maximum source sequence length.

 required
tgt_seq_len int

Maximum target sequence length.

 required
num_encoders int

Number of encoder blocks.

 required
num_decoders int

Number of decoder blocks.

 required
d_model int

Dimensionality of model embeddings.

 required
d_ff int

Dimensionality of feed-forward layer.

 required
h int

Number of attention heads.

 required
dropout_rate float

Rate of dropout for regularization.

 required
Source code in src/layers/nlp/transformer.py 
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
def __init__(
    self,
    src_vocab_size: int,
    tgt_vocab_size: int,
    src_seq_len: int,
    tgt_seq_len: int,
    num_encoders: int,
    num_decoders: int,
    d_model: int,
    d_ff: int,
    h: int,
    dropout_rate: float,
) -> None:
    """
    Initializes transformer model.

    Args:
        src_vocab_size: Size of source vocabulary.
        tgt_vocab_size: Size of target vocabulary.
        src_seq_len: Maximum source sequence length.
        tgt_seq_len: Maximum target sequence length.
        num_encoders: Number of encoder blocks.
        num_decoders: Number of decoder blocks.
        d_model: Dimensionality of model embeddings.
        d_ff: Dimensionality of feed-forward layer.
        h: Number of attention heads.
        dropout_rate: Rate of dropout for regularization.
    """

    super().__init__()

    # Parámetros
    self.src_vocab_size = src_vocab_size
    self.tgt_vocab_size = tgt_vocab_size
    self.src_seq_len = src_seq_len
    self.tgt_seq_len = tgt_seq_len
    self.num_encoders = num_encoders
    self.num_decoders = num_decoders
    self.d_model = d_model
    self.d_ff = d_ff
    self.h = h
    self.dropout_rate = dropout_rate

    # Embeddings y Positional Encoding
    self.src_embedding = InputEmbedding(
        d_model=self.d_model, vocab_size=self.src_vocab_size
    )
    self.tgt_embedding = InputEmbedding(
        d_model=self.d_model, vocab_size=self.tgt_vocab_size
    )
    self.src_positional_encoding = PositionalEncoding(
        d_model=self.d_model,
        sequence_length=self.src_seq_len,
        dropout_rate=self.dropout_rate,
    )
    self.tgt_positional_encoding = PositionalEncoding(
        d_model=self.d_model,
        sequence_length=self.tgt_seq_len,
        dropout_rate=self.dropout_rate,
    )

    # Capas del Encoder
    self.encoder_layers = nn.ModuleList(
        [
            EncoderBlock(
                d_model=self.d_model,
                d_ff=self.d_ff,
                h=self.h,
                dropout_rate=self.dropout_rate,
            )
            for _ in range(self.num_encoders)
        ]
    )

    # Capas del Decoder
    self.decoder_layers = nn.ModuleList(
        [
            DecoderBlock(
                d_model=self.d_model,
                d_ff=self.d_ff,
                h=self.h,
                dropout_rate=self.dropout_rate,
            )
            for _ in range(self.num_decoders)
        ]
    )

    # Capa de proyección final
    self.projection_layer = ProjectionLayer(
        d_model=self.d_model, vocab_size=self.tgt_vocab_size
    )

3.2.10.1 decode(decoder_input, encoder_output, src_mask=None, tgt_mask=None)

Decodes target input tensor using decoder blocks.

Parameters:

Name Type Description Default
decoder_input Tensor

Input tensor to the decoder.

 required
encoder_output Tensor

Output tensor from the encoder.

 required
src_mask Tensor | None

Optional source mask tensor.

 None
tgt_mask Tensor | None

Optional target mask tensor.

 None

Returns:

Type Description
Tensor

Decoded tensor.
Source code in src/layers/nlp/transformer.py 
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
def decode(
    self,
    decoder_input: torch.Tensor,
    encoder_output: torch.Tensor,
    src_mask: torch.Tensor | None = None,
    tgt_mask: torch.Tensor | None = None,
) -> torch.Tensor:
    """
    Decodes target input tensor using decoder blocks.

    Args:
        decoder_input: Input tensor to the decoder.
        encoder_output: Output tensor from the encoder.
        src_mask: Optional source mask tensor.
        tgt_mask: Optional target mask tensor.

    Returns:
        Decoded tensor.
    """

    # Aplicar embedding y positional encoding
    x = self.tgt_embedding(decoder_input)
    x = self.tgt_positional_encoding(x)

    # Pasar por todas las capas del decoder
    for decoder_layer in self.decoder_layers:
        x = decoder_layer(
            decoder_input=x,
            encoder_output=encoder_output,
            src_mask=src_mask,
            tgt_mask=tgt_mask,
        )

    return x

3.2.10.2 encode(encoder_input, src_mask=None)

Encodes source input tensor using encoder blocks.

Parameters:

Name Type Description Default
encoder_input Tensor

Input tensor to the encoder.

 required
src_mask Tensor | None

Optional source mask tensor.

 None

Returns:

Type Description
Tensor

Encoded tensor.
Source code in src/layers/nlp/transformer.py 
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
def encode(
    self, encoder_input: torch.Tensor, src_mask: torch.Tensor | None = None
) -> torch.Tensor:
    """
    Encodes source input tensor using encoder blocks.

    Args:
        encoder_input: Input tensor to the encoder.
        src_mask: Optional source mask tensor.

    Returns:
        Encoded tensor.
    """

    # Aplicar embedding y positional encoding
    x = self.src_embedding(encoder_input)
    x = self.src_positional_encoding(x)

    # Pasar por todas las capas del encoder
    for encoder_layer in self.encoder_layers:
        x = encoder_layer(input_tensor=x, mask=src_mask)

    return x

3.2.10.3 forward(src, tgt, src_mask=None, tgt_mask=None)

Processes input and target sequences through the encoder and decoder, applying optional source and target masks.

Parameters:

Name Type Description Default
src Tensor

Input sequence tensor.

 required
tgt Tensor

Target sequence tensor.

 required
src_mask Tensor | None

Optional mask for the input sequence.

 None
tgt_mask Tensor | None

Optional mask for the target sequence.

 None

Returns:

Type Description
Tensor

Tensor containing the final output after projection.
Source code in src/layers/nlp/transformer.py 
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
def forward(
    self,
    src: torch.Tensor,
    tgt: torch.Tensor,
    src_mask: torch.Tensor | None = None,
    tgt_mask: torch.Tensor | None = None,
) -> torch.Tensor:
    """
    Processes input and target sequences through the encoder
    and decoder, applying optional source and target masks.

    Args:
        src: Input sequence tensor.
        tgt: Target sequence tensor.
        src_mask: Optional mask for the input sequence.
        tgt_mask: Optional mask for the target sequence.

    Returns:
        Tensor containing the final output after projection.
    """

    # Encoder
    encoder_output = self.encode(src, src_mask)

    # Decoder
    decoder_output = self.decode(tgt, encoder_output, src_mask, tgt_mask)

    # Projection
    return self.projection_layer(decoder_output)