Transformer Encoder and Decoder Models

11import math
13import torch
14import torch.nn as nn
15from labml_helpers.module import Module
17from labml_nn.utils import clone_module_list
18from .feed_forward import FeedForward
19from .mha import MultiHeadAttention
20from .positional_encoding import get_positional_encoding
23class EmbeddingsWithPositionalEncoding(Module):
30    def __init__(self, d_model: int, n_vocab: int, max_len: int = 5000):
31        super().__init__()
32        self.linear = nn.Embedding(n_vocab, d_model)
33        self.d_model = d_model
34        self.register_buffer('positional_encodings', get_positional_encoding(d_model, max_len))
36    def forward(self, x: torch.Tensor):
37        pe = self.positional_encodings[:x.shape[0]].requires_grad_(False)
38        return self.linear(x) * math.sqrt(self.d_model) + pe
41class EmbeddingsWithLearnedPositionalEncoding(Module):
48    def __init__(self, d_model: int, n_vocab: int, max_len: int = 5000):
49        super().__init__()
50        self.linear = nn.Embedding(n_vocab, d_model)
51        self.d_model = d_model
52        self.positional_encodings = nn.Parameter(torch.zeros(max_len, 1, d_model), requires_grad=True)
54    def forward(self, x: torch.Tensor):
55        pe = self.positional_encodings[:x.shape[0]]
56        return self.linear(x) * math.sqrt(self.d_model) + pe

Transformer Layer

This can act as an encoder layer or a decoder layer.

🗒 Some implementations, including the paper seem to have differences in where the layer-normalization is done. Here we do a layer normalization before attention and feed-forward networks, and add the original residual vectors. Alternative is to do a layer normalization after adding the residuals. But we found this to be less stable when training. We found a detailed discussion about this in the paper On Layer Normalization in the Transformer Architecture.

59class TransformerLayer(Module):
  • d_model is the token embedding size
  • self_attn is the self attention module
  • src_attn is the source attention module (when this is used in a decoder)
  • feed_forward is the feed forward module
  • dropout_prob is the probability of dropping out after self attention and FFN
77    def __init__(self, *,
78                 d_model: int,
79                 self_attn: MultiHeadAttention,
80                 src_attn: MultiHeadAttention = None,
81                 feed_forward: FeedForward,
82                 dropout_prob: float):
90        super().__init__()
91        self.size = d_model
92        self.self_attn = self_attn
93        self.src_attn = src_attn
94        self.feed_forward = feed_forward
95        self.dropout = nn.Dropout(dropout_prob)
96        self.norm_self_attn = nn.LayerNorm([d_model])
97        if self.src_attn is not None:
98            self.norm_src_attn = nn.LayerNorm([d_model])
99        self.norm_ff = nn.LayerNorm([d_model])

Whether to save input to the feed forward layer

101        self.is_save_ff_input = False
103    def forward(self, *,
104                 x: torch.Tensor,
105                 mask: torch.Tensor,
106                 src: torch.Tensor = None,
107                 src_mask: torch.Tensor = None):

Normalize the vectors before doing self attention

109        z = self.norm_self_attn(x)

Run through self attention, i.e. keys and values are from self

111        self_attn = self.self_attn(query=z, key=z, value=z, mask=mask)

Add the self attention results

113        x = x + self.dropout(self_attn)

If a source is provided, get results from attention to source. This is when you have a decoder layer that pays attention to encoder outputs

118        if src is not None:

Normalize vectors

120            z = self.norm_src_attn(x)

Attention to source. i.e. keys and values are from source

122            attn_src = self.src_attn(query=z, key=src, value=src, mask=src_mask)

Add the source attention results

124            x = x + self.dropout(attn_src)

Normalize for feed-forward

127        z = self.norm_ff(x)

Save the input to the feed forward layer if specified

129        if self.is_save_ff_input:
130            self.ff_input = z.clone()

Pass through the feed-forward network

132        ff = self.feed_forward(z)

Add the feed-forward results back

134        x = x + self.dropout(ff)
136        return x
139class Encoder(Module):
146    def __init__(self, layer: TransformerLayer, n_layers: int):
147        super().__init__()

Make copies of the transformer layer

149        self.layers = clone_module_list(layer, n_layers)

Final normalization layer

151        self.norm = nn.LayerNorm([layer.size])
153    def forward(self, x: torch.Tensor, mask: torch.Tensor):

Run through each transformer layer

155        for layer in self.layers:
156            x = layer(x=x, mask=mask)

Finally, normalize the vectors

158        return self.norm(x)
161class Decoder(Module):
168    def __init__(self, layer: TransformerLayer, n_layers: int):
169        super().__init__()

Make copies of the transformer layer

171        self.layers = clone_module_list(layer, n_layers)

Final normalization layer

173        self.norm = nn.LayerNorm([layer.size])
175    def forward(self, x: torch.Tensor, memory: torch.Tensor, src_mask: torch.Tensor, tgt_mask: torch.Tensor):

Run through each transformer layer

177        for layer in self.layers:
178            x = layer(x=x, mask=tgt_mask, src=memory, src_mask=src_mask)

Finally, normalize the vectors

180        return self.norm(x)


This predicts the tokens and gives the lof softmax of those. You don’t need this if you are using nn.CrossEntropyLoss.

183class Generator(Module):
193    def __init__(self, n_vocab: int, d_model: int):
194        super().__init__()
195        self.projection = nn.Linear(d_model, n_vocab)
197    def forward(self, x):
198        return self.projection(x)
201class EncoderDecoder(Module):
208    def __init__(self, encoder: Encoder, decoder: Decoder, src_embed: Module, tgt_embed: Module, generator: Module):
209        super().__init__()
210        self.encoder = encoder
211        self.decoder = decoder
212        self.src_embed = src_embed
213        self.tgt_embed = tgt_embed
214        self.generator = generator

This was important from their code. Initialize parameters with Glorot / fan_avg.

218        for p in self.parameters():
219            if p.dim() > 1:
220                nn.init.xavier_uniform_(p)
222    def forward(self, src: torch.Tensor, tgt: torch.Tensor, src_mask: torch.Tensor, tgt_mask: torch.Tensor):

Run the source through encoder

224        enc = self.encode(src, src_mask)

Run encodings and targets through decoder

226        return self.decode(enc, src_mask, tgt, tgt_mask)
228    def encode(self, src: torch.Tensor, src_mask: torch.Tensor):
229        return self.encoder(self.src_embed(src), src_mask)
231    def decode(self, memory: torch.Tensor, src_mask: torch.Tensor, tgt: torch.Tensor, tgt_mask: torch.Tensor):
232        return self.decoder(self.tgt_embed(tgt), memory, src_mask, tgt_mask)