# 9.3. Transformer¶

Until now, we have covered the three major neural network architectures: the convolution neural network (CNN), the recurrent neural network (RNN), and the attention mechanism. Before we dive into the transformer architecture, let us quickly review the pros and cons for first two:

• The convolution neural network (CNN): Easy to parallelize at a layer but cannot capture the nonfixed sequential dependency very well.

• The recurrent neural network (RNN): Able to capture the long-range, variable-length sequential information, however, suffer from inability to parallelize within a sequence.

To combine the advantages from both CNN and RNN, [Vaswani et al., 2017] innovates a novel architecture with the attention mechanism that we just introduced in Section 9.1. This architecture, which is called as the transformer, achieves parallelization by capturing recurrence sequence with attention but at the same time encoding each item’s position in the sequence. As a result, the transformer leads to a compatible model with significantly shorter training time.

Similar to the seq2seq model in Section 8.14, the transformer is also based on the encoder-decoder architecture. However, the transformer differs to the former by replacing the recurrent layers in seq2seq with multi-head attention layers, incorporating the position-wise information through position encoding, and applying layer normalization.

Recall in Section 9.2, we introduced Sequence to Sequence (seq2seq) with attention model. To better visualize and compare the transformer architecture with it, we draw them side-by-side in Fig. 9.3.1. These two models are similar to each other overall: the source sequence embeddings are fed into $$n$$ repeated blocks. The outputs of the last block are then used as attention memory for the decoder. The target sequence embeddings is similarly fed into $$n$$ repeated blocks in the decoder, and the final outputs are obtained by applying a dense layer with vocabulary size to the last block’s outputs.

Fig. 9.3.1 The transformer architecture.

On the flip side, the transformer differs to the seq2seq with attention model in three major places:

1. Transformer block: a recurrent layer in seq2seq is replaced with a transformer block. This block contains a multi-head attention layer and a network with two position-wise feed-forward network layers for the encoder. For the decoder, another multi-head attention layer is used to take the encoder state.

2. Add and norm: the inputs and outputs of both the multi-head attention layer or the position-wise feed-forward network, are processed by two “add and norm” layer that contains a residual structure and a layer normalization layer.

3. Position encoding: since the self-attention layer does not distinguish the item order in a sequence, a positional encoding layer is used to add sequential information into each sequence item.

In the rest of this section, we will equip you with each new component introduced by the transformer, and get you up and running to construct a machine translation model.

import math
import d2l
from mxnet import np, npx, autograd
from mxnet.gluon import nn
npx.set_np()


Before the discussion of the multi-head attention layer, let us quick express the self-attention architecture. The self-attention model is a normal attention model, with its query, its key, and its value are copied exactly same from each item of the sequential inputs. As we illustrate in Fig. 9.3.2, self-attention outputs a same length sequential output for each input item. Compared to a recurrent layer, output items of a self-attention layer can be computed in parallel and, therefore, it is easy to obtain a high-efficient implementation.

Fig. 9.3.2 Self-attention architecture.

Based on the above, the multi-head attention layer consists of $$h$$ parallel self-attention layers, each one is called a head. For each head, before feeding into the attention layer, we project the queries, keys, and values with three dense layers with hidden sizes $$p_q$$, $$p_k$$, and $$p_v$$, respectively. The outputs of these $$h$$ attention heads are concatenated and then processed by a final dense layer.

To be more specific, assume that the dimension for a query, a key, and a value are $$d_q$$, $$d_k$$, and $$d_v$$, respectively. Then, for each head $$i=1,\ldots,h$$, we can train the learnable parameters $$\mathbf W_q^{(i)}\in\mathbb R^{p_q\times d_q}$$, $$\mathbf W_k^{(i)}\in\mathbb R^{p_k\times d_k}$$, and $$\mathbf W_v^{(i)}\in\mathbb R^{p_v\times d_v}$$. Therefore, the output for each head can be demonstrated by

(9.3.1)$\mathbf o^{(i)} = \textrm{attention}(\mathbf W_q^{(i)}\mathbf q, \mathbf W_k^{(i)}\mathbf k,\mathbf W_v^{(i)}\mathbf v),$

where the $$\text{attention}$$ in the formula can be any attention layer, such as the DotProductAttention and MLPAttention as we introduced in .. _sec_attention:.

After that, the output with length $$p_v$$ from each of the $$h$$ attention heads are concatenated to be a length $$h p_v$$ output, which is then passed the final dense layer with $$d_o$$ hidden units. The weights of this dense layer can be denoted by $$\mathbf W_o\in\mathbb R^{d_o\times h p_v}$$. As a result, the multi-head attention output will be

(9.3.2)$\begin{split}\mathbf o = \mathbf W_o \begin{bmatrix}\mathbf o^{(1)}\\\vdots\\\mathbf o^{(h)}\end{bmatrix}.\end{split}$

Let us implement the multi-head attention in MXNet. Assume that the multi-head attention contain the number heads num_heads $$=h$$, the hidden size hidden_size $$=p_q=p_k=p_v$$ are the same for the query, the key, and the value dense layers. In addition, since the multi-head attention keeps the same dimensionality between its input and its output, we have the output feature size $d_o =$ hidden_size as well.

class MultiHeadAttention(nn.Block):
def __init__(self, hidden_size, num_heads, dropout, **kwargs):
self.attention = d2l.DotProductAttention(dropout)
self.W_q = nn.Dense(hidden_size, use_bias=False, flatten=False)
self.W_k = nn.Dense(hidden_size, use_bias=False, flatten=False)
self.W_v = nn.Dense(hidden_size, use_bias=False, flatten=False)
self.W_o = nn.Dense(hidden_size, use_bias=False, flatten=False)

def forward(self, query, key, value, valid_length):
# query, key, and value shape: (batch_size, seq_len, dim),
# where seq_len is the length of input sequence
# valid_length shape is either (batch_size, ) or (batch_size, seq_len).

# Project and transpose query, key, and value from
# (batch_size, seq_len, hidden_size * num_heads) to
# (batch_size * num_heads, seq_len, hidden_size).

if valid_length is not None:
# Copy valid_length by num_heads times
if valid_length.ndim == 1:
else:

output = self.attention(query, key, value, valid_length)

# Transpose from (batch_size * num_heads, seq_len, hidden_size) back to
# (batch_size, seq_len, hidden_size * num_heads)
return self.W_o(output_concat)


Here are the definitions of the transpose functions transpose_qkv and transpose_output, who are the inverse of each other.

def transpose_qkv(X, num_heads):
# Original X shape: (batch_size, seq_len, hidden_size * num_heads),
# -1 means inferring its value,
# after first reshape, X shape: (batch_size, seq_len, num_heads, hidden_size)
X = X.reshape(X.shape[0], X.shape[1], num_heads, -1)

# After transpose, X shape: (batch_size, num_heads, seq_len, hidden_size)
X = X.transpose(0, 2, 1, 3)

# Merge the first two dimensions.
# Use reverse=True to infer shape from right to left.
# Output shape: (batch_size * num_heads, seq_len, hidden_size)
output = X.reshape(-1, X.shape[2], X.shape[3])
return output

# A reversed version of transpose_qkv
X = X.reshape(-1, num_heads, X.shape[1], X.shape[2])
X = X.transpose(0, 2, 1, 3)
return X.reshape(X.shape[0], X.shape[1], -1)


Let us validate the MultiHeadAttention model in the a toy example. Create a multi-head attention with the hidden size $$d_o = 100$$, the output will share the same batch size and sequence length as the input, but the last dimension will be equal to the hidden_size $$= 100$$.

cell = MultiHeadAttention(100, 10, 0.5)
cell.initialize()
X = np.ones((2, 4, 5))
valid_length = np.array([2,3])
cell(X, X, X, valid_length).shape

(2, 4, 100)


## 9.3.2. Position-wise Feed-Forward Networks¶

Another critical component in the transformer block is called position-wise feed-forward network (FFN). It accepts a $$3$$ dimensional input with shape (batch size, sequence length, feature size). The position-wise FFN consists of two dense layers that applies to the last dimension. Since the same two dense layers are used for each position item in the sequence, we referred it to as position-wise. Indeed, it equals to apply two $$Conv(1,1)$$, i.e., $$1 \times 1$$ convolution layers.

Below, the PositionWiseFFN shows how to implement a position-wise FFN with two dense layers of hidden size ffn_hidden_size and hidden_size_out, respectively.

class PositionWiseFFN(nn.Block):
def __init__(self, ffn_hidden_size, hidden_size_out, **kwargs):
super(PositionWiseFFN, self).__init__(**kwargs)
self.ffn_1 = nn.Dense(ffn_hidden_size, flatten=False, activation='relu')
self.ffn_2 = nn.Dense(hidden_size_out, flatten=False)

def forward(self, X):
return self.ffn_2(self.ffn_1(X))


Similar to the multi-head attention, the position-wise feed-forward network will only change the last dimension size of the input—the feature dimension. In addition, if two items in the input sequence are identical, the according outputs will be identical as well. Let us try a toy model!

ffn = PositionWiseFFN(4, 8)
ffn.initialize()
ffn(np.ones((2, 3, 4)))[0]

array([[-0.00073839,  0.00923239, -0.00016378,  0.00091236, -0.00763499,
0.00199923,  0.00446541,  0.00189135],
[-0.00073839,  0.00923239, -0.00016378,  0.00091236, -0.00763499,
0.00199923,  0.00446541,  0.00189135],
[-0.00073839,  0.00923239, -0.00016378,  0.00091236, -0.00763499,
0.00199923,  0.00446541,  0.00189135]])


Besides the above two components in the transformer block, the “add and norm” within the block also play a core role to connect the inputs and outputs of other layers smoothly. To be more clear, we add a layer that contains a residual structure and a layer normalization after both the multi-head attention layer and the position-wise FFN network. Layer normalization is similar to the batch normalization as we discussed in Section 7.5. One difference is that the mean and variances for the layer normalization are calculated along the last dimension, e.g X.mean(axis=-1) instead of the first batch dimension, e.g., X.mean(axis=0). Layer normalization prevents the range of values in the layers changing too much, which means that faster training and better generalization ability.

MXNet has both LayerNorm and BatchNorm implemented within the nn block. Let us call both of them and see the difference in the below example.

layer = nn.LayerNorm()
layer.initialize()
batch = nn.BatchNorm()
batch.initialize()
X = np.array([[1,2],[2,3]])
# compute mean and variance from X in the training mode.
print('layer norm:',layer(X), '\nbatch norm:', batch(X))

layer norm: [[-0.99998  0.99998]
[-0.99998  0.99998]]
batch norm: [[-0.99998 -0.99998]
[ 0.99998  0.99998]]


Now let us implement the connection block AddNorm together. AddNorm accepts two inputs $$X$$ and $$Y$$. We can imagine $$X$$ as the original input in the residual network, while $$Y$$ as the outputs from either the multi-head attention layer or the position-wise FFN network. In addition, we apply dropout on $$Y$$ for the purpose of regularization.

class AddNorm(nn.Block):
def __init__(self, dropout, **kwargs):
self.dropout = nn.Dropout(dropout)
self.norm = nn.LayerNorm()

def forward(self, X, Y):
return self.norm(self.dropout(Y) + X)


Due to the residual connection, $$X$$ and $$Y$$ should have the same shape.

add_norm = AddNorm(0.5)

(2, 3, 4)


## 9.3.4. Positional Encoding¶

Unlike the recurrent layer, both the multi-head attention layer and the position-wise feed-forward network compute the output of each item in the sequence independently. This functionality enables us to parallel the computation, but it fails to model the sequential information for a given sequence. To better capture the sequential information, the transformer model utilizes the positional encoding to maintain the positional information of the input sequence.

So what is the positional encoding? Assume that $$X\in\mathbb R^{l\times d}$$ is the embedding of an example, where $$l$$ is the sequence length and $$d$$ is the embedding size. This positional encoding layer encodes X’s position $$P\in\mathbb R^{l\times d}$$ and outputs $$P+X$$.

The position $$P$$ is a 2d matrix, where $$i$$ refers to the order in the sentence, and $$j$$ refers to the position along the embedding vector dimension. In this way, each value in the origin sequence is then maintained using the equations below:

(9.3.3)$P_{i,2j} = \sin(i/10000^{2j/d}),$
(9.3.4)$\quad P_{i,2j+1} = \cos(i/10000^{2j/d}),$

for $$i=0,\ldots,l-1$$ and $$j=0,\ldots,\lfloor(d-1)/2\rfloor$$.

An intuitive visualization and implementation of the positional encoding are showing below.

Fig. 9.3.4 Positional encoding.

class PositionalEncoding(nn.Block):
def __init__(self, embedding_size, dropout, max_len=1000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(dropout)
# Create a long enough P
self.P = np.zeros((1, max_len, embedding_size))
X = np.arange(0, max_len).reshape(-1,1) / np.power(
10000, np.arange(0, embedding_size, 2)/embedding_size)
self.P[:, :, 0::2] = np.sin(X)
self.P[:, :, 1::2] = np.cos(X)

def forward(self, X):
X = X + self.P[:, :X.shape[1], :].as_in_context(X.context)
return self.dropout(X)


Now we test the PositionalEncoding with a toy model for 4 dimensions. As we can see, the 4th dimension has the same frequency as the 5th but with different offset. The 5th and 6th dimension have a lower frequency.

pe = PositionalEncoding(20, 0)
pe.initialize()
Y = pe(np.zeros((1, 100, 20 )))
d2l.plot(np.arange(100), Y[0, :,4:8].T, figsize=(6, 2.5),
legend=["dim %d"%p for p in [4,5,6,7]])


## 9.3.5. Encoder¶

Armed with all the essential components of transformer, let us first build a encoder transformer block. This encoder contains a multi-head attention layer, a position-wise feed-forward network, and two “add and norm” connection blocks. As you can observer in the code, for both of the attention model and the positional FFN model in the EncoderBlock, their outputs’ dimension are equal to the embedding_size. This is due to the nature of the residual block, as we need to add these outputs back to the original value during “add and norm”.

class EncoderBlock(nn.Block):
def __init__(self, embedding_size, ffn_hidden_size, num_heads, dropout, **kwargs):
super(EncoderBlock, self).__init__(**kwargs)
self.ffn = PositionWiseFFN(ffn_hidden_size, embedding_size)

def forward(self, X, valid_length):
Y = self.addnorm_1(X, self.attention(X, X, X, valid_length))


Due to the residual connections, this block will not change the input shape. It means the embedding_size argument should be equal to the input size of the last dimension. In our toy example below, the embedding_size $$= 24$$, ffn_hidden_size $$=48$$, num_heads $$= 8$$, and dropout $$= 0.5$$.

X = np.ones((2, 100, 24))
encoder_blk = EncoderBlock(24, 48, 8, 0.5)
encoder_blk.initialize()
encoder_blk(X, valid_length).shape

(2, 100, 24)


Now it comes to the implementation of the whole encoder transformer as shown below. With the transformer encoder, $$n$$ blocks of EncoderBlock stack up one after another. Because of the residual connection, the embedding layer size $$d$$ is same as the transformer block output size. Also note that we multiple the embedding output by $$\sqrt{d}$$ to avoid its values are too small compared to positional encodings.

class TransformerEncoder(d2l.Encoder):
def __init__(self, vocab_size, embedding_size, ffn_hidden_size,
super(TransformerEncoder, self).__init__(**kwargs)
self.embedding_size = embedding_size
self.embed = nn.Embedding(vocab_size, embedding_size)
self.pos_encoding = PositionalEncoding(embedding_size, dropout)
self.blks = nn.Sequential()
for i in range(num_layers):

def forward(self, X, valid_length, *args):
X = self.pos_encoding(self.embed(X) * math.sqrt(self.embedding_size))
for blk in self.blks:
X = blk(X, valid_length)
return X


Let us create an encoder with two stacked encoder transformer blocks, whose hyper-parameters are same as before. Similar to the previous toy example’s parameters, we add two more parameters vocab_size to be $$200$$ and num_layers to be $$2$$ here.

encoder = TransformerEncoder(200, 24, 48, 8, 2, 0.5)
encoder.initialize()
encoder(np.ones((2, 100)), valid_length).shape

(2, 100, 24)


## 9.3.6. Decoder¶

The decoder transformer block looks similar to the encoder transformer block. However, besides the two sub-layers—the multi-head attention layer and the positional encoding network, the decoder transformer block contains a third sub-layer, which applys multi-head attention on the output of the encoder stack. Similar to the encoder transformer block, the decoder transformer block employ “add and norm”, i.e., the residual connections and the layer normalization to connect each of the sub-layers.

To be specific, at time step $$t$$, assume that $$\mathbf x_t$$ is the current input, i.e., the query. As illustrate in Fig. 9.3.5, the keys and values of the self-attention layer consist of the current query with all past queries $$\mathbf x_1, \ldots, \mathbf x_{t-1}$$.

Fig. 9.3.5 Predict at time step $$t$$ for a self-attention layer.

During the training, because the output for the $$t$$-query could observe all the previous key-value pairs, which results in an inconsistent behavior than prediction. We can eliminate the unnecessary information by specifying the valid length to be $$t$$ for the $$t^\textrm{th}$$ query.

class DecoderBlock(nn.Block):
# i means it is the i-th block in the decoder
def __init__(self, embedding_size, ffn_hidden_size, num_heads, dropout, i, **kwargs):
super(DecoderBlock, self).__init__(**kwargs)
self.i = i
self.ffn = PositionWiseFFN(ffn_hidden_size, embedding_size)

def forward(self, X, state):
enc_outputs, enc_valid_lengh = state[0], state[1]
# state[2][i] contains the past queries for this block
if state[2][self.i] is None:
key_values = X
else:
key_values = np.concatenate((state[2][self.i], X), axis=1)
state[2][self.i] = key_values
batch_size, seq_len, _ = X.shape
# shape: (batch_size, seq_len), the values in the j-th column
# are j+1
valid_length = np.tile(np.arange(1, seq_len+1, ctx=X.context),
(batch_size, 1))
else:
valid_length = None

X2 = self.attention_1(X, key_values, key_values, valid_length)
Y2 = self.attention_2(Y, enc_outputs, enc_outputs, enc_valid_lengh)


Similar to the encoder transformer block, embedding_size should be equal to the last dimension size of $$X$$.

decoder_blk = DecoderBlock(24, 48, 8, 0.5, 0)
decoder_blk.initialize()
X = np.ones((2, 100, 24))
state = [encoder_blk(X, valid_length), valid_length, [None]]
decoder_blk(X, state)[0].shape

(2, 100, 24)


The construction of the whole decoder transformer is identical to the encoder transformer, except for the additional last dense layer to obtain the output confident scores. Let us implement the decoder transformer TransformerDecoder together. Besides the regular hyper-parameters such as the vocab_size and embedding_size, the decoder transformer also needs the encoder transformer’s outputs enc_outputs and env_valid_lengh.

class TransformerDecoder(d2l.Decoder):
def __init__(self, vocab_size, embedding_size, ffn_hidden_size,
super(TransformerDecoder, self).__init__(**kwargs)
self.embedding_size = embedding_size
self.num_layers = num_layers
self.embed = nn.Embedding(vocab_size, embedding_size)
self.pos_encoding = PositionalEncoding(embedding_size, dropout)
self.blks = nn.Sequential()
for i in range(num_layers):
self.dense = nn.Dense(vocab_size, flatten=False)

def init_state(self, enc_outputs, env_valid_lengh, *args):
return [enc_outputs, env_valid_lengh, [None]*self.num_layers]

def forward(self, X, state):
X = self.pos_encoding(self.embed(X) * math.sqrt(self.embedding_size))
for blk in self.blks:
X, state = blk(X, state)
return self.dense(X), state


## 9.3.7. Training¶

Finally, we are fully prepared to build a encoder-decoder model with transformer architecture. Similar to the seq2seq with attention model we built in .. _sec_seq2seq_attention:, we use the following hyper-parameters: two transformer blocks with both the embedding size and the block output size to be $$32$$. The additional hyper-parameters are chosen as $$4$$ heads with the hidden size to be $$2$$ times larger than output size.

embed_size, embedding_size, num_layers, dropout = 32, 32, 2, 0.0
batch_size, num_steps = 64, 10
lr, num_epochs, ctx = 0.005, 100, d2l.try_gpu()

src_vocab, tgt_vocab, train_iter = d2l.load_data_nmt(batch_size, num_steps)

encoder = TransformerEncoder(
len(src_vocab), embedding_size, num_hiddens, num_heads, num_layers, dropout)
decoder = TransformerDecoder(
len(src_vocab), embedding_size, num_hiddens, num_heads, num_layers, dropout)
model = d2l.EncoderDecoder(encoder, decoder)
d2l.train_s2s_ch8(model, train_iter, lr, num_epochs, ctx)

loss 0.033, 3364 tokens/sec on gpu(0)


As we can see from the training time and accuracy, compared to the seq2seq model with attention model, the transformer runs faster per epoch, and converges faster at the beginning.

Last but not the least, let us translate some sentences. Unsurprisingly, this model outperforms the previous one we trained in the below examples.

for sentence in ['Go .', 'Wow !', "I'm OK .", 'I won !']:
print(sentence + ' => ' + d2l.predict_s2s_ch8(
model, sentence, src_vocab, tgt_vocab, num_steps, ctx))

Go . => va !
Wow ! => <unk> !
I'm OK . => ça va .
I won ! => je l'ai emporté !


## 9.3.8. Summary¶

• Transformer model is based on N*N encoder-decoder architecture. It differs from seq2seq with attention in 3 major places.

• Multi-head attention layer contains $$h$$ parallel attention layers.

• Position-wise feed-forward network consists of two dense layers that applies to the last dimension.

• Layer normalization differs from batch normalization by normalizing along the last dimension (the feature dimension) instead of the first (batch size) dimension.

• Positional encoding is the only place that adds positional information to the transformer model.

## 9.3.9. Exercises¶

• Try a large size of epochs and compare the loss between seq2seq model and transformer model in earlier stage and later stage.

• Can you think of another functions for positional encoding?

• Compare layer normalization and batch normalization, what are the suitable scenarios to apply them?