Author: Darshan Deshpande
Date created: 2021/10/05
Last modified: 2021/10/05
Description: FNet transformer for text generation in Keras.
The original transformer implementation (Vaswani et al., 2017) was one of the major breakthroughs in Natural Language Processing, giving rise to important architectures such BERT and GPT. However, the drawback of these architectures is that the self-attention mechanism they use is computationally expensive. The FNet architecture proposes to replace this self-attention attention with a leaner mechanism: a Fourier transformation-based linear mixer for input tokens.
The FNet model was able to achieve 92-97% of BERT's accuracy while training 80% faster on GPUs and almost 70% faster on TPUs. This type of design provides an efficient and small model size, leading to faster inference times.
In this example, we will implement and train this architecture on the Cornell Movie Dialog corpus to show the applicability of this model to text generation.
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import os
# Defining hyperparameters
VOCAB_SIZE = 8192
MAX_SAMPLES = 50000
BUFFER_SIZE = 20000
MAX_LENGTH = 40
EMBED_DIM = 256
LATENT_DIM = 512
NUM_HEADS = 8
BATCH_SIZE = 64
We will be using the Cornell Dialog Corpus. We will parse the movie conversations into questions and answers sets.
path_to_zip = keras.utils.get_file(
"cornell_movie_dialogs.zip",
origin="http://www.cs.cornell.edu/~cristian/data/cornell_movie_dialogs_corpus.zip",
extract=True,
)
path_to_dataset = os.path.join(
os.path.dirname(path_to_zip), "cornell movie-dialogs corpus"
)
path_to_movie_lines = os.path.join(path_to_dataset, "movie_lines.txt")
path_to_movie_conversations = os.path.join(path_to_dataset, "movie_conversations.txt")
def load_conversations():
# Helper function for loading the conversation splits
id2line = {}
with open(path_to_movie_lines, errors="ignore") as file:
lines = file.readlines()
for line in lines:
parts = line.replace("\n", "").split(" +++$+++ ")
id2line[parts[0]] = parts[4]
inputs, outputs = [], []
with open(path_to_movie_conversations, "r") as file:
lines = file.readlines()
for line in lines:
parts = line.replace("\n", "").split(" +++$+++ ")
# get conversation in a list of line ID
conversation = [line[1:-1] for line in parts[3][1:-1].split(", ")]
for i in range(len(conversation) - 1):
inputs.append(id2line[conversation[i]])
outputs.append(id2line[conversation[i + 1]])
if len(inputs) >= MAX_SAMPLES:
return inputs, outputs
return inputs, outputs
questions, answers = load_conversations()
# Splitting training and validation sets
train_dataset = tf.data.Dataset.from_tensor_slices((questions[:40000], answers[:40000]))
val_dataset = tf.data.Dataset.from_tensor_slices((questions[40000:], answers[40000:]))
Downloading data from http://www.cs.cornell.edu/~cristian/data/cornell_movie_dialogs_corpus.zip
9920512/9916637 [==============================] - 0s 0us/step
9928704/9916637 [==============================] - 0s 0us/step
def preprocess_text(sentence):
sentence = tf.strings.lower(sentence)
# Adding a space between the punctuation and the last word to allow better tokenization
sentence = tf.strings.regex_replace(sentence, r"([?.!,])", r" \1 ")
# Replacing multiple continuous spaces with a single space
sentence = tf.strings.regex_replace(sentence, r"\s\s+", " ")
# Replacing non english words with spaces
sentence = tf.strings.regex_replace(sentence, r"[^a-z?.!,]+", " ")
sentence = tf.strings.strip(sentence)
sentence = tf.strings.join(["[start]", sentence, "[end]"], separator=" ")
return sentence
vectorizer = layers.TextVectorization(
VOCAB_SIZE,
standardize=preprocess_text,
output_mode="int",
output_sequence_length=MAX_LENGTH,
)
# We will adapt the vectorizer to both the questions and answers
# This dataset is batched to parallelize and speed up the process
vectorizer.adapt(tf.data.Dataset.from_tensor_slices((questions + answers)).batch(128))
TextVectorization
def vectorize_text(inputs, outputs):
inputs, outputs = vectorizer(inputs), vectorizer(outputs)
# One extra padding token to the right to match the output shape
outputs = tf.pad(outputs, [[0, 1]])
return (
{"encoder_inputs": inputs, "decoder_inputs": outputs[:-1]},
{"outputs": outputs[1:]},
)
train_dataset = train_dataset.map(vectorize_text, num_parallel_calls=tf.data.AUTOTUNE)
val_dataset = val_dataset.map(vectorize_text, num_parallel_calls=tf.data.AUTOTUNE)
train_dataset = (
train_dataset.cache()
.shuffle(BUFFER_SIZE)
.batch(BATCH_SIZE)
.prefetch(tf.data.AUTOTUNE)
)
val_dataset = val_dataset.cache().batch(BATCH_SIZE).prefetch(tf.data.AUTOTUNE)
The FNet paper proposes a replacement for the standard attention mechanism used by the Transformer architecture (Vaswani et al., 2017).
The outputs of the FFT layer are complex numbers. To avoid dealing with complex layers, only the real part (the magnitude) is extracted.
The dense layers that follow the Fourier transformation act as convolutions applied on the frequency domain.
class FNetEncoder(layers.Layer):
def __init__(self, embed_dim, dense_dim, **kwargs):
super().__init__(**kwargs)
self.embed_dim = embed_dim
self.dense_dim = dense_dim
self.dense_proj = keras.Sequential(
[
layers.Dense(dense_dim, activation="relu"),
layers.Dense(embed_dim),
]
)
self.layernorm_1 = layers.LayerNormalization()
self.layernorm_2 = layers.LayerNormalization()
def call(self, inputs):
# Casting the inputs to complex64
inp_complex = tf.cast(inputs, tf.complex64)
# Projecting the inputs to the frequency domain using FFT2D and
# extracting the real part of the output
fft = tf.math.real(tf.signal.fft2d(inp_complex))
proj_input = self.layernorm_1(inputs + fft)
proj_output = self.dense_proj(proj_input)
return self.layernorm_2(proj_input + proj_output)
The decoder architecture remains the same as the one proposed by (Vaswani et al., 2017) in the original transformer architecture, consisting of an embedding, positional encoding, two masked multi-head attention layers and finally the dense output layers. The architecture that follows is taken from Deep Learning with Python, second edition, chapter 11.
class PositionalEmbedding(layers.Layer):
def __init__(self, sequence_length, vocab_size, embed_dim, **kwargs):
super().__init__(**kwargs)
self.token_embeddings = layers.Embedding(
input_dim=vocab_size, output_dim=embed_dim
)
self.position_embeddings = layers.Embedding(
input_dim=sequence_length, output_dim=embed_dim
)
self.sequence_length = sequence_length
self.vocab_size = vocab_size
self.embed_dim = embed_dim
def call(self, inputs):
length = tf.shape(inputs)[-1]
positions = tf.range(start=0, limit=length, delta=1)
embedded_tokens = self.token_embeddings(inputs)
embedded_positions = self.position_embeddings(positions)
return embedded_tokens + embedded_positions
def compute_mask(self, inputs, mask=None):
return tf.math.not_equal(inputs, 0)
class FNetDecoder(layers.Layer):
def __init__(self, embed_dim, latent_dim, num_heads, **kwargs):
super().__init__(**kwargs)
self.embed_dim = embed_dim
self.latent_dim = latent_dim
self.num_heads = num_heads
self.attention_1 = layers.MultiHeadAttention(
num_heads=num_heads, key_dim=embed_dim
)
self.attention_2 = layers.MultiHeadAttention(
num_heads=num_heads, key_dim=embed_dim
)
self.dense_proj = keras.Sequential(
[
layers.Dense(latent_dim, activation="relu"),
layers.Dense(embed_dim),
]
)
self.layernorm_1 = layers.LayerNormalization()
self.layernorm_2 = layers.LayerNormalization()
self.layernorm_3 = layers.LayerNormalization()
self.supports_masking = True
def call(self, inputs, encoder_outputs, mask=None):
causal_mask = self.get_causal_attention_mask(inputs)
if mask is not None:
padding_mask = tf.cast(mask[:, tf.newaxis, :], dtype="int32")
padding_mask = tf.minimum(padding_mask, causal_mask)
attention_output_1 = self.attention_1(
query=inputs, value=inputs, key=inputs, attention_mask=causal_mask
)
out_1 = self.layernorm_1(inputs + attention_output_1)
attention_output_2 = self.attention_2(
query=out_1,
value=encoder_outputs,
key=encoder_outputs,
attention_mask=padding_mask,
)
out_2 = self.layernorm_2(out_1 + attention_output_2)
proj_output = self.dense_proj(out_2)
return self.layernorm_3(out_2 + proj_output)
def get_causal_attention_mask(self, inputs):
input_shape = tf.shape(inputs)
batch_size, sequence_length = input_shape[0], input_shape[1]
i = tf.range(sequence_length)[:, tf.newaxis]
j = tf.range(sequence_length)
mask = tf.cast(i >= j, dtype="int32")
mask = tf.reshape(mask, (1, input_shape[1], input_shape[1]))
mult = tf.concat(
[tf.expand_dims(batch_size, -1), tf.constant([1, 1], dtype=tf.int32)],
axis=0,
)
return tf.tile(mask, mult)
def create_model():
encoder_inputs = keras.Input(shape=(None,), dtype="int32", name="encoder_inputs")
x = PositionalEmbedding(MAX_LENGTH, VOCAB_SIZE, EMBED_DIM)(encoder_inputs)
encoder_outputs = FNetEncoder(EMBED_DIM, LATENT_DIM)(x)
encoder = keras.Model(encoder_inputs, encoder_outputs)
decoder_inputs = keras.Input(shape=(None,), dtype="int32", name="decoder_inputs")
encoded_seq_inputs = keras.Input(
shape=(None, EMBED_DIM), name="decoder_state_inputs"
)
x = PositionalEmbedding(MAX_LENGTH, VOCAB_SIZE, EMBED_DIM)(decoder_inputs)
x = FNetDecoder(EMBED_DIM, LATENT_DIM, NUM_HEADS)(x, encoded_seq_inputs)
x = layers.Dropout(0.5)(x)
decoder_outputs = layers.Dense(VOCAB_SIZE, activation="softmax")(x)
decoder = keras.Model(
[decoder_inputs, encoded_seq_inputs], decoder_outputs, name="outputs"
)
decoder_outputs = decoder([decoder_inputs, encoder_outputs])
fnet = keras.Model([encoder_inputs, decoder_inputs], decoder_outputs, name="fnet")
return fnet
fnet = create_model()
fnet.compile("adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
Here, the epochs
parameter is set to a single epoch, but in practice the model will take around
20-30 epochs of training to start outputting comprehensible sentences. Although accuracy
is not a good measure for this task, we will use it just to get a hint of the improvement
of the network.
fnet.fit(train_dataset, epochs=1, validation_data=val_dataset)
625/625 [==============================] - 96s 133ms/step - loss: 1.3036 - accuracy: 0.4354 - val_loss: 0.7964 - val_accuracy: 0.6374
<keras.callbacks.History at 0x7f0d8d214c90>
VOCAB = vectorizer.get_vocabulary()
def decode_sentence(input_sentence):
# Mapping the input sentence to tokens and adding start and end tokens
tokenized_input_sentence = vectorizer(
tf.constant("[start] " + preprocess_text(input_sentence) + " [end]")
)
# Initializing the initial sentence consisting of only the start token.
tokenized_target_sentence = tf.expand_dims(VOCAB.index("[start]"), 0)
decoded_sentence = ""
for i in range(MAX_LENGTH):
# Get the predictions
predictions = fnet.predict(
{
"encoder_inputs": tf.expand_dims(tokenized_input_sentence, 0),
"decoder_inputs": tf.expand_dims(
tf.pad(
tokenized_target_sentence,
[[0, MAX_LENGTH - tf.shape(tokenized_target_sentence)[0]]],
),
0,
),
}
)
# Calculating the token with maximum probability and getting the corresponding word
sampled_token_index = tf.argmax(predictions[0, i, :])
sampled_token = VOCAB[sampled_token_index.numpy()]
# If sampled token is the end token then stop generating and return the sentence
if tf.equal(sampled_token_index, VOCAB.index("[end]")):
break
decoded_sentence += sampled_token + " "
tokenized_target_sentence = tf.concat(
[tokenized_target_sentence, [sampled_token_index]], 0
)
return decoded_sentence
decode_sentence("Where have you been all this time?")
'i m sorry .'
This example shows how to train and perform inference using the FNet model. For getting insight into the architecture or for further reading, you can refer to:
Thanks to François Chollet for his Keras example on English-to-Spanish translation with a sequence-to-sequence Transformer from which the decoder implementation was extracted.