Chapter 16 – Natural Language Processing with RNNs and Attention
This notebook contains all the sample code and solutions to the exercises in chapter 16.
|
|
Setup¶
This project requires Python 3.7 or above:
In [1]:
import sys
assert sys.version_info >= (3, 7)
Warning: the latest TensorFlow versions are based on Keras 3. For previous chapters, it wasn't too hard to update the code to support Keras 3, but unfortunately it's much harder for this chapter: for example, stateful RNNs work very differently, ragged tensors are no longer supported, TensorFlow Hub models are no longer supported, and more. So for this chapter I've had to revert to Keras 2. To do that, I set the TF_USE_LEGACY_KERAS environment variable to "1" and import the tf_keras package. This ensures that tf.keras points to tf_keras, which is Keras 2.*.
In [2]:
import os
os.environ["TF_USE_LEGACY_KERAS"] = "1"
import tf_keras
And TensorFlow ≥ 2.8:
In [3]:
from packaging import version
import tensorflow as tf
assert version.parse(tf.__version__) >= version.parse("2.8.0")
As we did in earlier chapters, let's define the default font sizes to make the figures prettier:
In [4]:
import matplotlib.pyplot as plt
plt.rc('font', size=14)
plt.rc('axes', labelsize=14, titlesize=14)
plt.rc('legend', fontsize=14)
plt.rc('xtick', labelsize=10)
plt.rc('ytick', labelsize=10)
And let's create the images/nlp folder (if it doesn't already exist), and define the save_fig() function which is used through this notebook to save the figures in high-res for the book:
In [5]:
from pathlib import Path
IMAGES_PATH = Path() / "images" / "nlp"
IMAGES_PATH.mkdir(parents=True, exist_ok=True)
def save_fig(fig_id, tight_layout=True, fig_extension="png", resolution=300):
path = IMAGES_PATH / f"{fig_id}.{fig_extension}"
if tight_layout:
plt.tight_layout()
plt.savefig(path, format=fig_extension, dpi=resolution)
This chapter can be very slow without a GPU, so let's make sure there's one, or else issue a warning:
In [6]:
if not tf.config.list_physical_devices('GPU'):
print("No GPU was detected. Neural nets can be very slow without a GPU.")
if "google.colab" in sys.modules:
print("Go to Runtime > Change runtime and select a GPU hardware "
"accelerator.")
if "kaggle_secrets" in sys.modules:
print("Go to Settings > Accelerator and select GPU.")
No GPU was detected. Neural nets can be very slow without a GPU.
Generating Shakespearean Text Using a Character RNN¶
Creating the Training Dataset¶
Let's download the Shakespeare data from Andrej Karpathy's char-rnn project
In [7]:
import tensorflow as tf
shakespeare_url = "https://homl.info/shakespeare" # shortcut URL
filepath = tf.keras.utils.get_file("shakespeare.txt", shakespeare_url)
with open(filepath) as f:
shakespeare_text = f.read()
Downloading data from https://homl.info/shakespeare 1115394/1115394 ━━━━━━━━━━━━━━━━━━━━ 0s 0us/step
In [8]:
# extra code – shows a short text sample
print(shakespeare_text[:80])
First Citizen: Before we proceed any further, hear me speak. All: Speak, speak.
In [9]:
# extra code – shows all 39 distinct characters (after converting to lower case)
"".join(sorted(set(shakespeare_text.lower())))
Out[9]:
"\n !$&',-.3:;?abcdefghijklmnopqrstuvwxyz"
In [10]:
text_vec_layer = tf.keras.layers.TextVectorization(split="character",
standardize="lower")
text_vec_layer.adapt([shakespeare_text])
encoded = text_vec_layer([shakespeare_text])[0]
In [11]:
encoded -= 2 # drop tokens 0 (pad) and 1 (unknown), which we will not use
n_tokens = text_vec_layer.vocabulary_size() - 2 # number of distinct chars = 39
dataset_size = len(encoded) # total number of chars = 1,115,394
In [12]:
n_tokens
Out[12]:
39
In [13]:
dataset_size
Out[13]:
1115394
In [14]:
def to_dataset(sequence, length, shuffle=False, seed=None, batch_size=32):
ds = tf.data.Dataset.from_tensor_slices(sequence)
ds = ds.window(length + 1, shift=1, drop_remainder=True)
ds = ds.flat_map(lambda window_ds: window_ds.batch(length + 1))
if shuffle:
ds = ds.shuffle(100_000, seed=seed)
ds = ds.batch(batch_size)
return ds.map(lambda window: (window[:, :-1], window[:, 1:])).prefetch(1)
In [15]:
# extra code – a simple example using to_dataset()
# There's just one sample in this dataset: the input represents "to b" and the
# output represents "o be"
list(to_dataset(text_vec_layer(["To be"])[0], length=4))
2025-02-09 15:24:14.780201: I tensorflow/core/framework/local_rendezvous.cc:405] Local rendezvous is aborting with status: OUT_OF_RANGE: End of sequence
Out[15]:
[(<tf.Tensor: shape=(1, 4), dtype=int64, numpy=array([[ 4, 5, 2, 23]])>, <tf.Tensor: shape=(1, 4), dtype=int64, numpy=array([[ 5, 2, 23, 3]])>)]
In [16]:
length = 100
tf.random.set_seed(42)
train_set = to_dataset(encoded[:1_000_000], length=length, shuffle=True,
seed=42)
valid_set = to_dataset(encoded[1_000_000:1_060_000], length=length)
test_set = to_dataset(encoded[1_060_000:], length=length)
Building and Training the Char-RNN Model¶
Warning: the following code may one or two hours to run, depending on your GPU. Without a GPU, it may take over 24 hours. If you don't want to wait, just skip the next two code cells and run the code below to download a pretrained model.
Note: the GRU class will only use cuDNN acceleration (assuming you have a GPU) when using the default values for the following arguments: activation, recurrent_activation, recurrent_dropout, unroll, use_bias and reset_after.
In [17]:
tf.random.set_seed(42) # extra code – ensures reproducibility on CPU
model = tf.keras.Sequential([
tf.keras.layers.Embedding(input_dim=n_tokens, output_dim=16),
tf.keras.layers.GRU(128, return_sequences=True),
tf.keras.layers.Dense(n_tokens, activation="softmax")
])
model.compile(loss="sparse_categorical_crossentropy", optimizer="nadam",
metrics=["accuracy"])
model_ckpt = tf.keras.callbacks.ModelCheckpoint(
"my_shakespeare_model", monitor="val_accuracy", save_best_only=True)
history = model.fit(train_set, validation_data=valid_set, epochs=10,
callbacks=[model_ckpt])
Epoch 1/10
INFO:tensorflow:Assets written to: my_shakespeare_model/assets
31247/31247 [==============================] - 1407s 45ms/step - loss: 1.3873 - accuracy: 0.5754 - val_loss: 1.6155 - val_accuracy: 0.5333 Epoch 2/10
INFO:tensorflow:Assets written to: my_shakespeare_model/assets
31247/31247 [==============================] - 1376s 44ms/step - loss: 1.2921 - accuracy: 0.5973 - val_loss: 1.5881 - val_accuracy: 0.5401 Epoch 3/10
INFO:tensorflow:Assets written to: my_shakespeare_model/assets
31247/31247 [==============================] - 1379s 44ms/step - loss: 1.2743 - accuracy: 0.6015 - val_loss: 1.5885 - val_accuracy: 0.5407 Epoch 4/10
INFO:tensorflow:Assets written to: my_shakespeare_model/assets
31247/31247 [==============================] - 1381s 44ms/step - loss: 1.2654 - accuracy: 0.6031 - val_loss: 1.5701 - val_accuracy: 0.5418 Epoch 5/10
INFO:tensorflow:Assets written to: my_shakespeare_model/assets
31247/31247 [==============================] - 1379s 44ms/step - loss: 1.2594 - accuracy: 0.6045 - val_loss: 1.5674 - val_accuracy: 0.5450 Epoch 6/10
INFO:tensorflow:Assets written to: my_shakespeare_model/assets
31247/31247 [==============================] - 1386s 44ms/step - loss: 1.2545 - accuracy: 0.6058 - val_loss: 1.5587 - val_accuracy: 0.5492 Epoch 7/10 31247/31247 [==============================] - 1381s 44ms/step - loss: 1.2514 - accuracy: 0.6062 - val_loss: 1.5532 - val_accuracy: 0.5460 Epoch 8/10 31247/31247 [==============================] - 1381s 44ms/step - loss: 1.2485 - accuracy: 0.6067 - val_loss: 1.5522 - val_accuracy: 0.5479 Epoch 9/10
INFO:tensorflow:Assets written to: my_shakespeare_model/assets
31247/31247 [==============================] - 1382s 44ms/step - loss: 1.2460 - accuracy: 0.6073 - val_loss: 1.5521 - val_accuracy: 0.5497 Epoch 10/10
INFO:tensorflow:Assets written to: my_shakespeare_model/assets
31247/31247 [==============================] - 1385s 44ms/step - loss: 1.2436 - accuracy: 0.6080 - val_loss: 1.5477 - val_accuracy: 0.5513
In [18]:
shakespeare_model = tf.keras.Sequential([
text_vec_layer,
tf.keras.layers.Lambda(lambda X: X - 2), # no <PAD> or <UNK> tokens
model
])
If you don't want to wait for training to complete, I've pretrained a model for you. The following code will download it. Uncomment the last line if you want to use it instead of the model trained above.
In [19]:
# extra code – downloads a pretrained model
url = "https://github.com/ageron/data/raw/main/shakespeare_model.tgz"
path = tf.keras.utils.get_file("shakespeare_model.tgz", url, extract=True)
if "_extracted" in path:
model_path = Path(path) / "shakespeare_model"
else:
model_path = Path(path).with_name("shakespeare_model")
#shakespeare_model = tf.keras.models.load_model(model_path)
Downloading data from https://github.com/ageron/data/raw/main/shakespeare_model.tgz 352865/352865 ━━━━━━━━━━━━━━━━━━━━ 0s 0us/step
In [20]:
y_proba = shakespeare_model.predict(["To be or not to b"])[0, -1]
y_pred = tf.argmax(y_proba) # choose the most probable character ID
text_vec_layer.get_vocabulary()[y_pred + 2]
Out[20]:
'e'
Generating Fake Shakespearean Text¶
In [21]:
log_probas = tf.math.log([[0.5, 0.4, 0.1]]) # probas = 50%, 40%, and 10%
tf.random.set_seed(42)
tf.random.categorical(log_probas, num_samples=8) # draw 8 samples
Out[21]:
<tf.Tensor: shape=(1, 8), dtype=int64, numpy=array([[0, 1, 0, 2, 1, 0, 0, 1]])>
In [22]:
def next_char(text, temperature=1):
y_proba = shakespeare_model.predict([text])[0, -1:]
rescaled_logits = tf.math.log(y_proba) / temperature
char_id = tf.random.categorical(rescaled_logits, num_samples=1)[0, 0]
return text_vec_layer.get_vocabulary()[char_id + 2]
In [23]:
def extend_text(text, n_chars=50, temperature=1):
for _ in range(n_chars):
text += next_char(text, temperature)
return text
In [24]:
tf.random.set_seed(42) # extra code – ensures reproducibility on CPU
In [25]:
print(extend_text("To be or not to be", temperature=0.01))
To be or not to be the duke as it is a proper strange death, and the
In [26]:
print(extend_text("To be or not to be", temperature=1))
To be or not to behold? second push: gremio, lord all, a sistermen,
In [27]:
print(extend_text("To be or not to be", temperature=100))
To be or not to bef ,mt'&o3fpadm!$ wh!nse?bws3est--vgerdjw?c-y-ewznq
Stateful RNN¶
In [28]:
def to_dataset_for_stateful_rnn(sequence, length):
ds = tf.data.Dataset.from_tensor_slices(sequence)
ds = ds.window(length + 1, shift=length, drop_remainder=True)
ds = ds.flat_map(lambda window: window.batch(length + 1)).batch(1)
return ds.map(lambda window: (window[:, :-1], window[:, 1:])).prefetch(1)
stateful_train_set = to_dataset_for_stateful_rnn(encoded[:1_000_000], length)
stateful_valid_set = to_dataset_for_stateful_rnn(encoded[1_000_000:1_060_000],
length)
stateful_test_set = to_dataset_for_stateful_rnn(encoded[1_060_000:], length)
In [29]:
# extra code – simple example using to_dataset_for_stateful_rnn()
list(to_dataset_for_stateful_rnn(tf.range(10), 3))
Out[29]:
[(<tf.Tensor: shape=(1, 3), dtype=int32, numpy=array([[0, 1, 2]], dtype=int32)>, <tf.Tensor: shape=(1, 3), dtype=int32, numpy=array([[1, 2, 3]], dtype=int32)>), (<tf.Tensor: shape=(1, 3), dtype=int32, numpy=array([[3, 4, 5]], dtype=int32)>, <tf.Tensor: shape=(1, 3), dtype=int32, numpy=array([[4, 5, 6]], dtype=int32)>), (<tf.Tensor: shape=(1, 3), dtype=int32, numpy=array([[6, 7, 8]], dtype=int32)>, <tf.Tensor: shape=(1, 3), dtype=int32, numpy=array([[7, 8, 9]], dtype=int32)>)]
If you'd like to have more than one window per batch, you can use the to_batched_dataset_for_stateful_rnn() function instead of to_dataset_for_stateful_rnn():
In [30]:
# extra code – shows one way to prepare a batched dataset for a stateful RNN
import numpy as np
def to_non_overlapping_windows(sequence, length):
ds = tf.data.Dataset.from_tensor_slices(sequence)
ds = ds.window(length + 1, shift=length, drop_remainder=True)
return ds.flat_map(lambda window: window.batch(length + 1))
def to_batched_dataset_for_stateful_rnn(sequence, length, batch_size=32):
parts = np.array_split(sequence, batch_size)
datasets = tuple(to_non_overlapping_windows(part, length) for part in parts)
ds = tf.data.Dataset.zip(datasets).map(lambda *windows: tf.stack(windows))
return ds.map(lambda window: (window[:, :-1], window[:, 1:])).prefetch(1)
list(to_batched_dataset_for_stateful_rnn(tf.range(20), length=3, batch_size=2))
Out[30]:
[(<tf.Tensor: shape=(2, 3), dtype=int32, numpy=
array([[ 0, 1, 2],
[10, 11, 12]], dtype=int32)>,
<tf.Tensor: shape=(2, 3), dtype=int32, numpy=
array([[ 1, 2, 3],
[11, 12, 13]], dtype=int32)>),
(<tf.Tensor: shape=(2, 3), dtype=int32, numpy=
array([[ 3, 4, 5],
[13, 14, 15]], dtype=int32)>,
<tf.Tensor: shape=(2, 3), dtype=int32, numpy=
array([[ 4, 5, 6],
[14, 15, 16]], dtype=int32)>),
(<tf.Tensor: shape=(2, 3), dtype=int32, numpy=
array([[ 6, 7, 8],
[16, 17, 18]], dtype=int32)>,
<tf.Tensor: shape=(2, 3), dtype=int32, numpy=
array([[ 7, 8, 9],
[17, 18, 19]], dtype=int32)>)]
In [31]:
tf.random.set_seed(42) # extra code – ensures reproducibility on CPU
model = tf.keras.Sequential([
tf.keras.layers.Embedding(input_dim=n_tokens, output_dim=16,
batch_input_shape=[1, None]),
tf.keras.layers.GRU(128, return_sequences=True, stateful=True),
tf.keras.layers.Dense(n_tokens, activation="softmax")
])
In [32]:
class ResetStatesCallback(tf.keras.callbacks.Callback):
def on_epoch_begin(self, epoch, logs):
self.model.reset_states()
In [33]:
# extra code – use a different directory to save the checkpoints
model_ckpt = tf.keras.callbacks.ModelCheckpoint(
"my_stateful_shakespeare_model",
monitor="val_accuracy",
save_best_only=True)
Warning: the following cell will take a while to run (possibly an hour if you are not using a GPU).
In [34]:
model.compile(loss="sparse_categorical_crossentropy", optimizer="nadam",
metrics=["accuracy"])
history = model.fit(stateful_train_set, validation_data=stateful_valid_set,
epochs=10, callbacks=[ResetStatesCallback(), model_ckpt])
INFO:tensorflow:Assets written to: my_stateful_shakespeare_model/assets
9999/9999 [==============================] - 213s 21ms/step - loss: 1.8690 - accuracy: 0.4494 - val_loss: 1.7632 - val_accuracy: 0.4672 Epoch 2/10
INFO:tensorflow:Assets written to: my_stateful_shakespeare_model/assets
9999/9999 [==============================] - 211s 21ms/step - loss: 1.5635 - accuracy: 0.5284 - val_loss: 1.6334 - val_accuracy: 0.4994 Epoch 3/10
INFO:tensorflow:Assets written to: my_stateful_shakespeare_model/assets
9999/9999 [==============================] - 209s 21ms/step - loss: 1.4875 - accuracy: 0.5478 - val_loss: 1.5788 - val_accuracy: 0.5153 Epoch 4/10
INFO:tensorflow:Assets written to: my_stateful_shakespeare_model/assets
9999/9999 [==============================] - 208s 21ms/step - loss: 1.4483 - accuracy: 0.5579 - val_loss: 1.5471 - val_accuracy: 0.5236 Epoch 5/10
INFO:tensorflow:Assets written to: my_stateful_shakespeare_model/assets
9999/9999 [==============================] - 213s 21ms/step - loss: 1.4241 - accuracy: 0.5643 - val_loss: 1.5270 - val_accuracy: 0.5286 Epoch 6/10
INFO:tensorflow:Assets written to: my_stateful_shakespeare_model/assets
9999/9999 [==============================] - 215s 21ms/step - loss: 1.4074 - accuracy: 0.5686 - val_loss: 1.5109 - val_accuracy: 0.5338 Epoch 7/10
INFO:tensorflow:Assets written to: my_stateful_shakespeare_model/assets
9999/9999 [==============================] - 210s 21ms/step - loss: 1.3953 - accuracy: 0.5714 - val_loss: 1.5008 - val_accuracy: 0.5361 Epoch 8/10
INFO:tensorflow:Assets written to: my_stateful_shakespeare_model/assets
9999/9999 [==============================] - 212s 21ms/step - loss: 1.3863 - accuracy: 0.5737 - val_loss: 1.4938 - val_accuracy: 0.5381 Epoch 9/10 9999/9999 [==============================] - 207s 21ms/step - loss: 1.3790 - accuracy: 0.5757 - val_loss: 1.4890 - val_accuracy: 0.5380 Epoch 10/10
INFO:tensorflow:Assets written to: my_stateful_shakespeare_model/assets
9999/9999 [==============================] - 208s 21ms/step - loss: 1.3729 - accuracy: 0.5770 - val_loss: 1.4786 - val_accuracy: 0.5420
Extra Material: converting the stateful RNN to a stateless RNN and using it
To use the model with different batch sizes, we need to create a stateless copy:
In [35]:
stateless_model = tf.keras.Sequential([
tf.keras.layers.Embedding(input_dim=n_tokens, output_dim=16),
tf.keras.layers.GRU(128, return_sequences=True),
tf.keras.layers.Dense(n_tokens, activation="softmax")
])
To set the weights, we first need to build the model (so the weights get created):
In [36]:
stateless_model.build(tf.TensorShape([None, None]))
In [37]:
stateless_model.set_weights(model.get_weights())
In [38]:
shakespeare_model = tf.keras.Sequential([
text_vec_layer,
tf.keras.layers.Lambda(lambda X: X - 2), # no <PAD> or <UNK> tokens
stateless_model
])
In [39]:
tf.random.set_seed(42)
print(extend_text("to be or not to be", temperature=0.01))
to be or not to be so in the world and the strangeness to see the wo
Sentiment Analysis¶
In [40]:
import tensorflow_datasets as tfds
raw_train_set, raw_valid_set, raw_test_set = tfds.load(
name="imdb_reviews",
split=["train[:90%]", "train[90%:]", "test"],
as_supervised=True
)
tf.random.set_seed(42)
train_set = raw_train_set.shuffle(5000, seed=42).batch(32).prefetch(1)
valid_set = raw_valid_set.batch(32).prefetch(1)
test_set = raw_test_set.batch(32).prefetch(1)
Downloading and preparing dataset 80.23 MiB (download: 80.23 MiB, generated: Unknown size, total: 80.23 MiB) to /home/ageron/tensorflow_datasets/imdb_reviews/plain_text/1.0.0...
Dl Completed...: 0 url [00:00, ? url/s]
Dl Size...: 0 MiB [00:00, ? MiB/s]
Generating splits...: 0%| | 0/3 [00:00<?, ? splits/s]
Generating train examples...: 0%| | 0/25000 [00:00<?, ? examples/s]
Shuffling /home/ageron/tensorflow_datasets/imdb_reviews/plain_text/1.0.0.incomplete0WPKUH/imdb_reviews-train.t…
Generating test examples...: 0%| | 0/25000 [00:00<?, ? examples/s]
Shuffling /home/ageron/tensorflow_datasets/imdb_reviews/plain_text/1.0.0.incomplete0WPKUH/imdb_reviews-test.tf…
Generating unsupervised examples...: 0%| | 0/50000 [00:00<?, ? examples/s]
Shuffling /home/ageron/tensorflow_datasets/imdb_reviews/plain_text/1.0.0.incomplete0WPKUH/imdb_reviews-unsuper…
Dataset imdb_reviews downloaded and prepared to /home/ageron/tensorflow_datasets/imdb_reviews/plain_text/1.0.0. Subsequent calls will reuse this data.
In [41]:
for review, label in raw_train_set.take(4):
print(review.numpy().decode("utf-8")[:200], "...")
print("Label:", label.numpy())
This was an absolutely terrible movie. Don't be lured in by Christopher Walken or Michael Ironside. Both are great actors, but this must simply be their worst role in history. Even their great acting ... Label: 0 I have been known to fall asleep during films, but this is usually due to a combination of things including, really tired, being warm and comfortable on the sette and having just eaten a lot. However ... Label: 0 Mann photographs the Alberta Rocky Mountains in a superb fashion, and Jimmy Stewart and Walter Brennan give enjoyable performances as they always seem to do. <br /><br />But come on Hollywood - a Moun ... Label: 0 This is the kind of film for a snowy Sunday afternoon when the rest of the world can go ahead with its own business as you descend into a big arm-chair and mellow for a couple of hours. Wonderful perf ... Label: 1
In [42]:
vocab_size = 1000
text_vec_layer = tf.keras.layers.TextVectorization(max_tokens=vocab_size)
text_vec_layer.adapt(train_set.map(lambda reviews, labels: reviews))
Warning: the following cell will take a few minutes to run and the model will probably not learn anything because we didn't mask the padding tokens (that's the point of the next section).
In [43]:
embed_size = 128
tf.random.set_seed(42)
model = tf.keras.Sequential([
text_vec_layer,
tf.keras.layers.Embedding(vocab_size, embed_size),
tf.keras.layers.GRU(128),
tf.keras.layers.Dense(1, activation="sigmoid")
])
model.compile(loss="binary_crossentropy", optimizer="nadam",
metrics=["accuracy"])
history = model.fit(train_set, validation_data=valid_set, epochs=2)
Epoch 1/2 704/704 [==============================] - 255s 359ms/step - loss: 0.6934 - accuracy: 0.4990 - val_loss: 0.6931 - val_accuracy: 0.5016 Epoch 2/2 704/704 [==============================] - 250s 355ms/step - loss: 0.6934 - accuracy: 0.5042 - val_loss: 0.6942 - val_accuracy: 0.5008
Masking¶
Warning: the following cell will take a while to run (possibly 30 minutes if you are not using a GPU).
In [44]:
embed_size = 128
tf.random.set_seed(42)
model = tf.keras.Sequential([
text_vec_layer,
tf.keras.layers.Embedding(vocab_size, embed_size, mask_zero=True),
tf.keras.layers.GRU(128),
tf.keras.layers.Dense(1, activation="sigmoid")
])
model.compile(loss="binary_crossentropy", optimizer="nadam",
metrics=["accuracy"])
history = model.fit(train_set, validation_data=valid_set, epochs=5)
Epoch 1/5 704/704 [==============================] - 303s 426ms/step - loss: 0.5296 - accuracy: 0.7234 - val_loss: 0.4045 - val_accuracy: 0.8244 Epoch 2/5 704/704 [==============================] - 295s 419ms/step - loss: 0.3702 - accuracy: 0.8418 - val_loss: 0.3390 - val_accuracy: 0.8532 Epoch 3/5 704/704 [==============================] - 298s 423ms/step - loss: 0.3057 - accuracy: 0.8747 - val_loss: 0.3196 - val_accuracy: 0.8696 Epoch 4/5 704/704 [==============================] - 294s 418ms/step - loss: 0.2784 - accuracy: 0.8871 - val_loss: 0.3162 - val_accuracy: 0.8596 Epoch 5/5 704/704 [==============================] - 293s 417ms/step - loss: 0.2597 - accuracy: 0.8961 - val_loss: 0.3209 - val_accuracy: 0.8548
Or using manual masking:
In [45]:
tf.random.set_seed(42) # extra code – ensures reproducibility on the CPU
inputs = tf.keras.layers.Input(shape=[], dtype=tf.string)
token_ids = text_vec_layer(inputs)
mask = tf.math.not_equal(token_ids, 0)
Z = tf.keras.layers.Embedding(vocab_size, embed_size)(token_ids)
Z = tf.keras.layers.GRU(128, dropout=0.2)(Z, mask=mask)
outputs = tf.keras.layers.Dense(1, activation="sigmoid")(Z)
model = tf.keras.Model(inputs=[inputs], outputs=[outputs])
Warning: the following cell will take a while to run (possibly 30 minutes if you are not using a GPU).
In [46]:
# extra code – compiles and trains the model, as usual
model.compile(loss="binary_crossentropy", optimizer="nadam",
metrics=["accuracy"])
history = model.fit(train_set, validation_data=valid_set, epochs=5)
Epoch 1/5 704/704 [==============================] - 303s 427ms/step - loss: 0.5447 - accuracy: 0.7198 - val_loss: 0.4604 - val_accuracy: 0.7720 Epoch 2/5 704/704 [==============================] - 301s 427ms/step - loss: 0.3469 - accuracy: 0.8512 - val_loss: 0.3214 - val_accuracy: 0.8608 Epoch 3/5 704/704 [==============================] - 295s 419ms/step - loss: 0.3054 - accuracy: 0.8713 - val_loss: 0.3069 - val_accuracy: 0.8672 Epoch 4/5 704/704 [==============================] - 295s 420ms/step - loss: 0.2798 - accuracy: 0.8828 - val_loss: 0.3028 - val_accuracy: 0.8672 Epoch 5/5 704/704 [==============================] - 298s 423ms/step - loss: 0.2622 - accuracy: 0.8920 - val_loss: 0.2953 - val_accuracy: 0.8700
Extra material: using ragged tensors
In [47]:
text_vec_layer_ragged = tf.keras.layers.TextVectorization(
max_tokens=vocab_size, ragged=True)
text_vec_layer_ragged.adapt(train_set.map(lambda reviews, labels: reviews))
text_vec_layer_ragged(["Great movie!", "This is DiCaprio's best role."])
Out[47]:
<tf.RaggedTensor [[86, 18], [11, 7, 1, 116, 217]]>
In [48]:
text_vec_layer(["Great movie!", "This is DiCaprio's best role."])
Out[48]:
<tf.Tensor: shape=(2, 5), dtype=int64, numpy=
array([[ 86, 18, 0, 0, 0],
[ 11, 7, 1, 116, 217]])>
Warning: the following cell will take a while to run (possibly 30 minutes if you are not using a GPU).
In [49]:
embed_size = 128
tf.random.set_seed(42)
model = tf.keras.Sequential([
text_vec_layer_ragged,
tf.keras.layers.Embedding(vocab_size, embed_size),
tf.keras.layers.GRU(128),
tf.keras.layers.Dense(1, activation="sigmoid")
])
model.compile(loss="binary_crossentropy", optimizer="nadam",
metrics=["accuracy"])
history = model.fit(train_set, validation_data=valid_set, epochs=5)
Epoch 1/5 704/704 [==============================] - 280s 395ms/step - loss: 0.5038 - accuracy: 0.7496 - val_loss: 0.6706 - val_accuracy: 0.6752 Epoch 2/5 704/704 [==============================] - 277s 393ms/step - loss: 0.4499 - accuracy: 0.7892 - val_loss: 0.3494 - val_accuracy: 0.8500 Epoch 3/5 704/704 [==============================] - 276s 392ms/step - loss: 0.3270 - accuracy: 0.8592 - val_loss: 0.3855 - val_accuracy: 0.8260 Epoch 4/5 704/704 [==============================] - 277s 394ms/step - loss: 0.2935 - accuracy: 0.8760 - val_loss: 0.3401 - val_accuracy: 0.8520 Epoch 5/5 704/704 [==============================] - 275s 390ms/step - loss: 0.2742 - accuracy: 0.8854 - val_loss: 0.3971 - val_accuracy: 0.8208
Reusing Pretrained Embeddings and Language Models¶
Warning: the following cell will take a while to run (possibly an hour if you are not using a GPU).
In [50]:
import os
import tensorflow_hub as hub
os.environ["TFHUB_CACHE_DIR"] = "my_tfhub_cache"
tf.random.set_seed(42) # extra code – ensures reproducibility on CPU
model = tf.keras.Sequential([
hub.KerasLayer("https://tfhub.dev/google/universal-sentence-encoder/4",
trainable=True, dtype=tf.string, input_shape=[]),
tf.keras.layers.Dense(64, activation="relu"),
tf.keras.layers.Dense(1, activation="sigmoid")
])
model.compile(loss="binary_crossentropy", optimizer="nadam",
metrics=["accuracy"])
model.fit(train_set, validation_data=valid_set, epochs=10)
Epoch 1/10 704/704 [==============================] - 224s 303ms/step - loss: 0.3141 - accuracy: 0.8648 - val_loss: 0.2397 - val_accuracy: 0.9008 Epoch 2/10 704/704 [==============================] - 205s 291ms/step - loss: 0.0489 - accuracy: 0.9852 - val_loss: 0.3257 - val_accuracy: 0.8936 Epoch 3/10 704/704 [==============================] - 204s 290ms/step - loss: 0.0061 - accuracy: 0.9988 - val_loss: 0.3963 - val_accuracy: 0.8944 Epoch 4/10 704/704 [==============================] - 204s 290ms/step - loss: 9.4918e-04 - accuracy: 0.9999 - val_loss: 0.4291 - val_accuracy: 0.8924 Epoch 5/10 704/704 [==============================] - 203s 289ms/step - loss: 5.1920e-04 - accuracy: 1.0000 - val_loss: 0.4691 - val_accuracy: 0.8932 Epoch 6/10 704/704 [==============================] - 204s 289ms/step - loss: 5.0053e-04 - accuracy: 1.0000 - val_loss: 0.4687 - val_accuracy: 0.8912 Epoch 7/10 704/704 [==============================] - 208s 296ms/step - loss: 3.7360e-04 - accuracy: 1.0000 - val_loss: 0.5034 - val_accuracy: 0.8984 Epoch 8/10 704/704 [==============================] - 209s 297ms/step - loss: 2.3907e-05 - accuracy: 1.0000 - val_loss: 0.5773 - val_accuracy: 0.8924 Epoch 9/10 704/704 [==============================] - 204s 290ms/step - loss: 9.0970e-06 - accuracy: 1.0000 - val_loss: 0.6163 - val_accuracy: 0.8972 Epoch 10/10 704/704 [==============================] - 205s 291ms/step - loss: 5.2528e-06 - accuracy: 1.0000 - val_loss: 0.6455 - val_accuracy: 0.8956
Out[50]:
<keras.callbacks.History at 0x7f89897f6d30>
An Encoder–Decoder Network for Neural Machine Translation¶
In [51]:
url = "https://storage.googleapis.com/download.tensorflow.org/data/spa-eng.zip"
path = tf.keras.utils.get_file("spa-eng.zip", origin=url, cache_dir="datasets",
extract=True)
text = (Path(path).with_name("spa-eng") / "spa.txt").read_text()
In [52]:
import numpy as np
text = text.replace("¡", "").replace("¿", "")
pairs = [line.split("\t") for line in text.splitlines()]
np.random.seed(42) # extra code – ensures reproducibility on CPU
np.random.shuffle(pairs)
sentences_en, sentences_es = zip(*pairs) # separates the pairs into 2 lists
In [53]:
for i in range(3):
print(sentences_en[i], "=>", sentences_es[i])
How boring! => Qué aburrimiento! I love sports. => Adoro el deporte. Would you like to swap jobs? => Te gustaría que intercambiemos los trabajos?
In [54]:
vocab_size = 1000
max_length = 50
text_vec_layer_en = tf.keras.layers.TextVectorization(
vocab_size, output_sequence_length=max_length)
text_vec_layer_es = tf.keras.layers.TextVectorization(
vocab_size, output_sequence_length=max_length)
text_vec_layer_en.adapt(sentences_en)
text_vec_layer_es.adapt([f"startofseq {s} endofseq" for s in sentences_es])
In [55]:
text_vec_layer_en.get_vocabulary()[:10]
Out[55]:
['', '[UNK]', 'the', 'i', 'to', 'you', 'tom', 'a', 'is', 'he']
In [56]:
text_vec_layer_es.get_vocabulary()[:10]
Out[56]:
['', '[UNK]', 'startofseq', 'endofseq', 'de', 'que', 'a', 'no', 'tom', 'la']
In [57]:
X_train = tf.constant(sentences_en[:100_000])
X_valid = tf.constant(sentences_en[100_000:])
X_train_dec = tf.constant([f"startofseq {s}" for s in sentences_es[:100_000]])
X_valid_dec = tf.constant([f"startofseq {s}" for s in sentences_es[100_000:]])
Y_train = text_vec_layer_es([f"{s} endofseq" for s in sentences_es[:100_000]])
Y_valid = text_vec_layer_es([f"{s} endofseq" for s in sentences_es[100_000:]])
In [58]:
tf.random.set_seed(42) # extra code – ensures reproducibility on CPU
encoder_inputs = tf.keras.layers.Input(shape=[], dtype=tf.string)
decoder_inputs = tf.keras.layers.Input(shape=[], dtype=tf.string)
In [59]:
embed_size = 128
encoder_input_ids = text_vec_layer_en(encoder_inputs)
decoder_input_ids = text_vec_layer_es(decoder_inputs)
encoder_embedding_layer = tf.keras.layers.Embedding(vocab_size, embed_size,
mask_zero=True)
decoder_embedding_layer = tf.keras.layers.Embedding(vocab_size, embed_size,
mask_zero=True)
encoder_embeddings = encoder_embedding_layer(encoder_input_ids)
decoder_embeddings = decoder_embedding_layer(decoder_input_ids)
In [60]:
encoder = tf.keras.layers.LSTM(512, return_state=True)
encoder_outputs, *encoder_state = encoder(encoder_embeddings)
In [61]:
decoder = tf.keras.layers.LSTM(512, return_sequences=True)
decoder_outputs = decoder(decoder_embeddings, initial_state=encoder_state)
In [62]:
output_layer = tf.keras.layers.Dense(vocab_size, activation="softmax")
Y_proba = output_layer(decoder_outputs)
Warning: the following cell will take a while to run (possibly a couple hours if you are not using a GPU).
In [63]:
model = tf.keras.Model(inputs=[encoder_inputs, decoder_inputs],
outputs=[Y_proba])
model.compile(loss="sparse_categorical_crossentropy", optimizer="nadam",
metrics=["accuracy"])
model.fit((X_train, X_train_dec), Y_train, epochs=10,
validation_data=((X_valid, X_valid_dec), Y_valid))
Epoch 1/10 3125/3125 [==============================] - 698s 221ms/step - loss: 0.4154 - accuracy: 0.4256 - val_loss: 0.3069 - val_accuracy: 0.5246 Epoch 2/10 3125/3125 [==============================] - 686s 219ms/step - loss: 0.2631 - accuracy: 0.5745 - val_loss: 0.2367 - val_accuracy: 0.6055 Epoch 3/10 3125/3125 [==============================] - 686s 220ms/step - loss: 0.2066 - accuracy: 0.6457 - val_loss: 0.2061 - val_accuracy: 0.6500 Epoch 4/10 3125/3125 [==============================] - 682s 218ms/step - loss: 0.1740 - accuracy: 0.6907 - val_loss: 0.1920 - val_accuracy: 0.6691 Epoch 5/10 3125/3125 [==============================] - 676s 216ms/step - loss: 0.1507 - accuracy: 0.7237 - val_loss: 0.1865 - val_accuracy: 0.6767 Epoch 6/10 3125/3125 [==============================] - 675s 216ms/step - loss: 0.1316 - accuracy: 0.7522 - val_loss: 0.1847 - val_accuracy: 0.6804 Epoch 7/10 3125/3125 [==============================] - 675s 216ms/step - loss: 0.1154 - accuracy: 0.7774 - val_loss: 0.1866 - val_accuracy: 0.6822 Epoch 8/10 3125/3125 [==============================] - 673s 215ms/step - loss: 0.1011 - accuracy: 0.8007 - val_loss: 0.1907 - val_accuracy: 0.6829 Epoch 9/10 3125/3125 [==============================] - 673s 215ms/step - loss: 0.0888 - accuracy: 0.8215 - val_loss: 0.1961 - val_accuracy: 0.6792 Epoch 10/10 3125/3125 [==============================] - 673s 215ms/step - loss: 0.0782 - accuracy: 0.8402 - val_loss: 0.2027 - val_accuracy: 0.6763
Out[63]:
<keras.callbacks.History at 0x7f897878ac10>
In [64]:
def translate(sentence_en):
translation = ""
for word_idx in range(max_length):
X = np.array([sentence_en]) # encoder input
X_dec = np.array(["startofseq " + translation]) # decoder input
y_proba = model.predict((X, X_dec))[0, word_idx] # last token's probas
predicted_word_id = np.argmax(y_proba)
predicted_word = text_vec_layer_es.get_vocabulary()[predicted_word_id]
if predicted_word == "endofseq":
break
translation += " " + predicted_word
return translation.strip()
In [65]:
translate("I like soccer")
Out[65]:
'me gusta el fútbol'
Nice! However, the model struggles with longer sentences:
In [66]:
translate("I like soccer and also going to the beach")
Out[66]:
'me gusta el fútbol y a veces mismo al bus'
Bidirectional RNNs¶
To create a bidirectional recurrent layer, just wrap a regular recurrent layer in a Bidirectional layer:
In [67]:
tf.random.set_seed(42) # extra code – ensures reproducibility on CPU
encoder = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(256, return_state=True))
In [68]:
encoder_outputs, *encoder_state = encoder(encoder_embeddings)
encoder_state = [tf.concat(encoder_state[::2], axis=-1), # short-term (0 & 2)
tf.concat(encoder_state[1::2], axis=-1)] # long-term (1 & 3)
Warning: the following cell will take a while to run (possibly a couple hours if you are not using a GPU).
In [69]:
# extra code — completes the model and trains it
decoder = tf.keras.layers.LSTM(512, return_sequences=True)
decoder_outputs = decoder(decoder_embeddings, initial_state=encoder_state)
output_layer = tf.keras.layers.Dense(vocab_size, activation="softmax")
Y_proba = output_layer(decoder_outputs)
model = tf.keras.Model(inputs=[encoder_inputs, decoder_inputs],
outputs=[Y_proba])
model.compile(loss="sparse_categorical_crossentropy", optimizer="nadam",
metrics=["accuracy"])
model.fit((X_train, X_train_dec), Y_train, epochs=10,
validation_data=((X_valid, X_valid_dec), Y_valid))
Epoch 1/10 3125/3125 [==============================] - 574s 181ms/step - loss: 0.3075 - accuracy: 0.5393 - val_loss: 0.2192 - val_accuracy: 0.6319 Epoch 2/10 3125/3125 [==============================] - 564s 180ms/step - loss: 0.1916 - accuracy: 0.6689 - val_loss: 0.1880 - val_accuracy: 0.6731 Epoch 3/10 3125/3125 [==============================] - 566s 181ms/step - loss: 0.1602 - accuracy: 0.7119 - val_loss: 0.1751 - val_accuracy: 0.6916 Epoch 4/10 3125/3125 [==============================] - 566s 181ms/step - loss: 0.1395 - accuracy: 0.7415 - val_loss: 0.1715 - val_accuracy: 0.6979 Epoch 5/10 3125/3125 [==============================] - 566s 181ms/step - loss: 0.1227 - accuracy: 0.7666 - val_loss: 0.1707 - val_accuracy: 0.7025 Epoch 6/10 3125/3125 [==============================] - 567s 181ms/step - loss: 0.1085 - accuracy: 0.7887 - val_loss: 0.1730 - val_accuracy: 0.6995 Epoch 7/10 3125/3125 [==============================] - 571s 183ms/step - loss: 0.0961 - accuracy: 0.8089 - val_loss: 0.1764 - val_accuracy: 0.7000 Epoch 8/10 3125/3125 [==============================] - 567s 181ms/step - loss: 0.0852 - accuracy: 0.8273 - val_loss: 0.1821 - val_accuracy: 0.6981 Epoch 9/10 3125/3125 [==============================] - 565s 181ms/step - loss: 0.0759 - accuracy: 0.8438 - val_loss: 0.1881 - val_accuracy: 0.6956 Epoch 10/10 3125/3125 [==============================] - 565s 181ms/step - loss: 0.0682 - accuracy: 0.8577 - val_loss: 0.1951 - val_accuracy: 0.6906
Out[69]:
<keras.callbacks.History at 0x7f892d2d5fa0>
In [70]:
translate("I like soccer")
Out[70]:
'me gusta el fútbol'
Beam Search¶
This is a very basic implementation of beam search. I tried to make it readable and understandable, but it's definitely not optimized for speed! The function first uses the model to find the top k words to start the translations (where k is the beam width). For each of the top k translations, it evaluates the conditional probabilities of all possible words it could add to that translation. These extended translations and their probabilities are added to the list of candidates. Once we've gone through all top k translations and all words that could complete them, we keep only the top k candidates with the highest probability, and we iterate over and over until they all finish with an EOS token. The top translation is then returned (after removing its EOS token).
- Note: If p(S) is the probability of sentence S, and p(W|S) is the conditional probability of the word W given that the translation starts with S, then the probability of the sentence S' = concat(S, W) is p(S') = p(S) * p(W|S). As we add more words, the probability gets smaller and smaller. To avoid the risk of it getting too small, which could cause floating point precision errors, the function keeps track of log probabilities instead of probabilities: recall that log(a*b) = log(a) + log(b), therefore log(p(S')) = log(p(S)) + log(p(W|S)).
In [71]:
# extra code – a basic implementation of beam search
def beam_search(sentence_en, beam_width, verbose=False):
X = np.array([sentence_en]) # encoder input
X_dec = np.array(["startofseq"]) # decoder input
y_proba = model.predict((X, X_dec))[0, 0] # first token's probas
top_k = tf.math.top_k(y_proba, k=beam_width)
top_translations = [ # list of best (log_proba, translation)
(np.log(word_proba), text_vec_layer_es.get_vocabulary()[word_id])
for word_proba, word_id in zip(top_k.values, top_k.indices)
]
# extra code – displays the top first words in verbose mode
if verbose:
print("Top first words:", top_translations)
for idx in range(1, max_length):
candidates = []
for log_proba, translation in top_translations:
if translation.endswith("endofseq"):
candidates.append((log_proba, translation))
continue # translation is finished, so don't try to extend it
X = np.array([sentence_en]) # encoder input
X_dec = np.array(["startofseq " + translation]) # decoder input
y_proba = model.predict((X, X_dec))[0, idx] # last token's proba
for word_id, word_proba in enumerate(y_proba):
word = text_vec_layer_es.get_vocabulary()[word_id]
candidates.append((log_proba + np.log(word_proba),
f"{translation} {word}"))
top_translations = sorted(candidates, reverse=True)[:beam_width]
# extra code – displays the top translation so far in verbose mode
if verbose:
print("Top translations so far:", top_translations)
if all([tr.endswith("endofseq") for _, tr in top_translations]):
return top_translations[0][1].replace("endofseq", "").strip()
In [72]:
# extra code – shows how the model making an error
sentence_en = "I love cats and dogs"
translate(sentence_en)
Out[72]:
'me [UNK] los gatos y los gatos'
In [73]:
# extra code – shows how beam search can help
beam_search(sentence_en, beam_width=3, verbose=True)
Top first words: [(-0.012974381, 'me'), (-4.592527, '[UNK]'), (-6.314033, 'yo')] Top translations so far: [(-0.4831518, 'me [UNK]'), (-1.4920667, 'me encanta'), (-1.986235, 'me gustan')] Top translations so far: [(-0.6793061, 'me [UNK] los'), (-1.9889652, 'me gustan los'), (-2.0470557, 'me encanta los')] Top translations so far: [(-0.7609749, 'me [UNK] los gatos'), (-2.0677316, 'me gustan los gatos'), (-2.26029, 'me encanta los gatos')] Top translations so far: [(-0.76985043, 'me [UNK] los gatos y'), (-2.0701222, 'me gustan los gatos y'), (-2.2649746, 'me encanta los gatos y')] Top translations so far: [(-0.81283045, 'me [UNK] los gatos y los'), (-2.118244, 'me gustan los gatos y los'), (-2.96167, 'me encanta los gatos y los')] Top translations so far: [(-1.2259341, 'me [UNK] los gatos y los gatos'), (-1.9556838, 'me [UNK] los gatos y los perros'), (-2.7524388, 'me gustan los gatos y los perros')] Top translations so far: [(-1.2261332, 'me [UNK] los gatos y los gatos endofseq'), (-1.9560521, 'me [UNK] los gatos y los perros endofseq'), (-2.7566314, 'me gustan los gatos y los perros endofseq')]
Out[73]:
'me [UNK] los gatos y los gatos'
The correct translation is in the top 3 sentences found by beam search, but it's not the first. Since we're using a small vocabulary, the [UNK] token is quite frequent, so you may want to penalize it (e.g., divide its probability by 2 in the beam search function): this will discourage beam search from using it too much.
Attention Mechanisms¶
We need to feed all the encoder's outputs to the Attention layer, so we must add return_sequences=True to the encoder:
In [74]:
tf.random.set_seed(42) # extra code – ensures reproducibility on CPU
encoder = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(256, return_sequences=True, return_state=True))
In [75]:
# extra code – this part of the model is exactly the same as earlier
encoder_outputs, *encoder_state = encoder(encoder_embeddings)
encoder_state = [tf.concat(encoder_state[::2], axis=-1), # short-term (0 & 2)
tf.concat(encoder_state[1::2], axis=-1)] # long-term (1 & 3)
decoder = tf.keras.layers.LSTM(512, return_sequences=True)
decoder_outputs = decoder(decoder_embeddings, initial_state=encoder_state)
And finally, let's add the Attention layer and the output layer:
In [76]:
attention_layer = tf.keras.layers.Attention()
attention_outputs = attention_layer([decoder_outputs, encoder_outputs])
output_layer = tf.keras.layers.Dense(vocab_size, activation="softmax")
Y_proba = output_layer(attention_outputs)
Warning: the following cell will take a while to run (possibly a couple hours if you are not using a GPU).
In [77]:
model = tf.keras.Model(inputs=[encoder_inputs, decoder_inputs],
outputs=[Y_proba])
model.compile(loss="sparse_categorical_crossentropy", optimizer="nadam",
metrics=["accuracy"])
model.fit((X_train, X_train_dec), Y_train, epochs=10,
validation_data=((X_valid, X_valid_dec), Y_valid))
Epoch 1/10 3125/3125 [==============================] - 597s 189ms/step - loss: 0.3074 - accuracy: 0.5469 - val_loss: 0.2106 - val_accuracy: 0.6487 Epoch 2/10 3125/3125 [==============================] - 585s 187ms/step - loss: 0.1902 - accuracy: 0.6789 - val_loss: 0.1865 - val_accuracy: 0.6830 Epoch 3/10 3125/3125 [==============================] - 585s 187ms/step - loss: 0.1659 - accuracy: 0.7123 - val_loss: 0.1759 - val_accuracy: 0.7005 Epoch 4/10 3125/3125 [==============================] - 584s 187ms/step - loss: 0.1493 - accuracy: 0.7359 - val_loss: 0.1728 - val_accuracy: 0.7060 Epoch 5/10 3125/3125 [==============================] - 582s 186ms/step - loss: 0.1358 - accuracy: 0.7548 - val_loss: 0.1724 - val_accuracy: 0.7084 Epoch 6/10 3125/3125 [==============================] - 583s 186ms/step - loss: 0.1245 - accuracy: 0.7712 - val_loss: 0.1738 - val_accuracy: 0.7103 Epoch 7/10 3125/3125 [==============================] - 582s 186ms/step - loss: 0.1148 - accuracy: 0.7863 - val_loss: 0.1770 - val_accuracy: 0.7111 Epoch 8/10 3125/3125 [==============================] - 582s 186ms/step - loss: 0.1064 - accuracy: 0.7992 - val_loss: 0.1806 - val_accuracy: 0.7110 Epoch 9/10 3125/3125 [==============================] - 582s 186ms/step - loss: 0.0991 - accuracy: 0.8101 - val_loss: 0.1862 - val_accuracy: 0.7088 Epoch 10/10 3125/3125 [==============================] - 581s 186ms/step - loss: 0.0929 - accuracy: 0.8205 - val_loss: 0.1903 - val_accuracy: 0.7077
Out[77]:
<keras.callbacks.History at 0x7f87e5c8ad90>
In [78]:
translate("I like soccer and also going to the beach")
Out[78]:
'me gusta el fútbol y también ir a la playa'
In [79]:
beam_search("I like soccer and also going to the beach", beam_width=3,
verbose=True)
Top first words: [(-0.26210824, 'me'), (-2.553061, 'prefiero'), (-3.2005944, 'yo')] Top translations so far: [(-0.32478744, 'me gusta'), (-3.0608056, 'prefiero el'), (-3.1685317, 'me gustan')] Top translations so far: [(-0.7464272, 'me gusta el'), (-2.4712462, 'me gusta fútbol'), (-2.9149299, 'me gusta al')] Top translations so far: [(-1.0369574, 'me gusta el fútbol'), (-2.3301778, 'me gusta el el'), (-2.9658434, 'me gusta fútbol y')] Top translations so far: [(-1.0404125, 'me gusta el fútbol y'), (-2.5983238, 'me gusta el el fútbol'), (-2.9736564, 'me gusta fútbol y también')] Top translations so far: [(-1.0520902, 'me gusta el fútbol y también'), (-2.6003318, 'me gusta el el fútbol y'), (-3.128903, 'me gusta fútbol y también me')] Top translations so far: [(-1.9568634, 'me gusta el fútbol y también ir'), (-2.6169589, 'me gusta el el fútbol y también'), (-2.6949644, 'me gusta el fútbol y también fuera')] Top translations so far: [(-1.9676423, 'me gusta el fútbol y también ir a'), (-2.8482866, 'me gusta el fútbol y también fuera a'), (-3.7197533, 'me gusta el el fútbol y también ir')] Top translations so far: [(-1.9692448, 'me gusta el fútbol y también ir a la'), (-2.8501132, 'me gusta el fútbol y también fuera a la'), (-3.7309551, 'me gusta el el fútbol y también ir a')] Top translations so far: [(-1.9733216, 'me gusta el fútbol y también ir a la playa'), (-2.851697, 'me gusta el fútbol y también fuera a la playa'), (-3.7333717, 'me gusta el el fútbol y también ir a la')] Top translations so far: [(-1.9737166, 'me gusta el fútbol y también ir a la playa endofseq'), (-2.8547554, 'me gusta el fútbol y también fuera a la playa endofseq'), (-3.737218, 'me gusta el el fútbol y también ir a la playa')] Top translations so far: [(-1.9737166, 'me gusta el fútbol y también ir a la playa endofseq'), (-2.8547554, 'me gusta el fútbol y también fuera a la playa endofseq'), (-3.7375438, 'me gusta el el fútbol y también ir a la playa endofseq')]
Out[79]:
'me gusta el fútbol y también ir a la playa'
In [80]:
max_length = 50 # max length in the whole training set
embed_size = 128
tf.random.set_seed(42) # extra code – ensures reproducibility on CPU
pos_embed_layer = tf.keras.layers.Embedding(max_length, embed_size)
batch_max_len_enc = tf.shape(encoder_embeddings)[1]
encoder_in = encoder_embeddings + pos_embed_layer(tf.range(batch_max_len_enc))
batch_max_len_dec = tf.shape(decoder_embeddings)[1]
decoder_in = decoder_embeddings + pos_embed_layer(tf.range(batch_max_len_dec))
Alternatively, we can use fixed, non-trainable positional encodings:
In [81]:
class PositionalEncoding(tf.keras.layers.Layer):
def __init__(self, max_length, embed_size, dtype=tf.float32, **kwargs):
super().__init__(dtype=dtype, **kwargs)
assert embed_size % 2 == 0, "embed_size must be even"
p, i = np.meshgrid(np.arange(max_length),
2 * np.arange(embed_size // 2))
pos_emb = np.empty((1, max_length, embed_size))
pos_emb[0, :, ::2] = np.sin(p / 10_000 ** (i / embed_size)).T
pos_emb[0, :, 1::2] = np.cos(p / 10_000 ** (i / embed_size)).T
self.pos_encodings = tf.constant(pos_emb.astype(self.dtype))
self.supports_masking = True
def call(self, inputs):
batch_max_length = tf.shape(inputs)[1]
return inputs + self.pos_encodings[:, :batch_max_length]
In [82]:
pos_embed_layer = PositionalEncoding(max_length, embed_size)
encoder_in = pos_embed_layer(encoder_embeddings)
decoder_in = pos_embed_layer(decoder_embeddings)
In [83]:
# extra code – this cells generates and saves Figure 16–9
figure_max_length = 201
figure_embed_size = 512
pos_emb = PositionalEncoding(figure_max_length, figure_embed_size)
zeros = np.zeros((1, figure_max_length, figure_embed_size), np.float32)
P = pos_emb(zeros)[0].numpy()
i1, i2, crop_i = 100, 101, 150
p1, p2, p3 = 22, 60, 35
fig, (ax1, ax2) = plt.subplots(nrows=2, ncols=1, sharex=True, figsize=(9, 5))
ax1.plot([p1, p1], [-1, 1], "k--", label="$p = {}$".format(p1))
ax1.plot([p2, p2], [-1, 1], "k--", label="$p = {}$".format(p2), alpha=0.5)
ax1.plot(p3, P[p3, i1], "bx", label="$p = {}$".format(p3))
ax1.plot(P[:,i1], "b-", label="$i = {}$".format(i1))
ax1.plot(P[:,i2], "r-", label="$i = {}$".format(i2))
ax1.plot([p1, p2], [P[p1, i1], P[p2, i1]], "bo")
ax1.plot([p1, p2], [P[p1, i2], P[p2, i2]], "ro")
ax1.legend(loc="center right", fontsize=14, framealpha=0.95)
ax1.set_ylabel("$P_{(p,i)}$", rotation=0, fontsize=16)
ax1.grid(True, alpha=0.3)
ax1.hlines(0, 0, figure_max_length - 1, color="k", linewidth=1, alpha=0.3)
ax1.axis([0, figure_max_length - 1, -1, 1])
ax2.imshow(P.T[:crop_i], cmap="gray", interpolation="bilinear", aspect="auto")
ax2.hlines(i1, 0, figure_max_length - 1, color="b", linewidth=3)
cheat = 2 # need to raise the red line a bit, or else it hides the blue one
ax2.hlines(i2+cheat, 0, figure_max_length - 1, color="r", linewidth=3)
ax2.plot([p1, p1], [0, crop_i], "k--")
ax2.plot([p2, p2], [0, crop_i], "k--", alpha=0.5)
ax2.plot([p1, p2], [i2+cheat, i2+cheat], "ro")
ax2.plot([p1, p2], [i1, i1], "bo")
ax2.axis([0, figure_max_length - 1, 0, crop_i])
ax2.set_xlabel("$p$", fontsize=16)
ax2.set_ylabel("$i$", rotation=0, fontsize=16)
save_fig("positional_embedding_plot")
plt.show()
Multi-Head Attention¶
In [84]:
N = 2 # instead of 6
num_heads = 8
dropout_rate = 0.1
n_units = 128 # for the first Dense layer in each Feed Forward block
encoder_pad_mask = tf.math.not_equal(encoder_input_ids, 0)[:, tf.newaxis]
Z = encoder_in
for _ in range(N):
skip = Z
attn_layer = tf.keras.layers.MultiHeadAttention(
num_heads=num_heads, key_dim=embed_size, dropout=dropout_rate)
Z = attn_layer(Z, value=Z, attention_mask=encoder_pad_mask)
Z = tf.keras.layers.LayerNormalization()(tf.keras.layers.Add()([Z, skip]))
skip = Z
Z = tf.keras.layers.Dense(n_units, activation="relu")(Z)
Z = tf.keras.layers.Dense(embed_size)(Z)
Z = tf.keras.layers.Dropout(dropout_rate)(Z)
Z = tf.keras.layers.LayerNormalization()(tf.keras.layers.Add()([Z, skip]))
In [85]:
decoder_pad_mask = tf.math.not_equal(decoder_input_ids, 0)[:, tf.newaxis]
causal_mask = tf.linalg.band_part( # creates a lower triangular matrix
tf.ones((batch_max_len_dec, batch_max_len_dec), tf.bool), -1, 0)
In [86]:
encoder_outputs = Z # let's save the encoder's final outputs
Z = decoder_in # the decoder starts with its own inputs
for _ in range(N):
skip = Z
attn_layer = tf.keras.layers.MultiHeadAttention(
num_heads=num_heads, key_dim=embed_size, dropout=dropout_rate)
Z = attn_layer(Z, value=Z, attention_mask=causal_mask & decoder_pad_mask)
Z = tf.keras.layers.LayerNormalization()(tf.keras.layers.Add()([Z, skip]))
skip = Z
attn_layer = tf.keras.layers.MultiHeadAttention(
num_heads=num_heads, key_dim=embed_size, dropout=dropout_rate)
Z = attn_layer(Z, value=encoder_outputs, attention_mask=encoder_pad_mask)
Z = tf.keras.layers.LayerNormalization()(tf.keras.layers.Add()([Z, skip]))
skip = Z
Z = tf.keras.layers.Dense(n_units, activation="relu")(Z)
Z = tf.keras.layers.Dense(embed_size)(Z)
Z = tf.keras.layers.LayerNormalization()(tf.keras.layers.Add()([Z, skip]))
Warning: the following cell will take a while to run (possibly 2 or 3 hours if you are not using a GPU).
In [87]:
Y_proba = tf.keras.layers.Dense(vocab_size, activation="softmax")(Z)
model = tf.keras.Model(inputs=[encoder_inputs, decoder_inputs],
outputs=[Y_proba])
model.compile(loss="sparse_categorical_crossentropy", optimizer="nadam",
metrics=["accuracy"])
model.fit((X_train, X_train_dec), Y_train, epochs=10,
validation_data=((X_valid, X_valid_dec), Y_valid))
Epoch 1/10 3125/3125 [==============================] - 828s 263ms/step - loss: 0.2982 - accuracy: 0.5545 - val_loss: 0.2105 - val_accuracy: 0.6476 Epoch 2/10 3125/3125 [==============================] - 820s 262ms/step - loss: 0.2006 - accuracy: 0.6601 - val_loss: 0.1876 - val_accuracy: 0.6802 Epoch 3/10 3125/3125 [==============================] - 820s 263ms/step - loss: 0.1842 - accuracy: 0.6816 - val_loss: 0.1766 - val_accuracy: 0.6975 Epoch 4/10 3125/3125 [==============================] - 820s 262ms/step - loss: 0.1748 - accuracy: 0.6942 - val_loss: 0.1704 - val_accuracy: 0.7055 Epoch 5/10 3125/3125 [==============================] - 820s 262ms/step - loss: 0.1683 - accuracy: 0.7021 - val_loss: 0.1657 - val_accuracy: 0.7102 Epoch 6/10 3125/3125 [==============================] - 821s 263ms/step - loss: 0.1628 - accuracy: 0.7096 - val_loss: 0.1628 - val_accuracy: 0.7130 Epoch 7/10 3125/3125 [==============================] - 826s 264ms/step - loss: 0.1588 - accuracy: 0.7154 - val_loss: 0.1595 - val_accuracy: 0.7205 Epoch 8/10 3125/3125 [==============================] - 822s 263ms/step - loss: 0.1550 - accuracy: 0.7205 - val_loss: 0.1590 - val_accuracy: 0.7199 Epoch 9/10 3125/3125 [==============================] - 821s 263ms/step - loss: 0.1518 - accuracy: 0.7249 - val_loss: 0.1547 - val_accuracy: 0.7258 Epoch 10/10 3125/3125 [==============================] - 821s 263ms/step - loss: 0.1492 - accuracy: 0.7279 - val_loss: 0.1538 - val_accuracy: 0.7281
Out[87]:
<keras.callbacks.History at 0x7f8946cdf9a0>
In [88]:
translate("I like soccer and also going to the beach")
Out[88]:
'me gusta el fútbol y yo también voy a la playa'
Hugging Face¶
In [89]:
from transformers import pipeline
classifier = pipeline("sentiment-analysis") # many other tasks are available
result = classifier("The actors were very convincing.")
No model was supplied, defaulted to distilbert/distilbert-base-uncased-finetuned-sst-2-english and revision 714eb0f (https://huggingface.co/distilbert/distilbert-base-uncased-finetuned-sst-2-english). Using a pipeline without specifying a model name and revision in production is not recommended. Device set to use cuda:0
Models can be very biased. For example, it may like or dislike some countries depending on the data it was trained on, and how it is used, so use it with care:
In [90]:
classifier(["I am from India.", "I am from Iraq."])
Out[90]:
[{'label': 'POSITIVE', 'score': 0.9896161556243896},
{'label': 'NEGATIVE', 'score': 0.9811071157455444}]
In [91]:
model_name = "huggingface/distilbert-base-uncased-finetuned-mnli"
classifier_mnli = pipeline("text-classification", model=model_name)
classifier_mnli("She loves me. [SEP] She loves me not.")
Device set to use cuda:0
Out[91]:
[{'label': 'contradiction', 'score': 0.9790192246437073}]
In [92]:
from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
WARNING: Sadly, the Transformers library has dropped support for TensorFlow and JAX. It only supports PyTorch now, so I have updated the code below to use PyTorch instead of TensorFlow (if you installed this project locally, you will need to install PyTorch). For more PyTorch code examples, please check out the PyTorch version of this repository at github.com/ageron/handson-mlp. These notebooks accompany the PyTorch version of this book, published in October 2025 by O'Reilly.
In [93]:
token_ids = tokenizer(["I like soccer. [SEP] We all love soccer!",
"Joe lived for a very long time. [SEP] Joe is old."],
padding=True, return_tensors="pt")
token_ids
Out[93]:
{'input_ids': tensor([[ 101, 1045, 2066, 4715, 1012, 102, 2057, 2035, 2293, 4715, 999, 102,
0, 0, 0],
[ 101, 3533, 2973, 2005, 1037, 2200, 2146, 2051, 1012, 102, 3533, 2003,
2214, 1012, 102]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])}
In [94]:
token_ids = tokenizer([("I like soccer.", "We all love soccer!"),
("Joe lived for a very long time.", "Joe is old.")],
padding=True, return_tensors="pt")
token_ids
Out[94]:
{'input_ids': tensor([[ 101, 1045, 2066, 4715, 1012, 102, 2057, 2035, 2293, 4715, 999, 102,
0, 0, 0],
[ 101, 3533, 2973, 2005, 1037, 2200, 2146, 2051, 1012, 102, 3533, 2003,
2214, 1012, 102]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])}
In [95]:
import torch
with torch.no_grad():
outputs = model(**token_ids)
outputs
Out[95]:
SequenceClassifierOutput(loss=None, logits=tensor([[-2.1124, 1.1787, 1.4101],
[-0.0148, 1.0962, -0.9920]]), hidden_states=None, attentions=None)
In [96]:
Y_logits = tf.constant(outputs.logits)
Y_logits
Out[96]:
<tf.Tensor: shape=(2, 3), dtype=float32, numpy=
array([[-2.1123817, 1.17868 , 1.4101001],
[-0.0147747, 1.0962385, -0.9919974]], dtype=float32)>
In [97]:
Y_probas = tf.keras.activations.softmax(Y_logits)
Y_probas
Out[97]:
<tf.Tensor: shape=(2, 3), dtype=float32, numpy=
array([[0.01619702, 0.43523622, 0.54856676],
[0.2265627 , 0.68816936, 0.08526798]], dtype=float32)>
In [98]:
Y_pred = tf.argmax(Y_probas, axis=1)
Y_pred # 0 = contradiction, 1 = entailment, 2 = neutral
Out[98]:
<tf.Tensor: shape=(2,), dtype=int64, numpy=array([2, 1])>
This TensorFlow code no longer works, so I commented it out:
In [99]:
# sentences = [("Sky is blue", "Sky is red"), ("I love her", "She loves me")]
# X_train = tokenizer(sentences, padding=True, return_tensors="tf").data
# y_train = tf.constant([0, 2]) # contradiction, neutral
# loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
# model.compile(loss=loss, optimizer="nadam", metrics=["accuracy"])
# history = model.fit(X_train, y_train, epochs=2)
The following code does the same thing using PyTorch and the Transformer library's Trainer API:
In [100]:
from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments
from torch.utils.data import TensorDataset
import torch
sentences = [("Sky is blue", "Sky is red"), ("I love her", "She loves me")]
X_train = tokenizer(sentences, padding=True, return_tensors="pt").data
y_train = torch.tensor([0, 2]) # 0=contradiction, 2=neutral
dataset = TensorDataset(X_train["input_ids"], X_train["attention_mask"],
y_train)
def collate_fn(batch):
input_ids, attention_mask, labels = zip(*batch)
return {
"input_ids": torch.stack(input_ids),
"attention_mask": torch.stack(attention_mask),
"labels": torch.stack(labels)
}
args = TrainingArguments(output_dir="./results", num_train_epochs=2,
per_device_train_batch_size=2, report_to="none")
trainer = Trainer(model=model, args=args, train_dataset=dataset,
data_collator=collate_fn)
trainer.train()
[2/2 00:21, Epoch 2/2]
| Step | Training Loss |
|---|
Out[100]:
TrainOutput(global_step=2, training_loss=0.0020126409363001585, metrics={'train_runtime': 21.8279, 'train_samples_per_second': 0.183, 'train_steps_per_second': 0.092, 'total_flos': 9314280072.0, 'train_loss': 0.0020126409363001585, 'epoch': 2.0})
Exercise solutions¶
1. to 7.¶
- Stateless RNNs can only capture patterns whose length is less than, or equal to, the size of the windows the RNN is trained on. Conversely, stateful RNNs can capture longer-term patterns. However, implementing a stateful RNN is much harder—especially preparing the dataset properly. Moreover, stateful RNNs do not always work better, in part because consecutive batches are not independent and identically distributed (IID). Gradient Descent is not fond of non-IID datasets.
- In general, if you translate a sentence one word at a time, the result will be terrible. For example, the French sentence "Je vous en prie" means "You are welcome," but if you translate it one word at a time, you get "I you in pray." Huh? It is much better to read the whole sentence first and then translate it. A plain sequence-to-sequence RNN would start translating a sentence immediately after reading the first word, while an Encoder–Decoder RNN will first read the whole sentence and then translate it. That said, one could imagine a plain sequence-to-sequence RNN that would output silence whenever it is unsure about what to say next (just like human translators do when they must translate a live broadcast).
- Variable-length input sequences can be handled by padding the shorter sequences so that all sequences in a batch have the same length, and using masking to ensure the RNN ignores the padding token. For better performance, you may also want to create batches containing sequences of similar sizes. Ragged tensors can hold sequences of variable lengths, and Keras now supports them, which simplifies handling variable-length input sequences (at the time of this writing, it still does not handle ragged tensors as targets on the GPU, though). Regarding variable-length output sequences, if the length of the output sequence is known in advance (e.g., if you know that it is the same as the input sequence), then you just need to configure the loss function so that it ignores tokens that come after the end of the sequence. Similarly, the code that will use the model should ignore tokens beyond the end of the sequence. But generally the length of the output sequence is not known ahead of time, so the solution is to train the model so that it outputs an end-of-sequence token at the end of each sequence.
- Beam search is a technique used to improve the performance of a trained Encoder–Decoder model, for example in a neural machine translation system. The algorithm keeps track of a short list of the k most promising output sentences (say, the top three), and at each decoder step it tries to extend them by one word; then it keeps only the k most likely sentences. The parameter k is called the beam width: the larger it is, the more CPU and RAM will be used, but also the more accurate the system will be. Instead of greedily choosing the most likely next word at each step to extend a single sentence, this technique allows the system to explore several promising sentences simultaneously. Moreover, this technique lends itself well to parallelization. You can implement beam search by writing a custom memory cell. Alternatively, TensorFlow Addons's seq2seq API provides an implementation.
- An attention mechanism is a technique initially used in Encoder–Decoder models to give the decoder more direct access to the input sequence, allowing it to deal with longer input sequences. At each decoder time step, the current decoder's state and the full output of the encoder are processed by an alignment model that outputs an alignment score for each input time step. This score indicates which part of the input is most relevant to the current decoder time step. The weighted sum of the encoder output (weighted by their alignment score) is then fed to the decoder, which produces the next decoder state and the output for this time step. The main benefit of using an attention mechanism is the fact that the Encoder–Decoder model can successfully process longer input sequences. Another benefit is that the alignment scores make the model easier to debug and interpret: for example, if the model makes a mistake, you can look at which part of the input it was paying attention to, and this can help diagnose the issue. An attention mechanism is also at the core of the Transformer architecture, in the Multi-Head Attention layers. See the next answer.
- The most important layer in the Transformer architecture is the Multi-Head Attention layer (the original Transformer architecture contains 18 of them, including 6 Masked Multi-Head Attention layers). It is at the core of language models such as BERT and GPT-2. Its purpose is to allow the model to identify which words are most aligned with each other, and then improve each word's representation using these contextual clues.
- Sampled softmax is used when training a classification model when there are many classes (e.g., thousands). It computes an approximation of the cross-entropy loss based on the logit predicted by the model for the correct class, and the predicted logits for a sample of incorrect words. This speeds up training considerably compared to computing the softmax over all logits and then estimating the cross-entropy loss. After training, the model can be used normally, using the regular softmax function to compute all the class probabilities based on all the logits.
8.¶
Exercise: Embedded Reber grammars were used by Hochreiter and Schmidhuber in their paper about LSTMs. They are artificial grammars that produce strings such as "BPBTSXXVPSEPE." Check out Jenny Orr's nice introduction to this topic. Choose a particular embedded Reber grammar (such as the one represented on Jenny Orr's page), then train an RNN to identify whether a string respects that grammar or not. You will first need to write a function capable of generating a training batch containing about 50% strings that respect the grammar, and 50% that don't.
First we need to build a function that generates strings based on a grammar. The grammar will be represented as a list of possible transitions for each state. A transition specifies the string to output (or a grammar to generate it) and the next state.
In [101]:
default_reber_grammar = [
[("B", 1)], # (state 0) =B=>(state 1)
[("T", 2), ("P", 3)], # (state 1) =T=>(state 2) or =P=>(state 3)
[("S", 2), ("X", 4)], # (state 2) =S=>(state 2) or =X=>(state 4)
[("T", 3), ("V", 5)], # and so on...
[("X", 3), ("S", 6)],
[("P", 4), ("V", 6)],
[("E", None)]] # (state 6) =E=>(terminal state)
embedded_reber_grammar = [
[("B", 1)],
[("T", 2), ("P", 3)],
[(default_reber_grammar, 4)],
[(default_reber_grammar, 5)],
[("T", 6)],
[("P", 6)],
[("E", None)]]
def generate_string(grammar):
state = 0
output = []
while state is not None:
index = np.random.randint(len(grammar[state]))
production, state = grammar[state][index]
if isinstance(production, list):
production = generate_string(grammar=production)
output.append(production)
return "".join(output)
Let's generate a few strings based on the default Reber grammar:
In [102]:
np.random.seed(42)
for _ in range(25):
print(generate_string(default_reber_grammar), end=" ")
BTXXTTVPXTVPXTTVPSE BPVPSE BTXSE BPVVE BPVVE BTSXSE BPTVPXTTTVVE BPVVE BTXSE BTXXVPSE BPTTTTTTTTVVE BTXSE BPVPSE BTXSE BPTVPSE BTXXTVPSE BPVVE BPVVE BPVVE BPTTVVE BPVVE BPVVE BTXXVVE BTXXVVE BTXXVPXVVE
Looks good. Now let's generate a few strings based on the embedded Reber grammar:
In [103]:
np.random.seed(42)
for _ in range(25):
print(generate_string(embedded_reber_grammar), end=" ")
BTBPTTTVPXTVPXTTVPSETE BPBPTVPSEPE BPBPVVEPE BPBPVPXVVEPE BPBTXXTTTTVVEPE BPBPVPSEPE BPBTXXVPSEPE BPBTSSSSSSSXSEPE BTBPVVETE BPBTXXVVEPE BPBTXXVPSEPE BTBTXXVVETE BPBPVVEPE BPBPVVEPE BPBTSXSEPE BPBPVVEPE BPBPTVPSEPE BPBTXXVVEPE BTBPTVPXVVETE BTBPVVETE BTBTSSSSSSSXXVVETE BPBTSSSXXTTTTVPSEPE BTBPTTVVETE BPBTXXTVVEPE BTBTXSETE
Okay, now we need a function to generate strings that do not respect the grammar. We could generate a random string, but the task would be a bit too easy, so instead we will generate a string that respects the grammar, and we will corrupt it by changing just one character:
In [104]:
POSSIBLE_CHARS = "BEPSTVX"
def generate_corrupted_string(grammar, chars=POSSIBLE_CHARS):
good_string = generate_string(grammar)
index = np.random.randint(len(good_string))
good_char = good_string[index]
bad_char = np.random.choice(sorted(set(chars) - set(good_char)))
return good_string[:index] + bad_char + good_string[index + 1:]
Let's look at a few corrupted strings:
In [105]:
np.random.seed(42)
for _ in range(25):
print(generate_corrupted_string(embedded_reber_grammar), end=" ")
BTBPTTTPPXTVPXTTVPSETE BPBTXEEPE BPBPTVVVEPE BPBTSSSSXSETE BPTTXSEPE BTBPVPXTTTTTTEVETE BPBTXXSVEPE BSBPTTVPSETE BPBXVVEPE BEBTXSETE BPBPVPSXPE BTBPVVVETE BPBTSXSETE BPBPTTTPTTTTTVPSEPE BTBTXXTTSTVPSETE BBBTXSETE BPBTPXSEPE BPBPVPXTTTTVPXTVPXVPXTTTVVEVE BTBXXXTVPSETE BEBTSSSSSXXVPXTVVETE BTBXTTVVETE BPBTXSTPE BTBTXXTTTVPSBTE BTBTXSETX BTBTSXSSTE
We cannot feed strings directly to an RNN, so we need to encode them somehow. One option would be to one-hot encode each character. Another option is to use embeddings. Let's go for the second option (but since there are just a handful of characters, one-hot encoding would probably be a good option as well). For embeddings to work, we need to convert each string into a sequence of character IDs. Let's write a function for that, using each character's index in the string of possible characters "BEPSTVX":
In [106]:
def string_to_ids(s, chars=POSSIBLE_CHARS):
return [chars.index(c) for c in s]
In [107]:
string_to_ids("BTTTXXVVETE")
Out[107]:
[0, 4, 4, 4, 6, 6, 5, 5, 1, 4, 1]
We can now generate the dataset, with 50% good strings, and 50% bad strings:
In [108]:
def generate_dataset(size):
good_strings = [
string_to_ids(generate_string(embedded_reber_grammar))
for _ in range(size // 2)
]
bad_strings = [
string_to_ids(generate_corrupted_string(embedded_reber_grammar))
for _ in range(size - size // 2)
]
all_strings = good_strings + bad_strings
X = tf.ragged.constant(all_strings, ragged_rank=1)
y = np.array([[1.] for _ in range(len(good_strings))] +
[[0.] for _ in range(len(bad_strings))])
return X, y
In [109]:
np.random.seed(42)
X_train, y_train = generate_dataset(10000)
X_valid, y_valid = generate_dataset(2000)
Let's take a look at the first training sequence:
In [110]:
X_train[0]
Out[110]:
<tf.Tensor: shape=(22,), dtype=int32, numpy=
array([0, 4, 0, 2, 4, 4, 4, 5, 2, 6, 4, 5, 2, 6, 4, 4, 5, 2, 3, 1, 4, 1],
dtype=int32)>
What class does it belong to?
In [111]:
y_train[0]
Out[111]:
array([1.])
Perfect! We are ready to create the RNN to identify good strings. We build a simple sequence binary classifier:
In [112]:
np.random.seed(42)
tf.random.set_seed(42)
embedding_size = 5
model = tf.keras.Sequential([
tf.keras.layers.InputLayer(input_shape=[None], dtype=tf.int32, ragged=True),
tf.keras.layers.Embedding(input_dim=len(POSSIBLE_CHARS),
output_dim=embedding_size),
tf.keras.layers.GRU(30),
tf.keras.layers.Dense(1, activation="sigmoid")
])
optimizer = tf.keras.optimizers.SGD(learning_rate=0.02, momentum = 0.95,
nesterov=True)
model.compile(loss="binary_crossentropy", optimizer=optimizer,
metrics=["accuracy"])
history = model.fit(X_train, y_train, epochs=20,
validation_data=(X_valid, y_valid))
Epoch 1/20 313/313 [==============================] - 4s 8ms/step - loss: 0.6910 - accuracy: 0.5095 - val_loss: 0.6825 - val_accuracy: 0.5645 Epoch 2/20 313/313 [==============================] - 2s 7ms/step - loss: 0.6678 - accuracy: 0.5659 - val_loss: 0.6635 - val_accuracy: 0.6105 Epoch 3/20 313/313 [==============================] - 2s 7ms/step - loss: 0.6504 - accuracy: 0.5766 - val_loss: 0.6521 - val_accuracy: 0.6110 Epoch 4/20 313/313 [==============================] - 2s 8ms/step - loss: 0.6347 - accuracy: 0.5980 - val_loss: 0.6224 - val_accuracy: 0.6445 Epoch 5/20 313/313 [==============================] - 2s 7ms/step - loss: 0.6054 - accuracy: 0.6361 - val_loss: 0.5779 - val_accuracy: 0.6980 Epoch 6/20 313/313 [==============================] - 2s 7ms/step - loss: 0.5414 - accuracy: 0.7093 - val_loss: 0.4695 - val_accuracy: 0.7795 Epoch 7/20 313/313 [==============================] - 2s 7ms/step - loss: 0.3756 - accuracy: 0.8418 - val_loss: 0.2685 - val_accuracy: 0.9115 Epoch 8/20 313/313 [==============================] - 2s 7ms/step - loss: 0.2601 - accuracy: 0.9044 - val_loss: 0.1534 - val_accuracy: 0.9615 Epoch 9/20 313/313 [==============================] - 2s 7ms/step - loss: 0.1774 - accuracy: 0.9427 - val_loss: 0.1063 - val_accuracy: 0.9735 Epoch 10/20 313/313 [==============================] - 2s 7ms/step - loss: 0.0624 - accuracy: 0.9826 - val_loss: 0.0219 - val_accuracy: 0.9975 Epoch 11/20 313/313 [==============================] - 2s 7ms/step - loss: 0.0371 - accuracy: 0.9914 - val_loss: 0.0055 - val_accuracy: 1.0000 Epoch 12/20 313/313 [==============================] - 2s 7ms/step - loss: 0.0029 - accuracy: 0.9995 - val_loss: 8.7265e-04 - val_accuracy: 1.0000 Epoch 13/20 313/313 [==============================] - 2s 7ms/step - loss: 6.7552e-04 - accuracy: 1.0000 - val_loss: 4.9408e-04 - val_accuracy: 1.0000 Epoch 14/20 313/313 [==============================] - 2s 7ms/step - loss: 4.4514e-04 - accuracy: 1.0000 - val_loss: 3.6322e-04 - val_accuracy: 1.0000 Epoch 15/20 313/313 [==============================] - 2s 7ms/step - loss: 3.3943e-04 - accuracy: 1.0000 - val_loss: 2.8524e-04 - val_accuracy: 1.0000 Epoch 16/20 313/313 [==============================] - 2s 7ms/step - loss: 2.7723e-04 - accuracy: 1.0000 - val_loss: 2.3880e-04 - val_accuracy: 1.0000 Epoch 17/20 313/313 [==============================] - 2s 7ms/step - loss: 2.3477e-04 - accuracy: 1.0000 - val_loss: 2.0363e-04 - val_accuracy: 1.0000 Epoch 18/20 313/313 [==============================] - 2s 7ms/step - loss: 2.0382e-04 - accuracy: 1.0000 - val_loss: 1.7760e-04 - val_accuracy: 1.0000 Epoch 19/20 313/313 [==============================] - 2s 7ms/step - loss: 1.8077e-04 - accuracy: 1.0000 - val_loss: 1.5916e-04 - val_accuracy: 1.0000 Epoch 20/20 313/313 [==============================] - 2s 8ms/step - loss: 1.6246e-04 - accuracy: 1.0000 - val_loss: 1.4362e-04 - val_accuracy: 1.0000
Now let's test our RNN on two tricky strings: the first one is bad while the second one is good. They only differ by the second to last character. If the RNN gets this right, it shows that it managed to notice the pattern that the second letter should always be equal to the second to last letter. That requires a fairly long short-term memory (which is the reason why we used a GRU cell).
In [113]:
test_strings = ["BPBTSSSSSSSXXTTVPXVPXTTTTTVVETE",
"BPBTSSSSSSSXXTTVPXVPXTTTTTVVEPE"]
X_test = tf.ragged.constant([string_to_ids(s) for s in test_strings], ragged_rank=1)
y_proba = model.predict(X_test)
print()
print("Estimated probability that these are Reber strings:")
for index, string in enumerate(test_strings):
print("{}: {:.2f}%".format(string, 100 * y_proba[index][0]))
Estimated probability that these are Reber strings: BPBTSSSSSSSXXTTVPXVPXTTTTTVVETE: 0.02% BPBTSSSSSSSXXTTVPXVPXTTTTTVVEPE: 99.99%
Ta-da! It worked fine. The RNN found the correct answers with very high confidence. :)
9.¶
Exercise: Train an Encoder–Decoder model that can convert a date string from one format to another (e.g., from "April 22, 2019" to "2019-04-22").
Let's start by creating the dataset. We will use random days between 1000-01-01 and 9999-12-31:
In [114]:
from datetime import date
# cannot use strftime()'s %B format since it depends on the locale
MONTHS = ["January", "February", "March", "April", "May", "June",
"July", "August", "September", "October", "November", "December"]
def random_dates(n_dates):
min_date = date(1000, 1, 1).toordinal()
max_date = date(9999, 12, 31).toordinal()
ordinals = np.random.randint(max_date - min_date, size=n_dates) + min_date
dates = [date.fromordinal(ordinal) for ordinal in ordinals]
x = [MONTHS[dt.month - 1] + " " + dt.strftime("%d, %Y") for dt in dates]
y = [dt.isoformat() for dt in dates]
return x, y
Here are a few random dates, displayed in both the input format and the target format:
In [115]:
np.random.seed(42)
n_dates = 3
x_example, y_example = random_dates(n_dates)
print("{:25s}{:25s}".format("Input", "Target"))
print("-" * 50)
for idx in range(n_dates):
print("{:25s}{:25s}".format(x_example[idx], y_example[idx]))
Input Target -------------------------------------------------- September 20, 7075 7075-09-20 May 15, 8579 8579-05-15 January 11, 7103 7103-01-11
Let's get the list of all possible characters in the inputs:
In [116]:
INPUT_CHARS = "".join(sorted(set("".join(MONTHS) + "0123456789, ")))
INPUT_CHARS
Out[116]:
' ,0123456789ADFJMNOSabceghilmnoprstuvy'
And here's the list of possible characters in the outputs:
In [117]:
OUTPUT_CHARS = "0123456789-"
Let's write a function to convert a string to a list of character IDs, as we did in the previous exercise:
In [118]:
def date_str_to_ids(date_str, chars=INPUT_CHARS):
return [chars.index(c) for c in date_str]
In [119]:
date_str_to_ids(x_example[0], INPUT_CHARS)
Out[119]:
[19, 23, 31, 34, 23, 28, 21, 23, 32, 0, 4, 2, 1, 0, 9, 2, 9, 7]
In [120]:
date_str_to_ids(y_example[0], OUTPUT_CHARS)
Out[120]:
[7, 0, 7, 5, 10, 0, 9, 10, 2, 0]
In [121]:
def prepare_date_strs(date_strs, chars=INPUT_CHARS):
X_ids = [date_str_to_ids(dt, chars) for dt in date_strs]
X = tf.ragged.constant(X_ids, ragged_rank=1)
return (X + 1).to_tensor() # using 0 as the padding token ID
def create_dataset(n_dates):
x, y = random_dates(n_dates)
return prepare_date_strs(x, INPUT_CHARS), prepare_date_strs(y, OUTPUT_CHARS)
In [122]:
np.random.seed(42)
X_train, Y_train = create_dataset(10000)
X_valid, Y_valid = create_dataset(2000)
X_test, Y_test = create_dataset(2000)
In [123]:
Y_train[0]
Out[123]:
<tf.Tensor: shape=(10,), dtype=int32, numpy=array([ 8, 1, 8, 6, 11, 1, 10, 11, 3, 1], dtype=int32)>
First version: a very basic seq2seq model¶
Let's first try the simplest possible model: we feed in the input sequence, which first goes through the encoder (an embedding layer followed by a single LSTM layer), which outputs a vector, then it goes through a decoder (a single LSTM layer, followed by a dense output layer), which outputs a sequence of vectors, each representing the estimated probabilities for all possible output character.
Since the decoder expects a sequence as input, we repeat the vector (which is output by the encoder) as many times as the longest possible output sequence.
In [124]:
embedding_size = 32
max_output_length = Y_train.shape[1]
np.random.seed(42)
tf.random.set_seed(42)
encoder = tf.keras.Sequential([
tf.keras.layers.Embedding(input_dim=len(INPUT_CHARS) + 1,
output_dim=embedding_size,
input_shape=[None]),
tf.keras.layers.LSTM(128)
])
decoder = tf.keras.Sequential([
tf.keras.layers.LSTM(128, return_sequences=True),
tf.keras.layers.Dense(len(OUTPUT_CHARS) + 1, activation="softmax")
])
model = tf.keras.Sequential([
encoder,
tf.keras.layers.RepeatVector(max_output_length),
decoder
])
optimizer = tf.keras.optimizers.Nadam()
model.compile(loss="sparse_categorical_crossentropy", optimizer=optimizer,
metrics=["accuracy"])
history = model.fit(X_train, Y_train, epochs=20,
validation_data=(X_valid, Y_valid))
Epoch 1/20 313/313 [==============================] - 10s 23ms/step - loss: 1.8150 - accuracy: 0.3489 - val_loss: 1.3726 - val_accuracy: 0.4939 Epoch 2/20 313/313 [==============================] - 7s 22ms/step - loss: 1.2447 - accuracy: 0.5510 - val_loss: 1.0725 - val_accuracy: 0.6115 Epoch 3/20 313/313 [==============================] - 7s 23ms/step - loss: 1.0937 - accuracy: 0.6125 - val_loss: 1.0548 - val_accuracy: 0.6130 Epoch 4/20 313/313 [==============================] - 7s 23ms/step - loss: 1.0032 - accuracy: 0.6413 - val_loss: 3.8747 - val_accuracy: 0.1788 Epoch 5/20 313/313 [==============================] - 8s 26ms/step - loss: 0.8159 - accuracy: 0.7023 - val_loss: 0.6623 - val_accuracy: 0.7474 Epoch 6/20 313/313 [==============================] - 8s 26ms/step - loss: 0.5645 - accuracy: 0.7795 - val_loss: 0.5005 - val_accuracy: 0.8032 Epoch 7/20 313/313 [==============================] - 8s 26ms/step - loss: 0.5037 - accuracy: 0.8103 - val_loss: 0.3798 - val_accuracy: 0.8500 Epoch 8/20 313/313 [==============================] - 8s 26ms/step - loss: 0.3131 - accuracy: 0.8795 - val_loss: 0.2582 - val_accuracy: 0.9043 Epoch 9/20 313/313 [==============================] - 8s 26ms/step - loss: 0.2141 - accuracy: 0.9280 - val_loss: 0.1637 - val_accuracy: 0.9498 Epoch 10/20 313/313 [==============================] - 9s 28ms/step - loss: 0.1282 - accuracy: 0.9650 - val_loss: 0.0918 - val_accuracy: 0.9774 Epoch 11/20 313/313 [==============================] - 9s 28ms/step - loss: 0.0669 - accuracy: 0.9871 - val_loss: 0.3368 - val_accuracy: 0.8871 Epoch 12/20 313/313 [==============================] - 10s 32ms/step - loss: 0.1551 - accuracy: 0.9662 - val_loss: 0.0398 - val_accuracy: 0.9949 Epoch 13/20 313/313 [==============================] - 9s 29ms/step - loss: 0.0291 - accuracy: 0.9969 - val_loss: 0.0240 - val_accuracy: 0.9984 Epoch 14/20 313/313 [==============================] - 9s 30ms/step - loss: 0.0182 - accuracy: 0.9986 - val_loss: 0.0161 - val_accuracy: 0.9993 Epoch 15/20 313/313 [==============================] - 9s 30ms/step - loss: 0.0119 - accuracy: 0.9995 - val_loss: 0.0112 - val_accuracy: 0.9997 Epoch 16/20 313/313 [==============================] - 10s 32ms/step - loss: 0.0082 - accuracy: 0.9998 - val_loss: 0.0083 - val_accuracy: 0.9999 Epoch 17/20 313/313 [==============================] - 10s 33ms/step - loss: 0.0059 - accuracy: 0.9999 - val_loss: 0.0058 - val_accuracy: 0.9999 Epoch 18/20 313/313 [==============================] - 11s 34ms/step - loss: 0.0042 - accuracy: 1.0000 - val_loss: 0.0043 - val_accuracy: 0.9999 Epoch 19/20 313/313 [==============================] - 10s 33ms/step - loss: 0.0031 - accuracy: 1.0000 - val_loss: 0.0034 - val_accuracy: 0.9999 Epoch 20/20 313/313 [==============================] - 12s 40ms/step - loss: 0.0024 - accuracy: 1.0000 - val_loss: 0.0026 - val_accuracy: 1.0000
Looks great, we reach 100% validation accuracy! Let's use the model to make some predictions. We will need to be able to convert a sequence of character IDs to a readable string:
In [125]:
def ids_to_date_strs(ids, chars=OUTPUT_CHARS):
return ["".join([("?" + chars)[index] for index in sequence])
for sequence in ids]
Now we can use the model to convert some dates
In [126]:
X_new = prepare_date_strs(["September 17, 2009", "July 14, 1789"])
In [127]:
ids = model.predict(X_new).argmax(axis=-1)
for date_str in ids_to_date_strs(ids):
print(date_str)
2009-09-17 1789-07-14
Perfect! :)
However, since the model was only trained on input strings of length 18 (which is the length of the longest date), it does not perform well if we try to use it to make predictions on shorter sequences:
In [128]:
X_new = prepare_date_strs(["May 02, 2020", "July 14, 1789"])
In [129]:
ids = model.predict(X_new).argmax(axis=-1)
for date_str in ids_to_date_strs(ids):
print(date_str)
2020-02-02 1789-01-14
Oops! We need to ensure that we always pass sequences of the same length as during training, using padding if necessary. Let's write a little helper function for that:
In [130]:
max_input_length = X_train.shape[1]
def prepare_date_strs_padded(date_strs):
X = prepare_date_strs(date_strs)
if X.shape[1] < max_input_length:
X = tf.pad(X, [[0, 0], [0, max_input_length - X.shape[1]]])
return X
def convert_date_strs(date_strs):
X = prepare_date_strs_padded(date_strs)
ids = model.predict(X).argmax(axis=-1)
return ids_to_date_strs(ids)
In [131]:
convert_date_strs(["May 02, 2020", "July 14, 1789"])
Out[131]:
['2020-05-02', '1789-07-14']
Cool! Granted, there are certainly much easier ways to write a date conversion tool (e.g., using regular expressions or even basic string manipulation), but you have to admit that using neural networks is way cooler. ;-)
However, real-life sequence-to-sequence problems will usually be harder, so for the sake of completeness, let's build a more powerful model.
Second version: feeding the shifted targets to the decoder (teacher forcing)¶
Instead of feeding the decoder a simple repetition of the encoder's output vector, we can feed it the target sequence, shifted by one time step to the right. This way, at each time step the decoder will know what the previous target character was. This should help is tackle more complex sequence-to-sequence problems.
Since the first output character of each target sequence has no previous character, we will need a new token to represent the start-of-sequence (sos).
During inference, we won't know the target, so what will we feed the decoder? We can just predict one character at a time, starting with an sos token, then feeding the decoder all the characters that were predicted so far (we will look at this in more details later in this notebook).
But if the decoder's LSTM expects to get the previous target as input at each step, how shall we pass it it the vector output by the encoder? Well, one option is to ignore the output vector, and instead use the encoder's LSTM state as the initial state of the decoder's LSTM (which requires that encoder's LSTM must have the same number of units as the decoder's LSTM).
Now let's create the decoder's inputs (for training, validation and testing). The sos token will be represented using the last possible output character's ID + 1.
In [132]:
sos_id = len(OUTPUT_CHARS) + 1
def shifted_output_sequences(Y):
sos_tokens = tf.fill(dims=(len(Y), 1), value=sos_id)
return tf.concat([sos_tokens, Y[:, :-1]], axis=1)
X_train_decoder = shifted_output_sequences(Y_train)
X_valid_decoder = shifted_output_sequences(Y_valid)
X_test_decoder = shifted_output_sequences(Y_test)
Let's take a look at the decoder's training inputs:
In [133]:
X_train_decoder
Out[133]:
<tf.Tensor: shape=(10000, 10), dtype=int32, numpy=
array([[12, 8, 1, ..., 10, 11, 3],
[12, 9, 6, ..., 6, 11, 2],
[12, 8, 2, ..., 2, 11, 2],
...,
[12, 10, 8, ..., 2, 11, 4],
[12, 2, 2, ..., 3, 11, 3],
[12, 8, 9, ..., 8, 11, 3]], dtype=int32)>
Now let's build the model. It's not a simple sequential model anymore, so let's use the functional API:
In [134]:
encoder_embedding_size = 32
decoder_embedding_size = 32
lstm_units = 128
np.random.seed(42)
tf.random.set_seed(42)
encoder_input = tf.keras.layers.Input(shape=[None], dtype=tf.int32)
encoder_embedding = tf.keras.layers.Embedding(
input_dim=len(INPUT_CHARS) + 1,
output_dim=encoder_embedding_size)(encoder_input)
_, encoder_state_h, encoder_state_c = tf.keras.layers.LSTM(
lstm_units, return_state=True)(encoder_embedding)
encoder_state = [encoder_state_h, encoder_state_c]
decoder_input = tf.keras.layers.Input(shape=[None], dtype=tf.int32)
decoder_embedding = tf.keras.layers.Embedding(
input_dim=len(OUTPUT_CHARS) + 2,
output_dim=decoder_embedding_size)(decoder_input)
decoder_lstm_output = tf.keras.layers.LSTM(lstm_units, return_sequences=True)(
decoder_embedding, initial_state=encoder_state)
decoder_output = tf.keras.layers.Dense(len(OUTPUT_CHARS) + 1,
activation="softmax")(decoder_lstm_output)
model = tf.keras.Model(inputs=[encoder_input, decoder_input],
outputs=[decoder_output])
optimizer = tf.keras.optimizers.Nadam()
model.compile(loss="sparse_categorical_crossentropy", optimizer=optimizer,
metrics=["accuracy"])
history = model.fit([X_train, X_train_decoder], Y_train, epochs=10,
validation_data=([X_valid, X_valid_decoder], Y_valid))
Epoch 1/10 313/313 [==============================] - 11s 27ms/step - loss: 1.6824 - accuracy: 0.3734 - val_loss: 1.4054 - val_accuracy: 0.4681 Epoch 2/10 313/313 [==============================] - 8s 26ms/step - loss: 1.1935 - accuracy: 0.5550 - val_loss: 0.8868 - val_accuracy: 0.6750 Epoch 3/10 313/313 [==============================] - 8s 26ms/step - loss: 0.6403 - accuracy: 0.7700 - val_loss: 0.3493 - val_accuracy: 0.8978 Epoch 4/10 313/313 [==============================] - 8s 26ms/step - loss: 0.2292 - accuracy: 0.9423 - val_loss: 0.1254 - val_accuracy: 0.9782 Epoch 5/10 313/313 [==============================] - 8s 26ms/step - loss: 0.0694 - accuracy: 0.9932 - val_loss: 0.0441 - val_accuracy: 0.9982 Epoch 6/10 313/313 [==============================] - 9s 29ms/step - loss: 0.0576 - accuracy: 0.9923 - val_loss: 0.0280 - val_accuracy: 0.9988 Epoch 7/10 313/313 [==============================] - 8s 26ms/step - loss: 0.0179 - accuracy: 0.9998 - val_loss: 0.0143 - val_accuracy: 0.9999 Epoch 8/10 313/313 [==============================] - 6s 18ms/step - loss: 0.0107 - accuracy: 0.9999 - val_loss: 0.0092 - val_accuracy: 0.9999 Epoch 9/10 313/313 [==============================] - 6s 20ms/step - loss: 0.0070 - accuracy: 1.0000 - val_loss: 0.0065 - val_accuracy: 0.9999 Epoch 10/10 313/313 [==============================] - 6s 18ms/step - loss: 0.0050 - accuracy: 1.0000 - val_loss: 0.0047 - val_accuracy: 0.9999
This model also reaches 100% validation accuracy, but it does so even faster.
Let's once again use the model to make some predictions. This time we need to predict characters one by one.
In [135]:
sos_id = len(OUTPUT_CHARS) + 1
def predict_date_strs(date_strs):
X = prepare_date_strs_padded(date_strs)
Y_pred = tf.fill(dims=(len(X), 1), value=sos_id)
for index in range(max_output_length):
pad_size = max_output_length - Y_pred.shape[1]
X_decoder = tf.pad(Y_pred, [[0, 0], [0, pad_size]])
Y_probas_next = model.predict([X, X_decoder])[:, index:index+1]
Y_pred_next = tf.argmax(Y_probas_next, axis=-1, output_type=tf.int32)
Y_pred = tf.concat([Y_pred, Y_pred_next], axis=1)
return ids_to_date_strs(Y_pred[:, 1:])
In [136]:
predict_date_strs(["July 14, 1789", "May 01, 2020"])
Out[136]:
['1789-07-14', '2020-05-01']
Works fine! Next, feel free to write a Transformer version. :)
10.¶
Exercise: Go through Keras's tutorial for Natural language image search with a Dual Encoder. You will learn how to build a model capable of representing both images and text within the same embedding space. This makes it possible to search for images using a text prompt, like in the CLIP model by OpenAI.
Just click the link and follow the instructions.
11.¶
Exercise: Use the Transformers library to download a pretrained language model capable of generating text (e.g., GPT), and try generating more convincing Shakespearean text. You will need to use the model's generate() method—see Hugging Face's documentation for more details.
WARNING: I've updated this solution to use PyTorch instead of TensorFlow since the Transformers library no longer supports TensorFlow (or JAX).
First, let's load a pretrained model. In this example, we will use OpenAI's GPT model, with an additional Language Model on top (just a linear layer with weights tied to the input embeddings). Let's import it and load the pretrained weights (this will download about 445MB of data to ~/.cache/torch/transformers):
In [137]:
from transformers import OpenAIGPTLMHeadModel
model = OpenAIGPTLMHeadModel.from_pretrained("openai-gpt")
generation_config.json: 0%| | 0.00/74.0 [00:00<?, ?B/s]
In [138]:
from transformers import OpenAIGPTTokenizer
tokenizer = OpenAIGPTTokenizer.from_pretrained("openai-gpt")
tokenizer_config.json: 0%| | 0.00/25.0 [00:00<?, ?B/s]
vocab.json: 0%| | 0.00/816k [00:00<?, ?B/s]
merges.txt: 0%| | 0.00/458k [00:00<?, ?B/s]
tokenizer.json: 0%| | 0.00/1.27M [00:00<?, ?B/s]
ftfy or spacy is not installed using BERT BasicTokenizer instead of SpaCy & ftfy.
Now let's use the tokenizer to tokenize and encode the prompt text:
In [139]:
tokenizer("hello everyone")
Out[139]:
{'input_ids': [3570, 1473], 'attention_mask': [1, 1]}
In [140]:
prompt_text = "This royal throne of kings, this sceptred isle"
encoded_prompt = tokenizer.encode(prompt_text,
add_special_tokens=False,
return_tensors="pt")
encoded_prompt
Out[140]:
tensor([[ 616, 5751, 6404, 498, 9606, 240, 616, 26271, 7428, 16187]])
Easy! Next, let's use the model to generate text after the prompt. We will generate 5 different sentences, each starting with the prompt text, followed by 40 additional tokens. For an explanation of what all the hyperparameters do, make sure to check out this great blog post by Patrick von Platen (from Hugging Face). You can play around with the hyperparameters to try to obtain better results.
In [141]:
num_sequences = 5
length = 40
generated_sequences = model.generate(
input_ids=encoded_prompt,
do_sample=True,
max_length=length + len(encoded_prompt[0]),
temperature=1.0,
top_k=0,
top_p=0.9,
repetition_penalty=1.0,
num_return_sequences=num_sequences,
)
generated_sequences
Out[141]:
tensor([[ 616, 5751, 6404, 498, 9606, 240, 616, 26271, 7428, 16187,
240, 544, 599, 606, 1370, 481, 1424, 5590, 498, 35,
480, 4785, 239, 1099, 1475, 606, 9029, 481, 1393, 4413,
8323, 2185, 984, 1496, 5566, 622, 905, 2283, 16015, 240,
617, 1129, 488, 592, 26980, 485, 481, 6558, 488, 13005],
[ 616, 5751, 6404, 498, 9606, 240, 616, 26271, 7428, 16187,
606, 761, 2954, 246, 618, 498, 1128, 1403, 16496, 239,
40477, 487, 535, 246, 1294, 260, 5633, 240, 2883, 260,
3207, 240, 1301, 260, 3207, 240, 1301, 260, 3207, 762,
763, 635, 963, 862, 580, 618, 498, 1306, 592, 753],
[ 616, 5751, 6404, 498, 9606, 240, 616, 26271, 7428, 16187,
240, 562, 655, 544, 246, 1820, 1311, 6260, 815, 498,
589, 1885, 1175, 240, 6260, 784, 815, 481, 1820, 498,
505, 2549, 3512, 240, 3491, 507, 635, 868, 4560, 481,
21467, 535, 2843, 562, 507, 239, 6202, 240, 871, 1234],
[ 616, 5751, 6404, 498, 9606, 240, 616, 26271, 7428, 16187,
509, 599, 487, 656, 2655, 562, 239, 487, 656, 1802,
240, 557, 558, 1241, 589, 498, 524, 2080, 239, 40477,
244, 249, 1048, 9306, 240, 244, 570, 18165, 603, 239,
244, 525, 512, 636, 21030, 704, 6650, 260, 498, 260],
[ 616, 5751, 6404, 498, 9606, 240, 616, 26271, 7428, 16187,
544, 725, 17107, 815, 775, 728, 1082, 500, 481, 1276,
239, 244, 40477, 244, 1272, 16943, 1558, 793, 500, 524,
8140, 535, 13988, 240, 244, 34506, 1014, 575, 239, 40477,
244, 500, 481, 2992, 240, 244, 481, 31623, 531, 4267]])
Now let's decode the generated sequences and print them:
In [142]:
for sequence in generated_sequences:
text = tokenizer.decode(sequence, clean_up_tokenization_spaces=True)
print(text)
print("-" * 80)
this royal throne of kings, this sceptred isle, is what we call the great desert of qandar. every morning we greet the big french noble lord which oversees our most important functions, from work and digestion to the health and maintenance -------------------------------------------------------------------------------- this royal throne of kings, this sceptred isle we're calling a king of mycenae. he's a dark - haired, gold - eyed, black - eyed, black - eyed man who could very well be king of magdiel -------------------------------------------------------------------------------- this royal throne of kings, this sceptred isle, for there is a power far greater than of all living men, greater even than the power of eleusis, indeed it could never survive the adversary's search for it. therefore, ever since -------------------------------------------------------------------------------- this royal throne of kings, this sceptred isle was what he 'd lived for. he 'd died, as had almost all of his children. " i am offended, " abriel said. " that you would disrespect your servant - of - -------------------------------------------------------------------------------- this royal throne of kings, this sceptred isle is more desirable than any other place in the world. " " many nobles stay here in his majesty's apartments, " cherek told him. " in the flesh, " the drasnian responded --------------------------------------------------------------------------------
You can try more recent (and larger) models, such as GPT-2, CTRL, Transformer-XL or XLNet, which are all available as pretrained models in the transformers library, including variants with Language Models on top. The preprocessing steps vary slightly between models, so make sure to check out this generation example from the transformers documentation.
Hope you enjoyed this chapter! :)