作者: Mohamad Merchant
创建日期 2020/08/15
上次修改日期 2020/08/29
描述:在 SNLI 语料库上微调 BERT 模型,进行自然语言推理。
语义相似度是指确定两个句子在含义上的相似程度的任务。本示例演示了如何使用 SNLI(斯坦福自然语言推理)语料库和 Transformer 来预测句子的语义相似度。我们将微调一个 BERT 模型,该模型将两个句子作为输入,并输出这两个句子的相似度得分。
注意:通过 `pip install transformers` 安装 HuggingFace `transformers`(版本 >= 2.11.0)。
import numpy as np
import pandas as pd
import tensorflow as tf
import transformers
max_length = 128 # Maximum length of input sentence to the model.
batch_size = 32
epochs = 2
# Labels in our dataset.
labels = ["contradiction", "entailment", "neutral"]
!curl -LO https://raw.githubusercontent.com/MohamadMerchant/SNLI/master/data.tar.gz
!tar -xvzf data.tar.gz
# There are more than 550k samples in total; we will use 100k for this example.
train_df = pd.read_csv("SNLI_Corpus/snli_1.0_train.csv", nrows=100000)
valid_df = pd.read_csv("SNLI_Corpus/snli_1.0_dev.csv")
test_df = pd.read_csv("SNLI_Corpus/snli_1.0_test.csv")
# Shape of the data
print(f"Total train samples : {train_df.shape[0]}")
print(f"Total validation samples: {valid_df.shape[0]}")
print(f"Total test samples: {valid_df.shape[0]}")
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 11.1M 100 11.1M 0 0 5231k 0 0:00:02 0:00:02 --:--:-- 5231k
SNLI_Corpus/
SNLI_Corpus/snli_1.0_dev.csv
SNLI_Corpus/snli_1.0_train.csv
SNLI_Corpus/snli_1.0_test.csv
Total train samples : 100000
Total validation samples: 10000
Total test samples: 10000
数据集概述
以下是我们数据集中“similarity”标签的值
让我们看看数据集中的一样本
print(f"Sentence1: {train_df.loc[1, 'sentence1']}")
print(f"Sentence2: {train_df.loc[1, 'sentence2']}")
print(f"Similarity: {train_df.loc[1, 'similarity']}")
Sentence1: A person on a horse jumps over a broken down airplane.
Sentence2: A person is at a diner, ordering an omelette.
Similarity: contradiction
# We have some NaN entries in our train data, we will simply drop them.
print("Number of missing values")
print(train_df.isnull().sum())
train_df.dropna(axis=0, inplace=True)
Number of missing values
similarity 0
sentence1 0
sentence2 3
dtype: int64
训练目标的分布。
print("Train Target Distribution")
print(train_df.similarity.value_counts())
Train Target Distribution
entailment 33384
contradiction 33310
neutral 33193
- 110
Name: similarity, dtype: int64
验证目标的分布。
print("Validation Target Distribution")
print(valid_df.similarity.value_counts())
Validation Target Distribution
entailment 3329
contradiction 3278
neutral 3235
- 158
Name: similarity, dtype: int64
值“-”出现在我们的训练和验证目标中。我们将跳过这些样本。
train_df = (
train_df[train_df.similarity != "-"]
.sample(frac=1.0, random_state=42)
.reset_index(drop=True)
)
valid_df = (
valid_df[valid_df.similarity != "-"]
.sample(frac=1.0, random_state=42)
.reset_index(drop=True)
)
对训练、验证和测试标签进行独热编码。
train_df["label"] = train_df["similarity"].apply(
lambda x: 0 if x == "contradiction" else 1 if x == "entailment" else 2
)
y_train = tf.keras.utils.to_categorical(train_df.label, num_classes=3)
valid_df["label"] = valid_df["similarity"].apply(
lambda x: 0 if x == "contradiction" else 1 if x == "entailment" else 2
)
y_val = tf.keras.utils.to_categorical(valid_df.label, num_classes=3)
test_df["label"] = test_df["similarity"].apply(
lambda x: 0 if x == "contradiction" else 1 if x == "entailment" else 2
)
y_test = tf.keras.utils.to_categorical(test_df.label, num_classes=3)
class BertSemanticDataGenerator(tf.keras.utils.Sequence):
"""Generates batches of data.
Args:
sentence_pairs: Array of premise and hypothesis input sentences.
labels: Array of labels.
batch_size: Integer batch size.
shuffle: boolean, whether to shuffle the data.
include_targets: boolean, whether to include the labels.
Returns:
Tuples `([input_ids, attention_mask, `token_type_ids], labels)`
(or just `[input_ids, attention_mask, `token_type_ids]`
if `include_targets=False`)
"""
def __init__(
self,
sentence_pairs,
labels,
batch_size=batch_size,
shuffle=True,
include_targets=True,
):
self.sentence_pairs = sentence_pairs
self.labels = labels
self.shuffle = shuffle
self.batch_size = batch_size
self.include_targets = include_targets
# Load our BERT Tokenizer to encode the text.
# We will use base-base-uncased pretrained model.
self.tokenizer = transformers.BertTokenizer.from_pretrained(
"bert-base-uncased", do_lower_case=True
)
self.indexes = np.arange(len(self.sentence_pairs))
self.on_epoch_end()
def __len__(self):
# Denotes the number of batches per epoch.
return len(self.sentence_pairs) // self.batch_size
def __getitem__(self, idx):
# Retrieves the batch of index.
indexes = self.indexes[idx * self.batch_size : (idx + 1) * self.batch_size]
sentence_pairs = self.sentence_pairs[indexes]
# With BERT tokenizer's batch_encode_plus batch of both the sentences are
# encoded together and separated by [SEP] token.
encoded = self.tokenizer.batch_encode_plus(
sentence_pairs.tolist(),
add_special_tokens=True,
max_length=max_length,
return_attention_mask=True,
return_token_type_ids=True,
pad_to_max_length=True,
return_tensors="tf",
)
# Convert batch of encoded features to numpy array.
input_ids = np.array(encoded["input_ids"], dtype="int32")
attention_masks = np.array(encoded["attention_mask"], dtype="int32")
token_type_ids = np.array(encoded["token_type_ids"], dtype="int32")
# Set to true if data generator is used for training/validation.
if self.include_targets:
labels = np.array(self.labels[indexes], dtype="int32")
return [input_ids, attention_masks, token_type_ids], labels
else:
return [input_ids, attention_masks, token_type_ids]
def on_epoch_end(self):
# Shuffle indexes after each epoch if shuffle is set to True.
if self.shuffle:
np.random.RandomState(42).shuffle(self.indexes)
# Create the model under a distribution strategy scope.
strategy = tf.distribute.MirroredStrategy()
with strategy.scope():
# Encoded token ids from BERT tokenizer.
input_ids = tf.keras.layers.Input(
shape=(max_length,), dtype=tf.int32, name="input_ids"
)
# Attention masks indicates to the model which tokens should be attended to.
attention_masks = tf.keras.layers.Input(
shape=(max_length,), dtype=tf.int32, name="attention_masks"
)
# Token type ids are binary masks identifying different sequences in the model.
token_type_ids = tf.keras.layers.Input(
shape=(max_length,), dtype=tf.int32, name="token_type_ids"
)
# Loading pretrained BERT model.
bert_model = transformers.TFBertModel.from_pretrained("bert-base-uncased")
# Freeze the BERT model to reuse the pretrained features without modifying them.
bert_model.trainable = False
bert_output = bert_model.bert(
input_ids, attention_mask=attention_masks, token_type_ids=token_type_ids
)
sequence_output = bert_output.last_hidden_state
pooled_output = bert_output.pooler_output
# Add trainable layers on top of frozen layers to adapt the pretrained features on the new data.
bi_lstm = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(64, return_sequences=True)
)(sequence_output)
# Applying hybrid pooling approach to bi_lstm sequence output.
avg_pool = tf.keras.layers.GlobalAveragePooling1D()(bi_lstm)
max_pool = tf.keras.layers.GlobalMaxPooling1D()(bi_lstm)
concat = tf.keras.layers.concatenate([avg_pool, max_pool])
dropout = tf.keras.layers.Dropout(0.3)(concat)
output = tf.keras.layers.Dense(3, activation="softmax")(dropout)
model = tf.keras.models.Model(
inputs=[input_ids, attention_masks, token_type_ids], outputs=output
)
model.compile(
optimizer=tf.keras.optimizers.Adam(),
loss="categorical_crossentropy",
metrics=["acc"],
)
print(f"Strategy: {strategy}")
model.summary()
HBox(children=(FloatProgress(value=0.0, description='Downloading', max=433.0, style=ProgressStyle(description_…
HBox(children=(FloatProgress(value=0.0, description='Downloading', max=536063208.0, style=ProgressStyle(descri…
Strategy: <tensorflow.python.distribute.mirrored_strategy.MirroredStrategy object at 0x7faf9dc63a90>
Model: "functional_1"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_ids (InputLayer) [(None, 128)] 0
__________________________________________________________________________________________________
attention_masks (InputLayer) [(None, 128)] 0
__________________________________________________________________________________________________
token_type_ids (InputLayer) [(None, 128)] 0
__________________________________________________________________________________________________
tf_bert_model (TFBertModel) ((None, 128, 768), ( 109482240 input_ids[0][0]
attention_masks[0][0]
token_type_ids[0][0]
__________________________________________________________________________________________________
bidirectional (Bidirectional) (None, 128, 128) 426496 tf_bert_model[0][0]
__________________________________________________________________________________________________
global_average_pooling1d (Globa (None, 128) 0 bidirectional[0][0]
__________________________________________________________________________________________________
global_max_pooling1d (GlobalMax (None, 128) 0 bidirectional[0][0]
__________________________________________________________________________________________________
concatenate (Concatenate) (None, 256) 0 global_average_pooling1d[0][0]
global_max_pooling1d[0][0]
__________________________________________________________________________________________________
dropout_37 (Dropout) (None, 256) 0 concatenate[0][0]
__________________________________________________________________________________________________
dense (Dense) (None, 3) 771 dropout_37[0][0]
==================================================================================================
Total params: 109,909,507
Trainable params: 427,267
Non-trainable params: 109,482,240
__________________________________________________________________________________________________
创建训练和验证数据生成器
train_data = BertSemanticDataGenerator(
train_df[["sentence1", "sentence2"]].values.astype("str"),
y_train,
batch_size=batch_size,
shuffle=True,
)
valid_data = BertSemanticDataGenerator(
valid_df[["sentence1", "sentence2"]].values.astype("str"),
y_val,
batch_size=batch_size,
shuffle=False,
)
HBox(children=(FloatProgress(value=0.0, description='Downloading', max=231508.0, style=ProgressStyle(descripti…
仅对顶层进行训练以执行“特征提取”,这将允许模型使用预训练模型的表示。
history = model.fit(
train_data,
validation_data=valid_data,
epochs=epochs,
use_multiprocessing=True,
workers=-1,
)
Epoch 1/2
3121/3121 [==============================] - 666s 213ms/step - loss: 0.6925 - acc: 0.7049 - val_loss: 0.5294 - val_acc: 0.7899
Epoch 2/2
3121/3121 [==============================] - 661s 212ms/step - loss: 0.5917 - acc: 0.7587 - val_loss: 0.4955 - val_acc: 0.8052
此步骤必须仅在特征提取模型在新数据上训练收敛后才能执行。
这是一个可选的最后步骤,其中解冻 `bert_model` 并以非常低的学习率重新训练。通过逐渐调整预训练特征以适应新数据,这可以带来有意义的改进。
# Unfreeze the bert_model.
bert_model.trainable = True
# Recompile the model to make the change effective.
model.compile(
optimizer=tf.keras.optimizers.Adam(1e-5),
loss="categorical_crossentropy",
metrics=["accuracy"],
)
model.summary()
Model: "functional_1"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_ids (InputLayer) [(None, 128)] 0
__________________________________________________________________________________________________
attention_masks (InputLayer) [(None, 128)] 0
__________________________________________________________________________________________________
token_type_ids (InputLayer) [(None, 128)] 0
__________________________________________________________________________________________________
tf_bert_model (TFBertModel) ((None, 128, 768), ( 109482240 input_ids[0][0]
attention_masks[0][0]
token_type_ids[0][0]
__________________________________________________________________________________________________
bidirectional (Bidirectional) (None, 128, 128) 426496 tf_bert_model[0][0]
__________________________________________________________________________________________________
global_average_pooling1d (Globa (None, 128) 0 bidirectional[0][0]
__________________________________________________________________________________________________
global_max_pooling1d (GlobalMax (None, 128) 0 bidirectional[0][0]
__________________________________________________________________________________________________
concatenate (Concatenate) (None, 256) 0 global_average_pooling1d[0][0]
global_max_pooling1d[0][0]
__________________________________________________________________________________________________
dropout_37 (Dropout) (None, 256) 0 concatenate[0][0]
__________________________________________________________________________________________________
dense (Dense) (None, 3) 771 dropout_37[0][0]
==================================================================================================
Total params: 109,909,507
Trainable params: 109,909,507
Non-trainable params: 0
__________________________________________________________________________________________________
history = model.fit(
train_data,
validation_data=valid_data,
epochs=epochs,
use_multiprocessing=True,
workers=-1,
)
Epoch 1/2
3121/3121 [==============================] - 1574s 504ms/step - loss: 0.4698 - accuracy: 0.8181 - val_loss: 0.3787 - val_accuracy: 0.8598
Epoch 2/2
3121/3121 [==============================] - 1569s 503ms/step - loss: 0.3516 - accuracy: 0.8702 - val_loss: 0.3416 - val_accuracy: 0.8757
test_data = BertSemanticDataGenerator(
test_df[["sentence1", "sentence2"]].values.astype("str"),
y_test,
batch_size=batch_size,
shuffle=False,
)
model.evaluate(test_data, verbose=1)
312/312 [==============================] - 55s 177ms/step - loss: 0.3697 - accuracy: 0.8629
[0.3696725070476532, 0.8628805875778198]
def check_similarity(sentence1, sentence2):
sentence_pairs = np.array([[str(sentence1), str(sentence2)]])
test_data = BertSemanticDataGenerator(
sentence_pairs, labels=None, batch_size=1, shuffle=False, include_targets=False,
)
proba = model.predict(test_data[0])[0]
idx = np.argmax(proba)
proba = f"{proba[idx]: .2f}%"
pred = labels[idx]
return pred, proba
检查一些示例句子对的结果。
sentence1 = "Two women are observing something together."
sentence2 = "Two women are standing with their eyes closed."
check_similarity(sentence1, sentence2)
('contradiction', ' 0.91%')
检查一些示例句子对的结果。
sentence1 = "A smiling costumed woman is holding an umbrella"
sentence2 = "A happy woman in a fairy costume holds an umbrella"
check_similarity(sentence1, sentence2)
('neutral', ' 0.88%')
检查一些示例句子对的结果
sentence1 = "A soccer game with multiple males playing"
sentence2 = "Some men are playing a sport"
check_similarity(sentence1, sentence2)
('entailment', ' 0.94%')
HuggingFace 上提供的示例
训练后的模型 | 演示 |
---|---|