保存一个“微调”的伯特模型
Posted
技术标签:
【中文标题】保存一个“微调”的伯特模型【英文标题】：Saving a 'fine-tuned' bert model 【发布时间】：2020-04-07 22:31:59 【问题描述】：我正在尝试保存一个经过微调的 bert 模型。我已经正确运行了代码 - 它工作正常，并且在 ipython 控制台中我可以调用 getPrediction 并让它产生结果。
我保存了我的体重文件（最高的是 model.ckpt-333.data-00000-of-00001
我不知道如何将模型保存为可重用。
我正在使用 bert-tensorflow。
import json
import pandas as pd
import tensorflow as tf
import tensorflow_hub as hub
from datetime import datetime
from sklearn.model_selection import train_test_split
import os
print("tensorflow version : ", tf.__version__)
print("tensorflow_hub version : ", hub.__version__)
#Importing BERT modules
import bert
from bert import run_classifier
from bert import optimization
from bert import tokenization
#set output directory of the model
OUTPUT_DIR = 'model'
#@markdown Whether or not to clear/delete the directory and create a new one
DO_DELETE = False #@param type:"boolean"
if DO_DELETE:
try:
tf.gfile.DeleteRecursively(OUTPUT_DIR)
except:
pass
tf.io.gfile.makedirs(OUTPUT_DIR)
print('***** Model output directory: *****'.format(OUTPUT_DIR))
### Load the data
data = pd.read_csv("data/bbc-text.csv")
data.columns = ['category', 'text']
print('*****Data Loaded: *****'.format(data.head()))
#check to see if any null values are present.
print('*****Empty Data: *****'.format(data[data.isnull().any(axis=1)]))
#encode category variable into numeric
data.category = pd.Categorical(data.category)
data['code'] = data.category.cat.codes
from sklearn.model_selection import train_test_split
train, test = train_test_split(data, test_size=0.2, random_state=200)
## 2 -- Data Visualisation
print(data.code.unique())
import matplotlib.pyplot as plt
train['code'].value_counts().plot(kind = 'bar')
DATA_COLUMN = 'text'
LABEL_COLUMN = 'code'
label_list = [0, 1, 2, 3, 4]
plt.show()
## 2 -- Data Preprocessing
train_InputExamples = train.apply(lambda x: bert.run_classifier.InputExample(guid=None,
text_a = x[DATA_COLUMN],
text_b = None,
label = x[LABEL_COLUMN]), axis = 1)
test_InputExamples = test.apply(lambda x: bert.run_classifier.InputExample(guid=None,
text_a = x[DATA_COLUMN],
text_b = None,
label = x[LABEL_COLUMN]), axis = 1)
# This is a path to an uncased (all lowercase) version of BERT
BERT_MODEL_HUB = "https://tfhub.dev/google/bert_uncased_L-12_H-768_A-12/1"
def create_tokenizer_from_hub_module():
"""Get the vocab file and casing info from the Hub module."""
with tf.Graph().as_default():
bert_module = hub.Module(BERT_MODEL_HUB)
tokenization_info = bert_module(signature="tokenization_info", as_dict=True)
with tf.compat.v1.Session() as sess:
vocab_file, do_lower_case = sess.run([tokenization_info["vocab_file"],
tokenization_info["do_lower_case"]])
return bert.tokenization.FullTokenizer(
vocab_file=vocab_file, do_lower_case=do_lower_case)
tokenizer = create_tokenizer_from_hub_module()
# We'll set sequences to be at most 128 tokens long.
MAX_SEQ_LENGTH = 128
# Convert our train and validation features to InputFeatures that BERT understands.
train_features = bert.run_classifier.convert_examples_to_features(train_InputExamples, label_list, MAX_SEQ_LENGTH, tokenizer)
test_features = bert.run_classifier.convert_examples_to_features(test_InputExamples, label_list, MAX_SEQ_LENGTH, tokenizer)
#Example on first observation in the training set
print("Example of train[0] as a training set")
print("Sentence : ", train_InputExamples.iloc[0].text_a)
print("-"*30)
print("Tokens : ", tokenizer.tokenize(train_InputExamples.iloc[0].text_a))
print("-"*30)
print("Input IDs : ", train_features[0].input_ids)
print("-"*30)
print("Input Masks : ", train_features[0].input_mask)
print("-"*30)
print("Segment IDs : ", train_features[0].segment_ids)
## 3. Creating a Multiclass Classifier
def create_model(is_predicting, input_ids, input_mask, segment_ids, labels,
num_labels):
bert_module = hub.Module(
BERT_MODEL_HUB,
trainable=True)
bert_inputs = dict(
input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids)
bert_outputs = bert_module(
inputs=bert_inputs,
signature="tokens",
as_dict=True)
# Use "pooled_output" for classification tasks on an entire sentence.
# Use "sequence_outputs" for token-level output.
output_layer = bert_outputs["pooled_output"]
hidden_size = output_layer.shape[-1].value
# Create our own layer to tune for politeness data.
output_weights = tf.compat.v1.get_variable(
"output_weights", [num_labels, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.compat.v1.get_variable(
"output_bias", [num_labels], initializer=tf.zeros_initializer())
with tf.compat.v1.variable_scope("loss"):
# Dropout helps prevent overfitting
output_layer = tf.nn.dropout(output_layer, keep_prob=0.9)
logits = tf.matmul(output_layer, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
log_probs = tf.nn.log_softmax(logits, axis=-1)
# Convert labels into one-hot encoding
one_hot_labels = tf.one_hot(labels, depth=num_labels, dtype=tf.float32)
predicted_labels = tf.squeeze(tf.argmax(log_probs, axis=-1, output_type=tf.int32))
# If we're predicting, we want predicted labels and the probabiltiies.
if is_predicting:
return (predicted_labels, log_probs)
# If we're train/eval, compute loss between predicted and actual label
per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1)
loss = tf.reduce_mean(per_example_loss)
return (loss, predicted_labels, log_probs)
#A function that adapts our model to work for training, evaluation, and prediction.
# model_fn_builder actually creates our model function
# using the passed parameters for num_labels, learning_rate, etc.
def model_fn_builder(num_labels, learning_rate, num_train_steps,
num_warmup_steps):
"""Returns `model_fn` closure for TPUEstimator."""
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
label_ids = features["label_ids"]
is_predicting = (mode == tf.estimator.ModeKeys.PREDICT)
# TRAIN and EVAL
if not is_predicting:
(loss, predicted_labels, log_probs) = create_model(
is_predicting, input_ids, input_mask, segment_ids, label_ids, num_labels)
train_op = bert.optimization.create_optimizer(
loss, learning_rate, num_train_steps, num_warmup_steps, use_tpu=False)
# Calculate evaluation metrics.
def metric_fn(label_ids, predicted_labels):
accuracy = tf.compat.v1.metrics.accuracy(label_ids, predicted_labels)
true_pos = tf.compat.v1.metrics.true_positives(
label_ids,
predicted_labels)
true_neg = tf.compat.v1.metrics.true_negatives(
label_ids,
predicted_labels)
false_pos = tf.compat.v1.metrics.false_positives(
label_ids,
predicted_labels)
false_neg = tf.compat.v1.metrics.false_negatives(
label_ids,
predicted_labels)
return
"eval_accuracy": accuracy,
"true_positives": true_pos,
"true_negatives": true_neg,
"false_positives": false_pos,
"false_negatives": false_neg
eval_metrics = metric_fn(label_ids, predicted_labels)
if mode == tf.estimator.ModeKeys.TRAIN:
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
else:
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
eval_metric_ops=eval_metrics)
else:
(predicted_labels, log_probs) = create_model(
is_predicting, input_ids, input_mask, segment_ids, label_ids, num_labels)
predictions =
'probabilities': log_probs,
'labels': predicted_labels
return tf.estimator.EstimatorSpec(mode, predictions=predictions)
# Return the actual model function in the closure
return model_fn
# Compute train and warmup steps from batch size
# These hyperparameters are copied from this colab notebook (https://colab.sandbox.google.com/github/tensorflow/tpu/blob/master/tools/colab/bert_finetuning_with_cloud_tpus.ipynb)
BATCH_SIZE = 16
LEARNING_RATE = 2e-5
NUM_TRAIN_EPOCHS = 3.0
# Warmup is a period of time where the learning rate is small and gradually increases--usually helps training.
WARMUP_PROPORTION = 0.1
# Model configs
SAVE_CHECKPOINTS_STEPS = 300
SAVE_SUMMARY_STEPS = 100
# Compute train and warmup steps from batch size
num_train_steps = int(len(train_features) / BATCH_SIZE * NUM_TRAIN_EPOCHS)
num_warmup_steps = int(num_train_steps * WARMUP_PROPORTION)
# Specify output directory and number of checkpoint steps to save
run_config = tf.estimator.RunConfig(
model_dir=OUTPUT_DIR,
save_summary_steps=SAVE_SUMMARY_STEPS,
save_checkpoints_steps=SAVE_CHECKPOINTS_STEPS)
# Specify output directory and number of checkpoint steps to save
run_config = tf.estimator.RunConfig(
model_dir=OUTPUT_DIR,
save_summary_steps=SAVE_SUMMARY_STEPS,
save_checkpoints_steps=SAVE_CHECKPOINTS_STEPS)
#Initializing the model and the estimator
model_fn = model_fn_builder(
num_labels=len(label_list),
learning_rate=LEARNING_RATE,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps)
estimator = tf.estimator.Estimator(
model_fn=model_fn,
config=run_config,
params="batch_size": BATCH_SIZE)
# Create an input function for training. drop_remainder = True for using TPUs.
train_input_fn = bert.run_classifier.input_fn_builder(
features=train_features,
seq_length=MAX_SEQ_LENGTH,
is_training=True,
drop_remainder=False)
# Create an input function for validating. drop_remainder = True for using TPUs.
test_input_fn = run_classifier.input_fn_builder(
features=test_features,
seq_length=MAX_SEQ_LENGTH,
is_training=False,
drop_remainder=False)
# #Training the model
print(f'Beginning Training!')
current_time = datetime.now()
estimator.train(input_fn=train_input_fn, max_steps=num_train_steps)
print("Training took time ", datetime.now() - current_time)
#Evaluating the model with Validation set
accuracy = estimator.evaluate(input_fn=test_input_fn, steps=None)
# A method to get predictions
def getPrediction(in_sentences):
# A list to map the actual labels to the predictions
labels = ["business", "entertainment", "politics", "sports", "tech"]
# Transforming the test data into BERT accepted form
input_examples = [run_classifier.InputExample(guid="", text_a=x, text_b=None, label=0) for x in in_sentences]
# Creating input features for Test data
input_features = run_classifier.convert_examples_to_features(input_examples, label_list, MAX_SEQ_LENGTH, tokenizer)
# Predicting the classes
predict_input_fn = run_classifier.input_fn_builder(features=input_features, seq_length=MAX_SEQ_LENGTH,
is_training=False, drop_remainder=False)
predictions = estimator.predict(predict_input_fn)
return [(sentence, prediction['probabilities'], prediction['labels'], labels[prediction['labels']]) for
sentence, prediction in zip(in_sentences, predictions)]
pred_sentences = list(test['text'])
predictions = getPrediction(pred_sentences)
enc_labels = []
act_labels = []
for i in range(len(predictions)):
enc_labels.append(predictions[i][2])
act_labels.append(predictions[i][3])
pd.DataFrame(enc_labels, columns = ['category']).to_excel('data/submission_bert.xlsx', index = False)
## Random tester
#Classifying random sentences
tests = getPrediction(['Mr.Modi is the Indian Prime Minister',
'Gaming machines are powered by efficient micro processores and GPUs',
'That HBO TV series is really good',
'A trillion dollar economy '
])
【问题讨论】：
【参考方案1】：正如问题清楚地表明要保存模型，它是这样工作的：
import torch
torch.save(model, 'path/to/model')
saved_model = torch.load('path/to/model')
【讨论】：
【参考方案2】：我认为您可以将您的 model.ckpt-333.data-00000-of-00001 重命名为 bert_model.ckpt，然后以与使用非微调模型相同的方式使用它。例如，运行
python run_classifier.py \
--task_name=MRPC \
--do_predict=true \
--data_dir=$GLUE_DIR/MRPC \
--vocab_file=$BERT_BASE_DIR/vocab.txt \
--bert_config_file=$BERT_BASE_DIR/bert_config.json \
--init_checkpoint=$TRAINED_CLASSIFIER
--init_checkpoint 指向模型的目录，或者运行 bert-as-service
bert-serving-start -model_dir $TRAINED_CLASSIFIER
用右边的-model_dir。
【讨论】：
问题是保存模型而不是检查点。那么，它们不是不同的吗？ 据我了解，它们并没有什么不同【参考方案3】：你可以使用这些方法：
model = MyModel(num_classes).to(device)
optimizer = AdamW(model.parameters(), lr=2e-5, weight_decay=1e-2)
output_model = './models/nameOfYourModel.pth'
# save
def save(model, optimizer):
# save
torch.save(
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict()
, output_model)
save(model, optimizer)
# load
checkpoint = torch.load(output_model, map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
来源：https://github.com/huggingface/transformers/issues/7849#issuecomment-718726121
【讨论】：
以上是关于保存一个“微调”的伯特模型的主要内容，如果未能解决你的问题，请参考以下文章