def one(id, train_i, args):
# 加载PaddleHub ERNIE预训练模型
module = hub.Module(name=args.model)
# ERNIE预训练模型输入变量inputs、输出变量outputs、以及模型program
inputs, outputs, program = module.context(trainable=True,
max_seq_len=args.max_seq_len)
# 加载竞赛数据集并使用ReadingComprehensionReader读取数据
dataset = DuReader(id)
reader = hub.reader.ReadingComprehensionReader(
dataset=dataset,
vocab_path=module.get_vocab_path(),
max_seq_len=args.max_seq_len,
doc_stride=128,
max_query_length=args.max_que_len)
# 取ERNIE的字级别预训练输出
seq_output = outputs["sequence_output"]
# 设置运行program所需的feed_list
feed_list = [
inputs["input_ids"].name,
inputs["position_ids"].name,
inputs["segment_ids"].name,
inputs["input_mask"].name,
]
# 选择Fine-tune优化策略
strategy = hub.AdamWeightDecayStrategy(
weight_decay=args.weight_decay,
learning_rate=args.learning_rate,
warmup_proportion=args.warmup_proportion)
# 设置运行配置
config = hub.RunConfig(eval_interval=200,
use_pyreader=False,
use_data_parallel=args.use_data_parallel,
use_cuda=args.use_gpu,
num_epoch=args.num_epoch,
batch_size=args.batch_size,
checkpoint_dir=args.checkpoint_dir + str(id),
strategy=strategy)
# 定义阅读理解Fine-tune Task
# 由于竞赛数据集与cmrc2018数据集格式比较相似，此处sub_task应为cmrc2018
# 否则运行可能出错
reading_comprehension_task = hub.ReadingComprehensionTask(
data_reader=reader,
feature=seq_output,
feed_list=feed_list,
config=config,
sub_task="cmrc2018",
)
reading_comprehension_task.main_program.random_seed = args.seed
change_task(reading_comprehension_task, id)
# 调用finetune_and_eval API，将会自动进行训练、评估以及保存最佳模型
reading_comprehension_task.finetune_and_eval()
# 竞赛数据集测试集部分数据用于预测
data = dataset.predict_examples
# 调用predict接口, 打开return_result(True)，将自动返回预测结果
all_prediction = reading_comprehension_task.predict(data=data,
return_result=True)
# 写入预测结果
json.dump(all_prediction,
open('./work/result/submit{}_{}.json'.format(train_i, id), 'w'),
ensure_ascii=False)
value = [id, reading_comprehension_task.best_score] + list(
args.__dict__.values())
value = [str(x) for x in value]
with open('./work/log/MRC_log.txt', 'a', encoding='utf-8') as f:
f.write(','.join(value) + ',-\n')
return reading_comprehension_task.best_score, value[2:]
# Use "sequence_output" for token-level output.
seq_output = outputs["sequence_output"]
# Setup feed list for data feeder
feed_list = [
inputs["input_ids"].name,
inputs["position_ids"].name,
inputs["segment_ids"].name,
inputs["input_mask"].name,
]
# Setup runing config for PaddleHub Finetune API
config = hub.RunConfig(
use_data_parallel=False,
use_cuda=args.use_gpu,
batch_size=args.batch_size,
checkpoint_dir=args.checkpoint_dir,
strategy=hub.AdamWeightDecayStrategy())
# Define a reading comprehension finetune task by PaddleHub's API
reading_comprehension_task = hub.ReadingComprehensionTask(
data_reader=reader,
feature=seq_output,
feed_list=feed_list,
config=config)
# Data to be predicted
data = dataset.dev_examples[:10]
reading_comprehension_task.predict(data=data)
dataset = hub.dataset.SQUAD(version_2_with_negative=False,
tokenizer=tokenizer,
max_seq_len=args.max_seq_len)
# dataset = hub.dataset.SQUAD(version_2_with_negative=True)
# Select fine-tune strategy, setup config and fine-tune
strategy = hub.AdamWeightDecayStrategy(
weight_decay=args.weight_decay,
learning_rate=args.learning_rate,
warmup_proportion=args.warmup_proportion)
# Setup RunConfig for PaddleHub Fine-tune API
config = hub.RunConfig(eval_interval=100,
use_data_parallel=args.use_data_parallel,
use_cuda=args.use_gpu,
num_epoch=args.num_epoch,
batch_size=args.batch_size,
checkpoint_dir=args.checkpoint_dir,
strategy=strategy)
# Define a reading comprehension fine-tune task by PaddleHub's API
reading_comprehension_task = hub.ReadingComprehensionTask(
dataset=dataset,
feature=outputs["sequence_output"],
config=config,
sub_task="squad",
)
# Fine-tune by PaddleHub's API
reading_comprehension_task.finetune_and_eval()
