提交 f3e8f301 编写于 作者: G Guo Sheng 提交者: guosheng

Merge pull request #54 from 0YuanZhang0/seq_tag

seq_tag
......@@ -6,7 +6,7 @@ Sequence Tagging,是一个序列标注模型,模型可用于实现,分词
|模型|Precision|Recall|F1-score|
|:-:|:-:|:-:|:-:|
|Lexical Analysis|88.26%|89.20%|88.73%|
|Lexical Analysis|89.57%|89.96%|89.76%|
## 2. 快速开始
......@@ -22,7 +22,7 @@ Sequence Tagging,是一个序列标注模型,模型可用于实现,分词
克隆工具集代码库到本地
```bash
git clone https://github.com/PaddlePaddle/hapi.git
cd hapi/sequence_tagging
cd hapi/examples/sequence_tagging
```
#### 3. 环境依赖
......@@ -70,7 +70,7 @@ python -u train.py \
--dynamic False
# --device: 使用gpu设备还是cpu设备
# --dynamic: 是否使用动态图模式进行训练,如果使用静态图训练,设置为True, 动态图设置为False
# --dynamic: 是否使用动态图模式进行训练,如果使用静态图训练,设置为False, 动态图设置为True
```
GPU上多卡训练
......@@ -84,7 +84,7 @@ python -m paddle.distributed.launch --selected_gpus=0,1,2,3 train.py \
--dynamic False
# --device: 使用gpu设备还是cpu设备
# --dynamic: 是否使用动态图模式进行训练,如果使用静态图训练,设置为True, 动态图设置为False
# --dynamic: 是否使用动态图模式进行训练,如果使用静态图训练,设置为False, 动态图设置为True
```
CPU上训练
......@@ -95,7 +95,7 @@ python -u train.py \
--dynamic False
# --device: 使用gpu设备还是cpu设备
# --dynamic: 是否使用动态图模式进行训练,如果使用静态图训练,设置为True, 动态图设置为False
# --dynamic: 是否使用动态图模式进行训练,如果使用静态图训练,设置为False, 动态图设置为True
```
### 模型预测
......@@ -105,15 +105,13 @@ python -u train.py \
python predict.py \
--init_from_checkpoint model_baseline/params \
--output_file predict.result \
--mode predict \
--device cpu \
--dynamic False
# --init_from_checkpoint: 初始化模型
# --output_file: 预测结果文件
# --device: 使用gpu还是cpu设备
# --mode: 开启模式, 设置为train时,进行训练,设置为predict时进行预测
# --dynamic: 是否使用动态图模式进行训练,如果使用静态图训练,设置为True, 动态图设置为False
# --dynamic: 是否使用动态图模式进行训练,如果使用静态图训练,设置为False, 动态图设置为True
```
### 模型评估
......@@ -123,14 +121,12 @@ python predict.py \
# baseline model
python eval.py \
--init_from_checkpoint ./model_baseline/params \
--mode predict \
--device cpu \
--dynamic False
# --init_from_checkpoint: 初始化模型
# --device: 使用gpu还是cpu设备
# --mode: 开启模式, 设置为train时,进行训练,设置为predict时进行预测
# --dynamic: 是否使用动态图模式进行训练,如果使用静态图训练,设置为True, 动态图设置为False
# --dynamic: 是否使用动态图模式进行训练,如果使用静态图训练,设置为False, 动态图设置为True
```
......@@ -168,7 +164,7 @@ Overall Architecture of GRU-CRF-MODEL
训练使用的数据可以由用户根据实际的应用场景,自己组织数据。除了第一行是 `text_a\tlabel` 固定的开头,后面的每行数据都是由两列组成,以制表符分隔,第一列是 utf-8 编码的中文文本,以 `\002` 分割,第二列是对应每个字的标注,以 `\002` 分隔。我们采用 IOB2 标注体系,即以 X-B 作为类型为 X 的词的开始,以 X-I 作为类型为 X 的词的持续,以 O 表示不关注的字(实际上,在词性、专名联合标注中,不存在 O )。示例如下:
```text
除\002了\002他\002续\002任\002十\002二\002届\002政\002协\002委\002员\002,\002马\002化\002腾\002,\002雷\002军\002,\002李\002彦\002宏\002也\002被\002推\002选\002为\002新\002一\002届\002全\002国\002人\002大\002代\002表\002或\002全\002国\002政\002协\002委\002员 p-B\002p-I\002r-B\002v-B\002v-I\002m-B\002m-I\002m-I\002ORG-B\002ORG-I\002n-B\002n-I\002w-B\002PER-B\002PER-I\002PER-I\002w-B\002PER-B\002PER-I\002w-B\002PER-B\002PER-I\002PER-I\002d-B\002p-B\002v-B\002v-I\002v-B\002a-B\002m-B\002m-I\002ORG-B\002ORG-I\002ORG-I\002ORG-I\002n-B\002n-I\002c-B\002n-B\002n-I\002ORG-B\002ORG-I\002n-B\002n-I
除\002了\002他\002续\002任\002十\002二\002届\002政\002协\002委\002员\002,\002马\002化\002腾\002,\002雷\002军\002,\002李\002彦\002宏\002也\002被\002推\002选\002为\002新\002一\002届\002全\002国\002人\002大\002代\002表\002或\002全\002国\002政\002协\002委\002员 p-B\002p-I\002r-B\002v-B\002v-I\002m-B\002m-I\002m-I\002ORG-B\002ORG-I\002n-B\002n-I\002w-B\002PER-B\002PER-I\002PER-I\002w-B\002PER-B\002PER-I\002w-B\002PER-B\002PER-I\002PER-I\002d-B\002p-B\002v-B\002v-I\002v-B\002a-B\002m-B\002m-I\002ORG-B\002ORG-I\002ORG-I\002ORG-I\002n-B\002n-I\002c-B\002n-B\002n-I\002ORG-B\002ORG-I\002n-B\002n-I
```
+ 我们随同代码一并发布了完全版的模型和相关的依赖数据。但是,由于模型的训练数据过于庞大,我们没有发布训练数据,仅在`data`目录下放置少数样本用以示例输入数据格式。
......@@ -196,6 +192,7 @@ Overall Architecture of GRU-CRF-MODEL
├── eval.py # 词法分析评估的脚本
├── downloads.py # 用于下载数据和模型的脚本
├── downloads.sh # 用于下载数据和模型的脚本
├── sequence_tagging.yaml # 模型训练、预测、评估相关配置参数
└──reader.py # 文件读取相关函数
```
......@@ -207,11 +204,11 @@ Overall Architecture of GRU-CRF-MODEL
```text
@article{jiao2018LAC,
title={Chinese Lexical Analysis with Deep Bi-GRU-CRF Network},
author={Jiao, Zhenyu and Sun, Shuqi and Sun, Ke},
journal={arXiv preprint arXiv:1807.01882},
year={2018},
url={https://arxiv.org/abs/1807.01882}
title={Chinese Lexical Analysis with Deep Bi-GRU-CRF Network},
author={Jiao, Zhenyu and Sun, Shuqi and Sun, Ke},
journal={arXiv preprint arXiv:1807.01882},
year={2018},
url={https://arxiv.org/abs/1807.01882}
}
```
### 如何贡献代码
......
......@@ -35,7 +35,7 @@ FILE_INFO = {
},
'MODEL': {
'name': 'sequence_tagging_dy.tar.gz',
'md5': "1125d374c03c8218b6e47325dcf607e3"
'md5': "6ba37ceea8f1f764ba1fe227295a6a3b"
},
}
......
......@@ -12,7 +12,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""
SequenceTagging network structure
SequenceTagging eval structure
"""
from __future__ import division
......@@ -25,18 +25,16 @@ import math
import argparse
import numpy as np
from train import SeqTagging
from train import SeqTagging, ChunkEval, LacLoss
from utils.configure import PDConfig
from utils.check import check_gpu, check_version
from utils.metrics import chunk_count
from reader import LacDataset, create_lexnet_data_generator, create_dataloader
from reader import LacDataset, LacDataLoader
work_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.append(os.path.join(work_dir, "../"))
from hapi.model import set_device, Input
import paddle.fluid as fluid
from paddle.fluid.optimizer import AdamOptimizer
from paddle.fluid.layers.utils import flatten
......@@ -44,51 +42,33 @@ def main(args):
place = set_device(args.device)
fluid.enable_dygraph(place) if args.dynamic else None
inputs = [Input([None, None], 'int64', name='words'),
Input([None], 'int64', name='length')]
inputs = [
Input(
[None, None], 'int64', name='words'), Input(
[None], 'int64', name='length'), Input(
[None, None], 'int64', name='target')
]
labels = [Input([None, None], 'int64', name='labels')]
feed_list = None if args.dynamic else [x.forward() for x in inputs]
dataset = LacDataset(args)
eval_path = args.test_file
chunk_evaluator = fluid.metrics.ChunkEvaluator()
chunk_evaluator.reset()
eval_generator = create_lexnet_data_generator(
args, reader=dataset, file_name=eval_path, place=place, mode="test")
eval_dataset = create_dataloader(
eval_generator, place, feed_list=feed_list)
eval_dataset = LacDataLoader(args, place, phase="test")
vocab_size = dataset.vocab_size
num_labels = dataset.num_labels
model = SeqTagging(args, vocab_size, num_labels)
optim = AdamOptimizer(
learning_rate=args.base_learning_rate,
parameter_list=model.parameters())
model = SeqTagging(args, vocab_size, num_labels, mode="test")
model.mode = "test"
model.prepare(inputs=inputs)
model.prepare(
metrics=ChunkEval(num_labels),
inputs=inputs,
labels=labels,
device=place)
model.load(args.init_from_checkpoint, skip_mismatch=True)
for data in eval_dataset():
if len(data) == 1:
batch_data = data[0]
targets = np.array(batch_data[2])
else:
batch_data = data
targets = batch_data[2].numpy()
inputs_data = [batch_data[0], batch_data[1]]
crf_decode, length = model.test(inputs=inputs_data)
num_infer_chunks, num_label_chunks, num_correct_chunks = chunk_count(crf_decode, targets, length, dataset.id2label_dict)
chunk_evaluator.update(num_infer_chunks, num_label_chunks, num_correct_chunks)
precision, recall, f1 = chunk_evaluator.eval()
print("[test] P: %.5f, R: %.5f, F1: %.5f" % (precision, recall, f1))
model.evaluate(eval_dataset.dataloader, batch_size=args.batch_size)
if __name__ == '__main__':
if __name__ == '__main__':
args = PDConfig(yaml_file="sequence_tagging.yaml")
args.build()
args.Print()
......
......@@ -12,7 +12,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""
SequenceTagging network structure
SequenceTagging predict structure
"""
from __future__ import division
......@@ -28,14 +28,13 @@ import numpy as np
from train import SeqTagging
from utils.check import check_gpu, check_version
from utils.configure import PDConfig
from reader import LacDataset, create_lexnet_data_generator, create_dataloader
from reader import LacDataset, LacDataLoader
work_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.append(os.path.join(work_dir, "../"))
from hapi.model import set_device, Input
import paddle.fluid as fluid
from paddle.fluid.optimizer import AdamOptimizer
from paddle.fluid.layers.utils import flatten
......@@ -43,26 +42,18 @@ def main(args):
place = set_device(args.device)
fluid.enable_dygraph(place) if args.dynamic else None
inputs = [Input([None, None], 'int64', name='words'),
Input([None], 'int64', name='length')]
inputs = [
Input(
[None, None], 'int64', name='words'), Input(
[None], 'int64', name='length')
]
feed_list = None if args.dynamic else [x.forward() for x in inputs]
dataset = LacDataset(args)
predict_path = args.predict_file
predict_generator = create_lexnet_data_generator(
args, reader=dataset, file_name=predict_path, place=place, mode="predict")
predict_dataset = create_dataloader(
predict_generator, place, feed_list=feed_list)
predict_dataset = LacDataLoader(args, place, phase="predict")
vocab_size = dataset.vocab_size
num_labels = dataset.num_labels
model = SeqTagging(args, vocab_size, num_labels)
optim = AdamOptimizer(
learning_rate=args.base_learning_rate,
parameter_list=model.parameters())
model = SeqTagging(args, vocab_size, num_labels, mode="predict")
model.mode = "test"
model.prepare(inputs=inputs)
......@@ -70,20 +61,20 @@ def main(args):
model.load(args.init_from_checkpoint, skip_mismatch=True)
f = open(args.output_file, "wb")
for data in predict_dataset():
if len(data) == 1:
for data in predict_dataset.dataloader:
if len(data) == 1:
input_data = data[0]
else:
else:
input_data = data
results, length = model.test(inputs=flatten(input_data))
for i in range(len(results)):
results, length = model.test_batch(inputs=flatten(input_data))
for i in range(len(results)):
word_len = length[i]
word_ids = results[i][: word_len]
word_ids = results[i][:word_len]
tags = [dataset.id2label_dict[str(id)] for id in word_ids]
f.write("\002".join(tags) + "\n")
if __name__ == '__main__':
if __name__ == '__main__':
args = PDConfig(yaml_file="sequence_tagging.yaml")
args.build()
args.Print()
......
......@@ -19,12 +19,19 @@ from __future__ import division
from __future__ import print_function
import io
import os
import leveldb
import numpy as np
import shutil
from functools import partial
import paddle
from paddle.io import BatchSampler, DataLoader, Dataset
from paddle.fluid.dygraph.parallel import ParallelEnv
from hapi.distributed import DistributedBatchSampler
class LacDataset(object):
class LacDataset(Dataset):
"""
Load lexical analysis dataset
"""
......@@ -34,6 +41,7 @@ class LacDataset(object):
self.label_dict_path = args.label_dict_path
self.word_rep_dict_path = args.word_rep_dict_path
self._load_dict()
self.examples = []
def _load_dict(self):
self.word2id_dict = self.load_kv_dict(
......@@ -108,152 +116,135 @@ class LacDataset(object):
label_ids.append(label_id)
return label_ids
def file_reader(self,
filename,
mode="train",
batch_size=32,
max_seq_len=126):
def file_reader(self, filename, phase="train"):
"""
yield (word_idx, target_idx) one by one from file,
or yield (word_idx, ) in `infer` mode
"""
def wrapper():
fread = io.open(filename, "r", encoding="utf-8")
if mode == "train":
headline = next(fread)
headline = headline.strip().split('\t')
assert len(headline) == 2 and headline[0] == "text_a" and headline[
1] == "label"
buf = []
for line in fread:
words, labels = line.strip("\n").split("\t")
if len(words) < 1:
continue
word_ids = self.word_to_ids(words.split("\002"))
label_ids = self.label_to_ids(labels.split("\002"))
assert len(word_ids) == len(label_ids)
words_len = np.int64(len(word_ids))
word_ids = word_ids[0:max_seq_len]
words_len = np.int64(len(word_ids))
word_ids += [0 for _ in range(max_seq_len - words_len)]
label_ids = label_ids[0:max_seq_len]
label_ids += [0 for _ in range(max_seq_len - words_len)]
assert len(word_ids) == len(label_ids)
yield word_ids, label_ids, words_len
elif mode == "test":
headline = next(fread)
self.phase = phase
with io.open(filename, "r", encoding="utf8") as fr:
if phase in ["train", "test"]:
headline = next(fr)
headline = headline.strip().split('\t')
assert len(headline) == 2 and headline[0] == "text_a" and headline[
1] == "label"
buf = []
for line in fread:
words, labels = line.strip("\n").split("\t")
if len(words) < 1:
continue
word_ids = self.word_to_ids(words.split("\002"))
label_ids = self.label_to_ids(labels.split("\002"))
assert len(word_ids) == len(label_ids)
words_len = np.int64(len(word_ids))
yield word_ids, label_ids, words_len
else:
for line in fread:
words = line.strip("\n").split('\t')[0]
if words == u"text_a":
assert len(headline) == 2 and headline[
0] == "text_a" and headline[1] == "label"
for line in fr:
line_str = line.strip("\n")
if len(line_str) < 1 and len(line_str.split('\t')) < 2:
continue
if "\002" not in words:
word_ids = self.word_to_ids(words)
else:
word_ids = self.word_to_ids(words.split("\002"))
words_len = np.int64(len(word_ids))
yield word_ids, words_len
fread.close()
self.examples.append(line_str)
else:
for idx, line in enumerate(fr):
words = line.strip("\n").split("\t")[0]
self.examples.append(words)
def __getitem__(self, idx):
line_str = self.examples[idx]
if self.phase in ["train", "test"]:
words, labels = line_str.split('\t')
word_ids = self.word_to_ids(words.split("\002"))
label_ids = self.label_to_ids(labels.split("\002"))
assert len(word_ids) == len(label_ids)
return word_ids, label_ids
else:
words = [w for w in line_str]
word_ids = self.word_to_ids(words)
return word_ids
def __len__(self):
return wrapper
return len(self.examples)
def create_lexnet_data_generator(args, reader, file_name, place, mode="train"):
def padding_data(max_len, batch_data):
def create_lexnet_data_generator(args, insts, phase="train"):
def padding_data(max_len, batch_data, if_len=False):
padding_batch_data = []
for data in batch_data:
padding_lens = []
for data in batch_data:
data = data[:max_len]
if if_len:
seq_len = np.int64(len(data))
padding_lens.append(seq_len)
data += [0 for _ in range(max_len - len(data))]
padding_batch_data.append(data)
return padding_batch_data
def wrapper():
if mode == "train":
batch_words, batch_labels, seq_lens = [], [], []
for epoch in xrange(args.epoch):
for instance in reader.file_reader(
file_name, mode, max_seq_len=args.max_seq_len)():
words, labels, words_len = instance
if len(seq_lens) < args.batch_size:
batch_words.append(words)
batch_labels.append(labels)
seq_lens.append(words_len)
if len(seq_lens) == args.batch_size:
yield batch_words, seq_lens, batch_labels, batch_labels
batch_words, batch_labels, seq_lens = [], [], []
if len(seq_lens) > 0:
yield batch_words, seq_lens, batch_labels, batch_labels
elif mode == "test":
batch_words, batch_labels, seq_lens, max_len = [], [], [], 0
for instance in reader.file_reader(
file_name, mode, max_seq_len=args.max_seq_len)():
words, labels, words_len = instance
max_len = words_len if words_len > max_len else max_len
if len(seq_lens) < args.batch_size:
batch_words.append(words)
seq_lens.append(words_len)
batch_labels.append(labels)
if len(seq_lens) == args.batch_size:
padding_batch_words = padding_data(max_len, batch_words)
padding_batch_labels = padding_data(max_len, batch_labels)
yield padding_batch_words, seq_lens, padding_batch_labels, padding_batch_labels
batch_words, batch_labels, seq_lens, max_len = [], [], [], 0
if len(seq_lens) > 0:
padding_batch_words = padding_data(max_len, batch_words)
padding_batch_labels = padding_data(max_len, batch_labels)
yield padding_batch_words, seq_lens, padding_batch_labels, padding_batch_labels
else:
batch_words, seq_lens, max_len = [], [], 0
for instance in reader.file_reader(
file_name, mode, max_seq_len=args.max_seq_len)():
words, words_len = instance
if len(seq_lens) < args.batch_size:
batch_words.append(words)
seq_lens.append(words_len)
max_len = words_len if words_len > max_len else max_len
if len(seq_lens) == args.batch_size:
padding_batch_words = padding_data(max_len, batch_words)
yield padding_batch_words, seq_lens
batch_words, seq_lens, max_len = [], [], 0
if len(seq_lens) > 0:
padding_batch_words = padding_data(max_len, batch_words)
yield padding_batch_words, seq_lens
return wrapper
def create_dataloader(generator, place, feed_list=None):
if not feed_list:
data_loader = paddle.io.DataLoader.from_generator(
capacity=50,
use_double_buffer=True,
iterable=True,
return_list=True)
if if_len:
return np.array(padding_batch_data), np.array(padding_lens)
else:
return np.array(padding_batch_data)
if phase == "train":
batch_words = [inst[0] for inst in insts]
batch_labels = [inst[1] for inst in insts]
padding_batch_words, padding_lens = padding_data(
args.max_seq_len, batch_words, if_len=True)
padding_batch_labels = padding_data(args.max_seq_len, batch_labels)
return [
padding_batch_words, padding_lens, padding_batch_labels,
padding_batch_labels
]
elif phase == "test":
batch_words = [inst[0] for inst in insts]
seq_len = [len(inst[0]) for inst in insts]
max_seq_len = max(seq_len)
batch_labels = [inst[1] for inst in insts]
padding_batch_words, padding_lens = padding_data(
max_seq_len, batch_words, if_len=True)
padding_batch_labels = padding_data(max_seq_len, batch_labels)
return [
padding_batch_words, padding_lens, padding_batch_labels,
padding_batch_labels
]
else:
data_loader = paddle.io.DataLoader.from_generator(
feed_list=feed_list,
capacity=50,
use_double_buffer=True,
iterable=True,
batch_words = insts
seq_len = [len(inst) for inst in insts]
max_seq_len = max(seq_len)
padding_batch_words, padding_lens = padding_data(
max_seq_len, batch_words, if_len=True)
return [padding_batch_words, padding_lens]
class LacDataLoader(object):
def __init__(self,
args,
place,
phase="train",
shuffle=False,
num_workers=0,
drop_last=False):
assert phase in [
"train", "test", "predict"
], "phase should be in [train, test, predict], but get %s" % phase
if phase == "train":
file_name = args.train_file
elif phase == "test":
file_name = args.test_file
elif phase == "predict":
file_name = args.predict_file
self.dataset = LacDataset(args)
self.dataset.file_reader(file_name, phase=phase)
if phase == "train":
self.sampler = DistributedBatchSampler(
dataset=self.dataset,
batch_size=args.batch_size,
shuffle=shuffle,
drop_last=drop_last)
else:
self.sampler = BatchSampler(
dataset=self.dataset,
batch_size=args.batch_size,
shuffle=shuffle,
drop_last=drop_last)
self.dataloader = DataLoader(
dataset=self.dataset,
batch_sampler=self.sampler,
places=place,
collate_fn=partial(
create_lexnet_data_generator, args, phase=phase),
num_workers=num_workers,
return_list=True)
data_loader.set_batch_generator(generator, places=place)
return data_loader
word_dict_path: "./conf/word.dic"
label_dict_path: "./conf/tag.dic"
word_rep_dict_path: "./conf/q2b.dic"
device: "cpu"
device: "gpu"
dynamic: True
epoch: 10
base_learning_rate: 0.001
......@@ -14,7 +14,7 @@ batch_size: 300
max_seq_len: 126
num_devices: 1
save_dir: "model"
init_from_checkpoint: "model_baseline/params"
init_from_checkpoint: ""
init_from_pretrain_model: ""
save_freq: 1
eval_freq: 1
......@@ -22,4 +22,3 @@ output_file: "predict.result"
test_file: "./data/test.tsv"
train_file: "./data/train.tsv"
predict_file: "./data/infer.tsv"
mode: "train"
......@@ -28,21 +28,23 @@ import numpy as np
work_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.append(os.path.join(work_dir, "../"))
from hapi.metrics import Metric
from hapi.model import Model, Input, Loss, set_device
from hapi.text.text import SequenceTagging
from utils.check import check_gpu, check_version
from utils.configure import PDConfig
from reader import LacDataset, create_lexnet_data_generator, create_dataloader
from reader import LacDataset, LacDataLoader
import paddle.fluid as fluid
from paddle.fluid.optimizer import AdamOptimizer
__all__ = ["SeqTagging", "LacLoss", "ChunkEval"]
class SeqTagging(Model):
def __init__(self, args, vocab_size, num_labels, length=None):
def __init__(self, args, vocab_size, num_labels, length=None,
mode="train"):
super(SeqTagging, self).__init__()
"""
define the lexical analysis network structure
......@@ -53,7 +55,7 @@ class SeqTagging(Model):
for infer: return the prediction
otherwise: return the prediction
"""
self.mode_type = args.mode
self.mode_type = mode
self.word_emb_dim = args.word_emb_dim
self.vocab_size = vocab_size
self.num_labels = num_labels
......@@ -65,19 +67,19 @@ class SeqTagging(Model):
self.bigru_num = args.bigru_num
self.batch_size = args.batch_size
self.init_bound = 0.1
self.length=length
self.length = length
self.sequence_tagging = SequenceTagging(
vocab_size=self.vocab_size,
num_labels=self.num_labels,
batch_size=self.batch_size,
word_emb_dim=self.word_emb_dim,
grnn_hidden_dim=self.grnn_hidden_dim,
emb_learning_rate=self.emb_lr,
crf_learning_rate=self.crf_lr,
bigru_num=self.bigru_num,
init_bound=self.init_bound,
length=self.length)
vocab_size=self.vocab_size,
num_labels=self.num_labels,
batch_size=self.batch_size,
word_emb_dim=self.word_emb_dim,
grnn_hidden_dim=self.grnn_hidden_dim,
emb_learning_rate=self.emb_lr,
crf_learning_rate=self.crf_lr,
bigru_num=self.bigru_num,
init_bound=self.init_bound,
length=self.length)
def forward(self, *inputs):
"""
......@@ -85,10 +87,10 @@ class SeqTagging(Model):
"""
word = inputs[0]
lengths = inputs[1]
if self.mode_type == "train" or self.mode_type == "test":
if self.mode_type == "train" or self.mode_type == "test":
target = inputs[2]
outputs = self.sequence_tagging(word, lengths, target)
else:
else:
outputs = self.sequence_tagging(word, lengths)
return outputs
......@@ -156,7 +158,7 @@ class ChunkEval(Metric):
int(math.ceil((num_labels - 1) / 2.0)), "IOB")
self.reset()
def add_metric_op(self, *args):
def add_metric_op(self, *args):
crf_decode = args[0]
lengths = args[2]
label = args[3]
......@@ -207,30 +209,25 @@ def main(args):
place = set_device(args.device)
fluid.enable_dygraph(place) if args.dynamic else None
inputs = [Input([None, None], 'int64', name='words'),
Input([None], 'int64', name='length'),
Input([None, None], 'int64', name='target')]
inputs = [
Input(
[None, None], 'int64', name='words'), Input(
[None], 'int64', name='length'), Input(
[None, None], 'int64', name='target')
]
labels = [Input([None, None], 'int64', name='labels')]
feed_list = None if args.dynamic else [x.forward() for x in inputs + labels]
dataset = LacDataset(args)
train_path = args.train_file
test_path = args.test_file
train_generator = create_lexnet_data_generator(
args, reader=dataset, file_name=train_path, place=place, mode="train")
test_generator = create_lexnet_data_generator(
args, reader=dataset, file_name=test_path, place=place, mode="test")
feed_list = None if args.dynamic else [
x.forward() for x in inputs + labels
]
train_dataset = create_dataloader(
train_generator, place, feed_list=feed_list)
test_dataset = create_dataloader(
test_generator, place, feed_list=feed_list)
dataset = LacDataset(args)
train_dataset = LacDataLoader(args, place, phase="train")
vocab_size = dataset.vocab_size
num_labels = dataset.num_labels
model = SeqTagging(args, vocab_size, num_labels)
model = SeqTagging(args, vocab_size, num_labels, mode="train")
optim = AdamOptimizer(
learning_rate=args.base_learning_rate,
......@@ -250,8 +247,7 @@ def main(args):
if args.init_from_pretrain_model:
model.load(args.init_from_pretrain_model, reset_optimizer=True)
model.fit(train_dataset,
test_dataset,
model.fit(train_dataset.dataloader,
epochs=args.epoch,
batch_size=args.batch_size,
eval_freq=args.eval_freq,
......@@ -263,7 +259,7 @@ if __name__ == '__main__':
args = PDConfig(yaml_file="sequence_tagging.yaml")
args.build()
args.Print()
use_gpu = True if args.device == "gpu" else False
check_gpu(use_gpu)
check_version()
......
......@@ -195,13 +195,19 @@ class PDConfig(object):
"Whether to perform predicting.")
self.default_g.add_arg("do_eval", bool, False,
"Whether to perform evaluating.")
self.default_g.add_arg("do_save_inference_model", bool, False,
"Whether to perform model saving for inference.")
self.default_g.add_arg(
"do_save_inference_model", bool, False,
"Whether to perform model saving for inference.")
# NOTE: args for profiler
self.default_g.add_arg("is_profiler", int, 0, "the switch of profiler tools. (used for benchmark)")
self.default_g.add_arg("profiler_path", str, './', "the profiler output file path. (used for benchmark)")
self.default_g.add_arg("max_iter", int, 0, "the max train batch num.(used for benchmark)")
self.default_g.add_arg(
"is_profiler", int, 0,
"the switch of profiler tools. (used for benchmark)")
self.default_g.add_arg(
"profiler_path", str, './',
"the profiler output file path. (used for benchmark)")
self.default_g.add_arg("max_iter", int, 0,
"the max train batch num.(used for benchmark)")
self.parser = parser
......
......@@ -23,7 +23,7 @@ import paddle.fluid as fluid
__all__ = ['chunk_count', "build_chunk"]
def build_chunk(data_list, id2label_dict):
def build_chunk(data_list, id2label_dict):
"""
Assembly entity
"""
......@@ -31,29 +31,29 @@ def build_chunk(data_list, id2label_dict):
ner_dict = {}
ner_str = ""
ner_start = 0
for i in range(len(tag_list)):
for i in range(len(tag_list)):
tag = tag_list[i]
if tag == u"O":
if i != 0:
if tag == u"O":
if i != 0:
key = "%d_%d" % (ner_start, i - 1)
ner_dict[key] = ner_str
ner_start = i
ner_str = tag
elif tag.endswith(u"B"):
if i != 0:
ner_str = tag
elif tag.endswith(u"B"):
if i != 0:
key = "%d_%d" % (ner_start, i - 1)
ner_dict[key] = ner_str
ner_start = i
ner_str = tag.split('-')[0]
elif tag.endswith(u"I"):
if tag.split('-')[0] != ner_str:
if i != 0:
elif tag.endswith(u"I"):
if tag.split('-')[0] != ner_str:
if i != 0:
key = "%d_%d" % (ner_start, i - 1)
ner_dict[key] = ner_str
ner_start = i
ner_str = tag.split('-')[0]
return ner_dict
def chunk_count(infer_numpy, label_numpy, seq_len, id2label_dict):
"""
......@@ -62,15 +62,14 @@ def chunk_count(infer_numpy, label_numpy, seq_len, id2label_dict):
num_infer_chunks, num_label_chunks, num_correct_chunks = 0, 0, 0
assert infer_numpy.shape[0] == label_numpy.shape[0]
for i in range(infer_numpy.shape[0]):
infer_list = infer_numpy[i][: seq_len[i]]
label_list = label_numpy[i][: seq_len[i]]
for i in range(infer_numpy.shape[0]):
infer_list = infer_numpy[i][:seq_len[i]]
label_list = label_numpy[i][:seq_len[i]]
infer_dict = build_chunk(infer_list, id2label_dict)
num_infer_chunks += len(infer_dict)
label_dict = build_chunk(label_list, id2label_dict)
num_label_chunks += len(label_dict)
for key in infer_dict:
if key in label_dict and label_dict[key] == infer_dict[key]:
for key in infer_dict:
if key in label_dict and label_dict[key] == infer_dict[key]:
num_correct_chunks += 1
return num_infer_chunks, num_label_chunks, num_correct_chunks
此差异已折叠。
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册