为什么 spacy 会忘记旧的训练数据以及如何解决

Question

我正在尝试为 ner 训练一个 spacy 模型。我有一个包含 2940 行的数据集，我用这些数据训练了一个基本模型，让它的名字为 current_model，我得到了另外 10 个不同的数据集，每个数据集的行范围从 200 到 530 行，所以我使用 spacy 的 spacy.load("current_model") 加载了我的 current_model然后我使用我的每个数据集进行训练。我尝试使用test data 预测 ner，它可以识别新数据集中的 ner，但它似乎忘记了最旧数据集中的 ner。我这样做是为了减少培训时间。请在下面查看我的代码以查看我尝试过的操作

基础模型训练代码

import spacy
from spacy.util import minibatch,compounding
import random
from pathlib import Path
from spacy import displacy
import re
import pandas as pd
from bs4 import BeautifulSoup
from datetime import datetime
imporcytoolz import partition_all
import os
from os import path
import shutil
import json


df = pd.read_csv("new_annotations/dataset_transfer_learning1.csv")
def populate_train_data(df):
train_data = []
i =0
for d_index, row in df.iterrows():
    print(row["annotations"])
    content = row["annotations"].replace("\\n", "\n").replace("\n", " ")
    content = re.sub(r"(?<=[:])(?=[^\s])", r" ", content)

# Finding tags and entities and store values in a entity list-----
soup = BeautifulSoup(content, "html.parser")
text = soup.get_text()
entities = []
for tag in soup.find_all():
    if tag.string is None:
        # failing silently for invalid tag
        print(f'Tagging is invalid: {row["_id"], tag.name}, on row {i+2}skipping..')
        continue

tag_index = content.split(str(tag))[0].count(tag.string)
try:

for index, match in enumerate(re.finditer(tag.string.replace("*", " "), text)):

if index == tag_index:

entities.append((match.start(), match.end(), tag.name))

except Exception as e:

print(e, f"at line no {i+2}")

continue

i += 1

if entities:

train_data.append((text, {"entities": entities}))

return train_data



def train(training_data,old_training_data=None,model_name=None):

nlp = ""

pretrained_weights = Path('weights/model999.bin')

if model_name is not None:

nlp = spacy.load(model_name,weights=pretrained_weights)

else:

print("no model specified using default model")

nlp = spacy.load("en_core_web_sm")

if "ner" not in nlp.pipe_names:

print("there is no ner creating ner")

ner = nlp.create_pipe("ner")

nlp.add_pipe(ner,last=True)

else:

print("there is ner")

ner = nlp.get_pipe("ner")

for _,annotations in training_data:

for ent in annotations.get("entities"):

ner.add_label(ent[2])

start_time = time.time()

if model_name is not None:

# nlp.resume_training()

# TRAINING_DATA = populate_train_data(pd.read_csv(old_training_data))

TRAINING_DATA = old_training_data

revision_data =[]

for doc in nlp.pipe(list(zip(*TRAINING_DATA))[0]):

tags = [w.tag_ for w in doc]

heads = [w.head.i for w in doc]

deps = [w.dep_ for w in doc]

entities = [(e.start_char, e.end_char, e.label_) for e in doc.ents]

revision_data.append((doc, GoldParse(doc, entities=entities)))

fine_tune_data = []

for raw_text, entity_offsets in training_data:

doc = nlp.make_doc(raw_text)

try:

gold = GoldParse(doc,entities=entity_offsets['entities'])

except ValueError:

pass

fine_tune_data.append((doc,gold))

other_pipes = [pipe for pipe in nlp.pipe_names if pipe != "ner"]

optimizer = nlp.entity.create_optimizer()

with nlp.disable_pipes(*other_pipes):

# pretrained_weights = Path('weights/model999.bin')

# with pretrained_weights.open("rb") as file_:

# ner.model.tok2vec.from_bytes(file_.read())

for i in range(20):

example_data = revision_data+fine_tune_data

# example_data = training_data

losses = {}

random.shuffle(example_data)

for batch in partition_all(2,example_data):

docs, golds = zip(*batch)

# print(docs, golds)

try:


nlp.update(docs,golds)

except ValueError:

pass

# print(losses)

else:

for i in range(20):

random.shuffle(training_data)

correct = 1

for text, annotations in training_data:

try:

nlp.update([text],[annotations])

print(correct)

correct +=1

except ValueError:

pass

# print("skipping..")

no_of_stars = i

print("*"*no_of_stars)

end_time = time.time()

print("this code took {}".format(end_time - start_time))

return nlp


def save_to_directory(nlp,directory_name):

save_directory = directory_name

for directory in save_directory:

if directory is not None:

directory_full_path = Path(directory+"_"+datetime.today().strftime('%Y_%m_%d'))

if path.exists(directory_full_path):

shutil.rmtree(directory_full_path)

print("folder already existed so removed")

if not directory_full_path.exists():

directory_full_path.mkdir()

nlp.to_disk(directory_full_path)

print("Saved model to output directory",directory)



if __name__ == "__main__":

training_data = populate_train_data(df)


# training_data = [

# ("I Like Today and Evening", {"entities":[(7,12,"DAY"),(17,24,"DAY")]}),

# ("Today is my lucky day", {"entities":[(1,5,"DAY")]}),

# ("Yesterday and Today are two same days of a month",{"entities":[(14,19,"DAY")]}),

# ("May Today is Best Day",{"entities":[(4,9,"DAY")]}),

# ("Have a Nice Today and Every Day",{"entities":[(12,17,"DAY")]}),

# ("Hey How are feeling Today",{"entities":[(20,25,"DAY")]}),

# ]

# print(training_data)

nlp = train(training_data)

save_to_directory(nlp,["trained_model_with_transfer_learning"])cytoolz import partition_all

import os

from os import path

import shutil

import json



df = pd.read_csv("new_annotations/dataset_transfer_learning1.csv")

def populate_train_data(df):

train_data = []

i =0

for d_index, row in df.iterrows():

print(row["annotations"])

content = row["annotations"].replace("\\n", "\n").replace("\n", " ")

content = re.sub(r"(?<=[:])(?=[^\s])", r" ", content)


# Finding tags and entities and store values in a entity list-----

soup = BeautifulSoup(content, "html.parser")

text = soup.get_text()

entities = []

for tag in soup.find_all():

if tag.string is None:

# failing silently for invalid tag

print(f'Tagging is invalid: {row["_id"], tag.name}, on row {i+2}skipping..')

continue


tag_index = content.split(str(tag))[0].count(tag.string)

try:

for index, match in enumerate(re.finditer(tag.string.replace("*", " "), text)):

if index == tag_index:

entities.append((match.start(), match.end(), tag.name))

except Exception as e:

print(e, f"at line no {i+2}")

continue

i += 1

if entities:

train_data.append((text, {"entities": entities}))

return train_data



def train(training_data,old_training_data=None,model_name=None):

nlp = ""

pretrained_weights = Path('weights/model999.bin')

if model_name is not None:

nlp = spacy.load(model_name,weights=pretrained_weights)

else:

print("no model specified using default model")

nlp = spacy.load("en_core_web_sm")

if "ner" not in nlp.pipe_names:

print("there is no ner creating ner")

ner = nlp.create_pipe("ner")

nlp.add_pipe(ner,last=True)

else:

print("there is ner")

ner = nlp.get_pipe("ner")

for _,annotations in training_data:

for ent in annotations.get("entities"):

ner.add_label(ent[2])

start_time = time.time()

if model_name is not None:

# nlp.resume_training()

# TRAINING_DATA = populate_train_data(pd.read_csv(old_training_data))

TRAINING_DATA = old_training_data

revision_data =[]

for doc in nlp.pipe(list(zip(*TRAINING_DATA))[0]):

tags = [w.tag_ for w in doc]

heads = [w.head.i for w in doc]

deps = [w.dep_ for w in doc]

entities = [(e.start_char, e.end_char, e.label_) for e in doc.ents]

revision_data.append((doc, GoldParse(doc, entities=entities)))

fine_tune_data = []

for raw_text, entity_offsets in training_data:

doc = nlp.make_doc(raw_text)

try:

gold = GoldParse(doc,entities=entity_offsets['entities'])

except ValueError:

pass

fine_tune_data.append((doc,gold))

other_pipes = [pipe for pipe in nlp.pipe_names if pipe != "ner"]

optimizer = nlp.entity.create_optimizer()

with nlp.disable_pipes(*other_pipes):

# pretrained_weights = Path('weights/model999.bin')

# with pretrained_weights.open("rb") as file_:

# ner.model.tok2vec.from_bytes(file_.read())

for i in range(20):

example_data = revision_data+fine_tune_data

# example_data = training_data

losses = {}

random.shuffle(example_data)

for batch in partition_all(2,example_data):

docs, golds = zip(*batch)

# print(docs, golds)

try:


nlp.update(docs,golds)

except ValueError:

pass

# print(losses)

else:

for i in range(20):

random.shuffle(training_data)

correct = 1

for text, annotations in training_data:

try:

nlp.update([text],[annotations])

print(correct)

correct +=1

except ValueError:

pass

# print("skipping..")

no_of_stars = i

print("*"*no_of_stars)

end_time = time.time()

print("this code took {}".format(end_time - start_time))

return nlp


def save_to_directory(nlp,directory_name):

save_directory = directory_name

for directory in save_directory:

if directory is not None:

directory_full_path = Path(directory+"_"+datetime.today().strftime('%Y_%m_%d'))

if path.exists(directory_full_path):

shutil.rmtree(directory_full_path)

print("folder already existed so removed")

if not directory_full_path.exists():

directory_full_path.mkdir()

nlp.to_disk(directory_full_path)

print("Saved model to output directory",directory)



if __name__ == "__main__":

training_data = populate_train_data(df)


# training_data = [

# ("I Like Today and Evening", {"entities":[(7,12,"DAY"),(17,24,"DAY")]}),

# ("Today is my lucky day", {"entities":[(1,5,"DAY")]}),

# ("Yesterday and Today are two same days of a month",{"entities":[(14,19,"DAY")]}),

# ("May Today is Best Day",{"entities":[(4,9,"DAY")]}),

# ("Have a Nice Today and Every Day",{"entities":[(12,17,"DAY")]}),

# ("Hey How are feeling Today",{"entities":[(20,25,"DAY")]}),

# ]

# print(training_data)

nlp = train(training_data)

save_to_directory(nlp,["trained_model_with_transfer_learning"])


#to do train using batched

#add drop rate

使用新数据集进行训练并保存到另一个目录的代码

注意：下面的代码是写新文件的。

import spacy

from spacy import displacy

import pandas as pd

from annotations_training_spacy_31_oct_2019 import populate_train_data,train,save_to_directory



# test_texts = "I Like Today and Evening"

# base_training_data = [

# ("I Like Today and Evening", {"entities":[(7,12,"DAY"),(17,24,"DAY")]}),

# ("Today is my lucky day", {"entities":[(1,5,"DAY")]}),

# ("Yesterday and Today are two same days of a month",{"entities":[(14,19,"DAY")]}),

# ("May Today is Best Day",{"entities":[(4,9,"DAY")]}),

# ("Have a Nice Today and Every Day",{"entities":[(12,17,"DAY")]}),

# ("Hey How are feeling Today",{"entities":[(20,25,"DAY")]}),

# ]

test_text = test_texts


# new_data_set = [

# ("Today is an Awsome Day", {"entities":[(1,5,"DAY")]}),

# ]


nlp = train(training_data=new_data_set,old_training_data=base_training_data,model_name="trained_model_with_transfer_learning_8_2019_12_05")

save_to_directory(nlp,["trained_model_with_transfer_learning_9"])


doc = nlp(test_text)

print("ENTITIES in '%s'" % test_text)

nlp.add_pipe(nlp.create_pipe('sentencizer'))

sentence = list(doc.sents)

for ent in doc.ents:

print(ent.label_,ent.text)



displacy.serve(sentence, style='ent')

如您所见，我还尝试加载旧数据集标签。但是我还是有这个问题

我知道有些人遇到了这个问题，如果有人解决了这个问题，请帮助我。

提前感谢您帮助的朋友。

嗨，

我正在尝试为 ner 训练一个 spacy 模型。我有一个包含 2940 行的数据集，并且我训练了一个基础

模型用这些数据命名为current_model，我得到了另外10个不同的数据集

每个都有从 200 到 530 行的行，所以我使用 spacy 的 spacy.load("current_model") 加载了我的 current_model，然后我使用我的每个数据集进行了训练。我试图预测 ner

使用 test data 它可以识别新数据集中的 ner，但它似乎忘记了最旧数据集中的 ner

我这样做是为了减少培训时间。请在下面查看我的代码以查看我尝试过的内容

要做和我

基础模型训练代码

import spacy

from spacy.util import minibatch,compounding

import random

from pathlib import Path

from spacy import displacy

import re

import pandas as pd

from bs4 import BeautifulSoup

from datetime import datetime

imporcytoolz import partition_all

import os

from os import path

import shutil

import json



df = pd.read_csv("new_annotations/dataset_transfer_learning1.csv")

def populate_train_data(df):

train_data = []

i =0

for d_index, row in df.iterrows():

print(row["annotations"])

content = row["annotations"].replace("\\n", "\n").replace("\n", " ")

content = re.sub(r"(?<=[:])(?=[^\s])", r" ", content)


# Finding tags and entities and store values in a entity list-----

soup = BeautifulSoup(content, "html.parser")

text = soup.get_text()

entities = []

for tag in soup.find_all():

if tag.string is None:

# failing silently for invalid tag

print(f'Tagging is invalid: {row["_id"], tag.name}, on row {i+2}skipping..')

continue


tag_index = content.split(str(tag))[0].count(tag.string)

try:

for index, match in enumerate(re.finditer(tag.string.replace("*", " "), text)):

if index == tag_index:

entities.append((match.start(), match.end(), tag.name))

except Exception as e:

print(e, f"at line no {i+2}")

continue

i += 1

if entities:

train_data.append((text, {"entities": entities}))

return train_data



def train(training_data,old_training_data=None,model_name=None):

nlp = ""

pretrained_weights = Path('weights/model999.bin')

if model_name is not None:

nlp = spacy.load(model_name,weights=pretrained_weights)

else:

print("no model specified using default model")

nlp = spacy.load("en_core_web_sm")

if "ner" not in nlp.pipe_names:

print("there is no ner creating ner")

ner = nlp.create_pipe("ner")

nlp.add_pipe(ner,last=True)

else:

print("there is ner")

ner = nlp.get_pipe("ner")

for _,annotations in training_data:

for ent in annotations.get("entities"):

ner.add_label(ent[2])

start_time = time.time()

if model_name is not None:

# nlp.resume_training()

# TRAINING_DATA = populate_train_data(pd.read_csv(old_training_data))

TRAINING_DATA = old_training_data

revision_data =[]

for doc in nlp.pipe(list(zip(*TRAINING_DATA))[0]):

tags = [w.tag_ for w in doc]

heads = [w.head.i for w in doc]

deps = [w.dep_ for w in doc]

entities = [(e.start_char, e.end_char, e.label_) for e in doc.ents]

revision_data.append((doc, GoldParse(doc, entities=entities)))

fine_tune_data = []

for raw_text, entity_offsets in training_data:

doc = nlp.make_doc(raw_text)

try:

gold = GoldParse(doc,entities=entity_offsets['entities'])

except ValueError:

pass

fine_tune_data.append((doc,gold))

other_pipes = [pipe for pipe in nlp.pipe_names if pipe != "ner"]

optimizer = nlp.entity.create_optimizer()

with nlp.disable_pipes(*other_pipes):

# pretrained_weights = Path('weights/model999.bin')

# with pretrained_weights.open("rb") as file_:

# ner.model.tok2vec.from_bytes(file_.read())

for i in range(20):

example_data = revision_data+fine_tune_data

# example_data = training_data

losses = {}

random.shuffle(example_data)

for batch in partition_all(2,example_data):

docs, golds = zip(*batch)

# print(docs, golds)

try:


nlp.update(docs,golds)

except ValueError:

pass

# print(losses)

else:

for i in range(20):

random.shuffle(training_data)

correct = 1

for text, annotations in training_data:

try:

nlp.update([text],[annotations])

print(correct)

correct +=1

except ValueError:

pass

# print("skipping..")

no_of_stars = i

print("*"*no_of_stars)

end_time = time.time()

print("this code took {}".format(end_time - start_time))

return nlp


def save_to_directory(nlp,directory_name):

save_directory = directory_name

for directory in save_directory:

if directory is not None:

directory_full_path = Path(directory+"_"+datetime.today().strftime('%Y_%m_%d'))

if path.exists(directory_full_path):

shutil.rmtree(directory_full_path)

print("folder already existed so removed")

if not directory_full_path.exists():

directory_full_path.mkdir()

nlp.to_disk(directory_full_path)

print("Saved model to output directory",directory)



if __name__ == "__main__":

training_data = populate_train_data(df)


# training_data = [

# ("I Like Today and Evening", {"entities":[(7,12,"DAY"),(17,24,"DAY")]}),

# ("Today is my lucky day", {"entities":[(1,5,"DAY")]}),

# ("Yesterday and Today are two same days of a month",{"entities":[(14,19,"DAY")]}),

# ("May Today is Best Day",{"entities":[(4,9,"DAY")]}),

# ("Have a Nice Today and Every Day",{"entities":[(12,17,"DAY")]}),

# ("Hey How are feeling Today",{"entities":[(20,25,"DAY")]}),

# ]

# print(training_data)

nlp = train(training_data)

save_to_directory(nlp,["trained_model_with_transfer_learning"])cytoolz import partition_all

import os

from os import path

import shutil

import json



df = pd.read_csv("new_annotations/dataset_transfer_learning1.csv")

def populate_train_data(df):

train_data = []

i =0

for d_index, row in df.iterrows():

print(row["annotations"])

content = row["annotations"].replace("\\n", "\n").replace("\n", " ")

content = re.sub(r"(?<=[:])(?=[^\s])", r" ", content)


# Finding tags and entities and store values in a entity list-----

soup = BeautifulSoup(content, "html.parser")

text = soup.get_text()

entities = []

for tag in soup.find_all():

if tag.string is None:

# failing silently for invalid tag

print(f'Tagging is invalid: {row["_id"], tag.name}, on row {i+2}skipping..')

continue


tag_index = content.split(str(tag))[0].count(tag.string)

try:

for index, match in enumerate(re.finditer(tag.string.replace("*", " "), text)):

if index == tag_index:

entities.append((match.start(), match.end(), tag.name))

except Exception as e:

print(e, f"at line no {i+2}")

continue

i += 1

if entities:

train_data.append((text, {"entities": entities}))

return train_data



def train(training_data,old_training_data=None,model_name=None):

nlp = ""

pretrained_weights = Path('weights/model999.bin')

if model_name is not None:

nlp = spacy.load(model_name,weights=pretrained_weights)

else:

print("no model specified using default model")

nlp = spacy.load("en_core_web_sm")

if "ner" not in nlp.pipe_names:

print("there is no ner creating ner")

ner = nlp.create_pipe("ner")

nlp.add_pipe(ner,last=True)

else:

print("there is ner")

ner = nlp.get_pipe("ner")

for _,annotations in training_data:

for ent in annotations.get("entities"):

ner.add_label(ent[2])

start_time = time.time()

if model_name is not None:

# nlp.resume_training()

# TRAINING_DATA = populate_train_data(pd.read_csv(old_training_data))

TRAINING_DATA = old_training_data

revision_data =[]

for doc in nlp.pipe(list(zip(*TRAINING_DATA))[0]):

tags = [w.tag_ for w in doc]

heads = [w.head.i for w in doc]

deps = [w.dep_ for w in doc]

entities = [(e.start_char, e.end_char, e.label_) for e in doc.ents]

revision_data.append((doc, GoldParse(doc, entities=entities)))

fine_tune_data = []

for raw_text, entity_offsets in training_data:

doc = nlp.make_doc(raw_text)

try:

gold = GoldParse(doc,entities=entity_offsets['entities'])

except ValueError:

pass

fine_tune_data.append((doc,gold))

other_pipes = [pipe for pipe in nlp.pipe_names if pipe != "ner"]

optimizer = nlp.entity.create_optimizer()

with nlp.disable_pipes(*other_pipes):

# pretrained_weights = Path('weights/model999.bin')

# with pretrained_weights.open("rb") as file_:

# ner.model.tok2vec.from_bytes(file_.read())

for i in range(20):

example_data = revision_data+fine_tune_data

# example_data = training_data

losses = {}

random.shuffle(example_data)

for batch in partition_all(2,example_data):

docs, golds = zip(*batch)

# print(docs, golds)

try:


nlp.update(docs,golds)

except ValueError:

pass

# print(losses)

else:

for i in range(20):

random.shuffle(training_data)

correct = 1

for text, annotations in training_data:

try:

nlp.update([text],[annotations])

print(correct)

correct +=1

except ValueError:

pass

# print("skipping..")

no_of_stars = i

print("*"*no_of_stars)

end_time = time.time()

print("this code took {}".format(end_time - start_time))

return nlp


def save_to_directory(nlp,directory_name):

save_directory = directory_name

for directory in save_directory:

if directory is not None:

directory_full_path = Path(directory+"_"+datetime.today().strftime('%Y_%m_%d'))

if path.exists(directory_full_path):

shutil.rmtree(directory_full_path)

print("folder already existed so removed")

if not directory_full_path.exists():

directory_full_path.mkdir()

nlp.to_disk(directory_full_path)

print("Saved model to output directory",directory)



if __name__ == "__main__":

training_data = populate_train_data(df)


# training_data = [

# ("I Like Today and Evening", {"entities":[(7,12,"DAY"),(17,24,"DAY")]}),

# ("Today is my lucky day", {"entities":[(1,5,"DAY")]}),

# ("Yesterday and Today are two same days of a month",{"entities":[(14,19,"DAY")]}),

# ("May Today is Best Day",{"entities":[(4,9,"DAY")]}),

# ("Have a Nice Today and Every Day",{"entities":[(12,17,"DAY")]}),

# ("Hey How are feeling Today",{"entities":[(20,25,"DAY")]}),

# ]

# print(training_data)

nlp = train(training_data)

save_to_directory(nlp,["trained_model_with_transfer_learning"])


#to do train using batched

#add drop rate

使用新数据集进行训练并保存到另一个目录的代码

注意：下面的代码是写新文件的。

import spacy

from spacy import displacy

import pandas as pd

from annotations_training_spacy_31_oct_2019 import populate_train_data,train,save_to_directory



# test_texts = "I Like Today and Evening"

# base_training_data = [

# ("I Like Today and Evening", {"entities":[(7,12,"DAY"),(17,24,"DAY")]}),

# ("Today is my lucky day", {"entities":[(1,5,"DAY")]}),

# ("Yesterday and Today are two same days of a month",{"entities":[(14,19,"DAY")]}),

# ("May Today is Best Day",{"entities":[(4,9,"DAY")]}),

# ("Have a Nice Today and Every Day",{"entities":[(12,17,"DAY")]}),

# ("Hey How are feeling Today",{"entities":[(20,25,"DAY")]}),

# ]

test_text = test_texts


# new_data_set = [

# ("Today is an Awsome Day", {"entities":[(1,5,"DAY")]}),

# ]


nlp = train(training_data=new_data_set,old_training_data=base_training_data,model_name="trained_model_with_transfer_learning_8_2019_12_05")

save_to_directory(nlp,["trained_model_with_transfer_learning_9"])


doc = nlp(test_text)

print("ENTITIES in '%s'" % test_text)

nlp.add_pipe(nlp.create_pipe('sentencizer'))

sentence = list(doc.sents)

for ent in doc.ents:

print(ent.label_,ent.text)



displacy.serve(sentence, style='ent')

如您所见，我还尝试加载旧数据集标签。但是我还是有这个问题

我知道有些人遇到了这个问题，如果有人解决了这个问题，请帮助我。

提前感谢您帮助的朋友。

Answer 1

您是在训练新模型还是附加到现有的 spacy 模型？如果你在后面做，所有的神经网络（学习权重，特征）都将被遗忘和错位，从而导致准确性损失。当我想训练 spacy 无法识别的韩国和日本名字时，我是根据经验告诉我的。你也可以试试 FastText、Flair 和 Polyglot，看看它是否能达到你的目的。尝试使用所有这些工具，你应该有很好的输出。这就是我最终使用的解决方案。