@@ -19,7 +19,7 @@ class DotAttention(nn.Module): | |||||
补上文档 | 补上文档 | ||||
""" | """ | ||||
def __init__(self, key_size, value_size, dropout=0): | |||||
def __init__(self, key_size, value_size, dropout=0.0): | |||||
super(DotAttention, self).__init__() | super(DotAttention, self).__init__() | ||||
self.key_size = key_size | self.key_size = key_size | ||||
self.value_size = value_size | self.value_size = value_size | ||||
@@ -37,7 +37,7 @@ class DotAttention(nn.Module): | |||||
""" | """ | ||||
output = torch.matmul(Q, K.transpose(1, 2)) / self.scale | output = torch.matmul(Q, K.transpose(1, 2)) / self.scale | ||||
if mask_out is not None: | if mask_out is not None: | ||||
output.masked_fill_(mask_out, -1e8) | |||||
output.masked_fill_(mask_out, -1e18) | |||||
output = self.softmax(output) | output = self.softmax(output) | ||||
output = self.drop(output) | output = self.drop(output) | ||||
return torch.matmul(output, V) | return torch.matmul(output, V) | ||||
@@ -67,9 +67,8 @@ class MultiHeadAttention(nn.Module): | |||||
self.k_in = nn.Linear(input_size, in_size) | self.k_in = nn.Linear(input_size, in_size) | ||||
self.v_in = nn.Linear(input_size, in_size) | self.v_in = nn.Linear(input_size, in_size) | ||||
# follow the paper, do not apply dropout within dot-product | # follow the paper, do not apply dropout within dot-product | ||||
self.attention = DotAttention(key_size=key_size, value_size=value_size, dropout=0) | |||||
self.attention = DotAttention(key_size=key_size, value_size=value_size, dropout=dropout) | |||||
self.out = nn.Linear(value_size * num_head, input_size) | self.out = nn.Linear(value_size * num_head, input_size) | ||||
self.drop = TimestepDropout(dropout) | |||||
self.reset_parameters() | self.reset_parameters() | ||||
def reset_parameters(self): | def reset_parameters(self): | ||||
@@ -105,7 +104,7 @@ class MultiHeadAttention(nn.Module): | |||||
# concat all heads, do output linear | # concat all heads, do output linear | ||||
atte = atte.permute(1, 2, 0, 3).contiguous().view(batch, sq, -1) | atte = atte.permute(1, 2, 0, 3).contiguous().view(batch, sq, -1) | ||||
output = self.drop(self.out(atte)) | |||||
output = self.out(atte) | |||||
return output | return output | ||||
@@ -0,0 +1,111 @@ | |||||
import torch | |||||
import torch.nn as nn | |||||
import torch.nn.functional as F | |||||
from fastNLP.modules.decoder import ConditionalRandomField | |||||
from fastNLP.modules.encoder import Embedding | |||||
from fastNLP.core.utils import seq_len_to_mask | |||||
from fastNLP.core.const import Const as C | |||||
class IDCNN(nn.Module): | |||||
def __init__(self, init_embed, char_embed, | |||||
num_cls, | |||||
repeats, num_layers, num_filters, kernel_size, | |||||
use_crf=False, use_projection=False, block_loss=False, | |||||
input_dropout=0.3, hidden_dropout=0.2, inner_dropout=0.0): | |||||
super(IDCNN, self).__init__() | |||||
self.word_embeddings = Embedding(init_embed) | |||||
self.char_embeddings = Embedding(char_embed) | |||||
embedding_size = self.word_embeddings.embedding_dim + \ | |||||
self.char_embeddings.embedding_dim | |||||
self.conv0 = nn.Sequential( | |||||
nn.Conv1d(in_channels=embedding_size, | |||||
out_channels=num_filters, | |||||
kernel_size=kernel_size, | |||||
stride=1, dilation=1, | |||||
padding=kernel_size//2, | |||||
bias=True), | |||||
nn.ReLU(), | |||||
) | |||||
block = [] | |||||
for layer_i in range(num_layers): | |||||
dilated = 2 ** layer_i | |||||
block.append(nn.Conv1d( | |||||
in_channels=num_filters, | |||||
out_channels=num_filters, | |||||
kernel_size=kernel_size, | |||||
stride=1, dilation=dilated, | |||||
padding=(kernel_size//2) * dilated, | |||||
bias=True)) | |||||
block.append(nn.ReLU()) | |||||
self.block = nn.Sequential(*block) | |||||
if use_projection: | |||||
self.projection = nn.Sequential( | |||||
nn.Conv1d( | |||||
in_channels=num_filters, | |||||
out_channels=num_filters//2, | |||||
kernel_size=1, | |||||
bias=True), | |||||
nn.ReLU(),) | |||||
encode_dim = num_filters // 2 | |||||
else: | |||||
self.projection = None | |||||
encode_dim = num_filters | |||||
self.input_drop = nn.Dropout(input_dropout) | |||||
self.hidden_drop = nn.Dropout(hidden_dropout) | |||||
self.inner_drop = nn.Dropout(inner_dropout) | |||||
self.repeats = repeats | |||||
self.out_fc = nn.Conv1d( | |||||
in_channels=encode_dim, | |||||
out_channels=num_cls, | |||||
kernel_size=1, | |||||
bias=True) | |||||
self.crf = ConditionalRandomField( | |||||
num_tags=num_cls) if use_crf else None | |||||
self.block_loss = block_loss | |||||
def forward(self, words, chars, seq_len, target=None): | |||||
e1 = self.word_embeddings(words) | |||||
e2 = self.char_embeddings(chars) | |||||
x = torch.cat((e1, e2), dim=-1) # b,l,h | |||||
mask = seq_len_to_mask(seq_len) | |||||
x = x.transpose(1, 2) # b,h,l | |||||
last_output = self.conv0(x) | |||||
output = [] | |||||
for repeat in range(self.repeats): | |||||
last_output = self.block(last_output) | |||||
hidden = self.projection(last_output) if self.projection is not None else last_output | |||||
output.append(self.out_fc(hidden)) | |||||
def compute_loss(y, t, mask): | |||||
if self.crf is not None and target is not None: | |||||
loss = self.crf(y, t, mask) | |||||
else: | |||||
t.masked_fill_(mask == 0, -100) | |||||
loss = F.cross_entropy(y, t, ignore_index=-100) | |||||
return loss | |||||
if self.block_loss: | |||||
losses = [compute_loss(o, target, mask) for o in output] | |||||
loss = sum(losses) | |||||
else: | |||||
loss = compute_loss(output[-1], target, mask) | |||||
scores = output[-1] | |||||
if self.crf is not None: | |||||
pred = self.crf.viterbi_decode(scores, target, mask) | |||||
else: | |||||
pred = scores.max(1)[1] * mask.long() | |||||
return { | |||||
C.LOSS: loss, | |||||
C.OUTPUT: pred, | |||||
} | |||||
def predict(self, words, chars, seq_len): | |||||
return self.forward(words, chars, seq_len)[C.OUTPUT] |
@@ -9,6 +9,7 @@ from fastNLP import Const | |||||
# from reproduction.utils import check_dataloader_paths | # from reproduction.utils import check_dataloader_paths | ||||
from functools import partial | from functools import partial | ||||
class IMDBLoader(DataSetLoader): | class IMDBLoader(DataSetLoader): | ||||
""" | """ | ||||
读取IMDB数据集,DataSet包含以下fields: | 读取IMDB数据集,DataSet包含以下fields: | ||||
@@ -33,6 +34,7 @@ class IMDBLoader(DataSetLoader): | |||||
target = parts[0] | target = parts[0] | ||||
words = parts[1].lower().split() | words = parts[1].lower().split() | ||||
dataset.append(Instance(words=words, target=target)) | dataset.append(Instance(words=words, target=target)) | ||||
if len(dataset)==0: | if len(dataset)==0: | ||||
raise RuntimeError(f"{path} has no valid data.") | raise RuntimeError(f"{path} has no valid data.") | ||||
@@ -42,19 +44,32 @@ class IMDBLoader(DataSetLoader): | |||||
paths: Union[str, Dict[str, str]], | paths: Union[str, Dict[str, str]], | ||||
src_vocab_opt: VocabularyOption = None, | src_vocab_opt: VocabularyOption = None, | ||||
tgt_vocab_opt: VocabularyOption = None, | tgt_vocab_opt: VocabularyOption = None, | ||||
src_embed_opt: EmbeddingOption = None): | |||||
# paths = check_dataloader_paths(paths) | |||||
src_embed_opt: EmbeddingOption = None, | |||||
char_level_op=False): | |||||
datasets = {} | datasets = {} | ||||
info = DataInfo() | info = DataInfo() | ||||
for name, path in paths.items(): | for name, path in paths.items(): | ||||
dataset = self.load(path) | dataset = self.load(path) | ||||
datasets[name] = dataset | datasets[name] = dataset | ||||
def wordtochar(words): | |||||
chars = [] | |||||
for word in words: | |||||
word = word.lower() | |||||
for char in word: | |||||
chars.append(char) | |||||
return chars | |||||
if char_level_op: | |||||
for dataset in datasets.values(): | |||||
dataset.apply_field(wordtochar, field_name="words", new_field_name='chars') | |||||
datasets["train"], datasets["dev"] = datasets["train"].split(0.1, shuffle=False) | datasets["train"], datasets["dev"] = datasets["train"].split(0.1, shuffle=False) | ||||
src_vocab = Vocabulary() if src_vocab_opt is None else Vocabulary(**src_vocab_opt) | src_vocab = Vocabulary() if src_vocab_opt is None else Vocabulary(**src_vocab_opt) | ||||
src_vocab.from_dataset(datasets['train'], field_name='words') | src_vocab.from_dataset(datasets['train'], field_name='words') | ||||
src_vocab.index_dataset(*datasets.values(), field_name='words') | src_vocab.index_dataset(*datasets.values(), field_name='words') | ||||
tgt_vocab = Vocabulary(unknown=None, padding=None) \ | tgt_vocab = Vocabulary(unknown=None, padding=None) \ | ||||
@@ -78,3 +93,18 @@ class IMDBLoader(DataSetLoader): | |||||
dataset.set_target("target") | dataset.set_target("target") | ||||
return info | return info | ||||
if __name__=="__main__": | |||||
datapath = {"train": "/remote-home/ygwang/IMDB_data/train.csv", | |||||
"test": "/remote-home/ygwang/IMDB_data/test.csv"} | |||||
datainfo=IMDBLoader().process(datapath,char_level_op=True) | |||||
#print(datainfo.datasets["train"]) | |||||
len_count = 0 | |||||
for instance in datainfo.datasets["train"]: | |||||
len_count += len(instance["chars"]) | |||||
ave_len = len_count / len(datainfo.datasets["train"]) | |||||
print(ave_len) | |||||
@@ -0,0 +1,98 @@ | |||||
import csv | |||||
from typing import Iterable | |||||
from fastNLP import DataSet, Instance, Vocabulary | |||||
from fastNLP.core.vocabulary import VocabularyOption | |||||
from fastNLP.io.base_loader import DataInfo,DataSetLoader | |||||
from fastNLP.io.embed_loader import EmbeddingOption | |||||
from fastNLP.io.file_reader import _read_json | |||||
from typing import Union, Dict | |||||
from reproduction.Star_transformer.datasets import EmbedLoader | |||||
from reproduction.utils import check_dataloader_paths | |||||
class sst2Loader(DataSetLoader): | |||||
''' | |||||
数据来源"SST":'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FSST-2.zip?alt=media&token=aabc5f6b-e466-44a2-b9b4-cf6337f84ac8', | |||||
''' | |||||
def __init__(self): | |||||
super(sst2Loader, self).__init__() | |||||
def _load(self, path: str) -> DataSet: | |||||
ds = DataSet() | |||||
all_count=0 | |||||
csv_reader = csv.reader(open(path, encoding='utf-8'),delimiter='\t') | |||||
skip_row = 0 | |||||
for idx,row in enumerate(csv_reader): | |||||
if idx<=skip_row: | |||||
continue | |||||
target = row[1] | |||||
words = row[0].split() | |||||
ds.append(Instance(words=words,target=target)) | |||||
all_count+=1 | |||||
print("all count:", all_count) | |||||
return ds | |||||
def process(self, | |||||
paths: Union[str, Dict[str, str]], | |||||
src_vocab_opt: VocabularyOption = None, | |||||
tgt_vocab_opt: VocabularyOption = None, | |||||
src_embed_opt: EmbeddingOption = None, | |||||
char_level_op=False): | |||||
paths = check_dataloader_paths(paths) | |||||
datasets = {} | |||||
info = DataInfo() | |||||
for name, path in paths.items(): | |||||
dataset = self.load(path) | |||||
datasets[name] = dataset | |||||
def wordtochar(words): | |||||
chars=[] | |||||
for word in words: | |||||
word=word.lower() | |||||
for char in word: | |||||
chars.append(char) | |||||
return chars | |||||
input_name, target_name = 'words', 'target' | |||||
info.vocabs={} | |||||
# 就分隔为char形式 | |||||
if char_level_op: | |||||
for dataset in datasets.values(): | |||||
dataset.apply_field(wordtochar, field_name="words", new_field_name='chars') | |||||
src_vocab = Vocabulary() if src_vocab_opt is None else Vocabulary(**src_vocab_opt) | |||||
src_vocab.from_dataset(datasets['train'], field_name='words') | |||||
src_vocab.index_dataset(*datasets.values(), field_name='words') | |||||
tgt_vocab = Vocabulary(unknown=None, padding=None) \ | |||||
if tgt_vocab_opt is None else Vocabulary(**tgt_vocab_opt) | |||||
tgt_vocab.from_dataset(datasets['train'], field_name='target') | |||||
tgt_vocab.index_dataset(*datasets.values(), field_name='target') | |||||
info.vocabs = { | |||||
"words": src_vocab, | |||||
"target": tgt_vocab | |||||
} | |||||
info.datasets = datasets | |||||
if src_embed_opt is not None: | |||||
embed = EmbedLoader.load_with_vocab(**src_embed_opt, vocab=src_vocab) | |||||
info.embeddings['words'] = embed | |||||
return info | |||||
if __name__=="__main__": | |||||
datapath = {"train": "/remote-home/ygwang/workspace/GLUE/SST-2/train.tsv", | |||||
"dev": "/remote-home/ygwang/workspace/GLUE/SST-2/dev.tsv"} | |||||
datainfo=sst2Loader().process(datapath,char_level_op=True) | |||||
#print(datainfo.datasets["train"]) | |||||
len_count = 0 | |||||
for instance in datainfo.datasets["train"]: | |||||
len_count += len(instance["chars"]) | |||||
ave_len = len_count / len(datainfo.datasets["train"]) | |||||
print(ave_len) |
@@ -1,77 +1,203 @@ | |||||
from fastNLP.io.embed_loader import EmbeddingOption, EmbedLoader | |||||
import ast | |||||
import csv | |||||
from typing import Iterable | |||||
from fastNLP import DataSet, Instance, Vocabulary | |||||
from fastNLP.core.vocabulary import VocabularyOption | from fastNLP.core.vocabulary import VocabularyOption | ||||
from fastNLP.io.base_loader import DataSetLoader, DataInfo | |||||
from typing import Union, Dict, List, Iterator | |||||
from fastNLP import DataSet | |||||
from fastNLP import Instance | |||||
from fastNLP import Vocabulary | |||||
from fastNLP import Const | |||||
# from reproduction.utils import check_dataloader_paths | |||||
from functools import partial | |||||
import pandas as pd | |||||
from fastNLP.io import JsonLoader | |||||
from fastNLP.io.base_loader import DataInfo,DataSetLoader | |||||
from fastNLP.io.embed_loader import EmbeddingOption | |||||
from fastNLP.io.file_reader import _read_json | |||||
from typing import Union, Dict | |||||
from reproduction.utils import check_dataloader_paths | |||||
class yelpLoader(DataSetLoader): | |||||
def get_tokenizer(): | |||||
try: | |||||
import spacy | |||||
en = spacy.load('en') | |||||
print('use spacy tokenizer') | |||||
return lambda x: [w.text for w in en.tokenizer(x)] | |||||
except Exception as e: | |||||
print('use raw tokenizer') | |||||
return lambda x: x.split() | |||||
def clean_str(sentence, tokenizer, char_lower=False): | |||||
""" | """ | ||||
读取IMDB数据集,DataSet包含以下fields: | |||||
heavily borrowed from github | |||||
https://github.com/LukeZhuang/Hierarchical-Attention-Network/blob/master/yelp-preprocess.ipynb | |||||
:param sentence: is a str | |||||
:return: | |||||
""" | |||||
if char_lower: | |||||
sentence = sentence.lower() | |||||
import re | |||||
nonalpnum = re.compile('[^0-9a-zA-Z?!\']+') | |||||
words = tokenizer(sentence) | |||||
words_collection = [] | |||||
for word in words: | |||||
if word in ['-lrb-', '-rrb-', '<sssss>', '-r', '-l', 'b-']: | |||||
continue | |||||
tt = nonalpnum.split(word) | |||||
t = ''.join(tt) | |||||
if t != '': | |||||
words_collection.append(t) | |||||
words: list(str), 需要分类的文本 | |||||
target: str, 文本的标签 | |||||
return words_collection | |||||
class yelpLoader(DataSetLoader): | |||||
""" | """ | ||||
读取Yelp_full/Yelp_polarity数据集, DataSet包含fields: | |||||
words: list(str), 需要分类的文本 | |||||
target: str, 文本的标签 | |||||
chars:list(str),未index的字符列表 | |||||
def __init__(self): | |||||
数据集:yelp_full/yelp_polarity | |||||
:param fine_grained: 是否使用SST-5标准,若 ``False`` , 使用SST-2。Default: ``False`` | |||||
""" | |||||
def __init__(self, fine_grained=False,lower=False): | |||||
super(yelpLoader, self).__init__() | super(yelpLoader, self).__init__() | ||||
tag_v = {'1.0': 'very negative', '2.0': 'negative', '3.0': 'neutral', | |||||
'4.0': 'positive', '5.0': 'very positive'} | |||||
if not fine_grained: | |||||
tag_v['1.0'] = tag_v['2.0'] | |||||
tag_v['5.0'] = tag_v['4.0'] | |||||
self.fine_grained = fine_grained | |||||
self.tag_v = tag_v | |||||
self.lower = lower | |||||
self.tokenizer = get_tokenizer() | |||||
def _load(self, path): | |||||
dataset = DataSet() | |||||
data = pd.read_csv(path, header=None, sep=",").values | |||||
for line in data: | |||||
target = str(line[0]) | |||||
words = str(line[1]).lower().split() | |||||
dataset.append(Instance(words=words, target=target)) | |||||
if len(dataset)==0: | |||||
raise RuntimeError(f"{path} has no valid data.") | |||||
return dataset | |||||
''' | |||||
读取Yelp数据集, DataSet包含fields: | |||||
def process(self, | |||||
paths: Union[str, Dict[str, str]], | |||||
src_vocab_opt: VocabularyOption = None, | |||||
tgt_vocab_opt: VocabularyOption = None, | |||||
src_embed_opt: EmbeddingOption = None): | |||||
# paths = check_dataloader_paths(paths) | |||||
datasets = {} | |||||
info = DataInfo() | |||||
for name, path in paths.items(): | |||||
dataset = self.load(path) | |||||
datasets[name] = dataset | |||||
review_id: str, 22 character unique review id | |||||
user_id: str, 22 character unique user id | |||||
business_id: str, 22 character business id | |||||
useful: int, number of useful votes received | |||||
funny: int, number of funny votes received | |||||
cool: int, number of cool votes received | |||||
date: str, date formatted YYYY-MM-DD | |||||
words: list(str), 需要分类的文本 | |||||
target: str, 文本的标签 | |||||
数据来源: https://www.yelp.com/dataset/download | |||||
def _load_json(self, path): | |||||
ds = DataSet() | |||||
for idx, d in _read_json(path, fields=self.fields_list, dropna=self.dropna): | |||||
d = ast.literal_eval(d) | |||||
d["words"] = d.pop("text").split() | |||||
d["target"] = self.tag_v[str(d.pop("stars"))] | |||||
ds.append(Instance(**d)) | |||||
return ds | |||||
def _load_yelp2015_broken(self,path): | |||||
ds = DataSet() | |||||
with open (path,encoding='ISO 8859-1') as f: | |||||
row=f.readline() | |||||
all_count=0 | |||||
exp_count=0 | |||||
while row: | |||||
row=row.split("\t\t") | |||||
all_count+=1 | |||||
if len(row)>=3: | |||||
words=row[-1].split() | |||||
try: | |||||
target=self.tag_v[str(row[-2])+".0"] | |||||
ds.append(Instance(words=words, target=target)) | |||||
except KeyError: | |||||
exp_count+=1 | |||||
else: | |||||
exp_count+=1 | |||||
row = f.readline() | |||||
print("error sample count:",exp_count) | |||||
print("all count:",all_count) | |||||
return ds | |||||
''' | |||||
def _load(self, path): | |||||
ds = DataSet() | |||||
csv_reader=csv.reader(open(path,encoding='utf-8')) | |||||
all_count=0 | |||||
real_count=0 | |||||
for row in csv_reader: | |||||
all_count+=1 | |||||
if len(row)==2: | |||||
target=self.tag_v[row[0]+".0"] | |||||
words=clean_str(row[1],self.tokenizer,self.lower) | |||||
if len(words)!=0: | |||||
ds.append(Instance(words=words,target=target)) | |||||
real_count += 1 | |||||
print("all count:", all_count) | |||||
print("real count:", real_count) | |||||
return ds | |||||
datasets["train"], datasets["dev"] = datasets["train"].split(0.1, shuffle=False) | |||||
src_vocab = Vocabulary() if src_vocab_opt is None else Vocabulary(**src_vocab_opt) | |||||
src_vocab.from_dataset(datasets['train'], field_name='words') | |||||
src_vocab.index_dataset(*datasets.values(), field_name='words') | |||||
def process(self, paths: Union[str, Dict[str, str]], | |||||
train_ds: Iterable[str] = None, | |||||
src_vocab_op: VocabularyOption = None, | |||||
tgt_vocab_op: VocabularyOption = None, | |||||
embed_opt: EmbeddingOption = None, | |||||
char_level_op=False): | |||||
paths = check_dataloader_paths(paths) | |||||
datasets = {} | |||||
info = DataInfo(datasets=self.load(paths)) | |||||
src_vocab = Vocabulary() if src_vocab_op is None else Vocabulary(**src_vocab_op) | |||||
tgt_vocab = Vocabulary(unknown=None, padding=None) \ | tgt_vocab = Vocabulary(unknown=None, padding=None) \ | ||||
if tgt_vocab_opt is None else Vocabulary(**tgt_vocab_opt) | |||||
tgt_vocab.from_dataset(datasets['train'], field_name='target') | |||||
tgt_vocab.index_dataset(*datasets.values(), field_name='target') | |||||
if tgt_vocab_op is None else Vocabulary(**tgt_vocab_op) | |||||
_train_ds = [info.datasets[name] | |||||
for name in train_ds] if train_ds else info.datasets.values() | |||||
def wordtochar(words): | |||||
info.vocabs = { | |||||
"words": src_vocab, | |||||
"target": tgt_vocab | |||||
} | |||||
chars=[] | |||||
for word in words: | |||||
word=word.lower() | |||||
for char in word: | |||||
chars.append(char) | |||||
return chars | |||||
info.datasets = datasets | |||||
input_name, target_name = 'words', 'target' | |||||
info.vocabs={} | |||||
#就分隔为char形式 | |||||
if char_level_op: | |||||
for dataset in info.datasets.values(): | |||||
dataset.apply_field(wordtochar, field_name="words",new_field_name='chars') | |||||
# if embed_opt is not None: | |||||
# embed = EmbedLoader.load_with_vocab(**embed_opt, vocab=vocab) | |||||
# info.embeddings['words'] = embed | |||||
else: | |||||
src_vocab.from_dataset(*_train_ds, field_name=input_name) | |||||
src_vocab.index_dataset(*info.datasets.values(),field_name=input_name, new_field_name=input_name) | |||||
info.vocabs[input_name]=src_vocab | |||||
if src_embed_opt is not None: | |||||
embed = EmbedLoader.load_with_vocab(**src_embed_opt, vocab=src_vocab) | |||||
info.embeddings['words'] = embed | |||||
tgt_vocab.from_dataset(*_train_ds, field_name=target_name) | |||||
tgt_vocab.index_dataset( | |||||
*info.datasets.values(), | |||||
field_name=target_name, new_field_name=target_name) | |||||
for name, dataset in info.datasets.items(): | |||||
dataset.set_input("words") | |||||
dataset.set_target("target") | |||||
info.vocabs[target_name]=tgt_vocab | |||||
return info | return info | ||||
if __name__=="__main__": | |||||
testloader=yelpLoader() | |||||
# datapath = {"train": "/remote-home/ygwang/yelp_full/train.csv", | |||||
# "test": "/remote-home/ygwang/yelp_full/test.csv"} | |||||
#datapath={"train": "/remote-home/ygwang/yelp_full/test.csv"} | |||||
datapath = {"train": "/remote-home/ygwang/yelp_polarity/train.csv", | |||||
"test": "/remote-home/ygwang/yelp_polarity/test.csv"} | |||||
datainfo=testloader.process(datapath,char_level_op=True) | |||||
len_count=0 | |||||
for instance in datainfo.datasets["train"]: | |||||
len_count+=len(instance["chars"]) | |||||
ave_len=len_count/len(datainfo.datasets["train"]) | |||||
print(ave_len) |
@@ -1 +1,90 @@ | |||||
# TODO | |||||
''' | |||||
@author: https://github.com/ahmedbesbes/character-based-cnn | |||||
这里借鉴了上述链接中char-cnn model的代码,改动主要为将其改动为符合fastnlp的pipline | |||||
''' | |||||
import torch | |||||
import torch.nn as nn | |||||
from fastNLP.core.const import Const as C | |||||
class CharacterLevelCNN(nn.Module): | |||||
def __init__(self, args,embedding): | |||||
super(CharacterLevelCNN, self).__init__() | |||||
self.config=args.char_cnn_config | |||||
self.embedding=embedding | |||||
conv_layers = [] | |||||
for i, conv_layer_parameter in enumerate(self.config['model_parameters'][args.model_size]['conv']): | |||||
if i == 0: | |||||
#in_channels = args.number_of_characters + len(args.extra_characters) | |||||
in_channels = args.embedding_dim | |||||
out_channels = conv_layer_parameter[0] | |||||
else: | |||||
in_channels, out_channels = conv_layer_parameter[0], conv_layer_parameter[0] | |||||
if conv_layer_parameter[2] != -1: | |||||
conv_layer = nn.Sequential(nn.Conv1d(in_channels, | |||||
out_channels, | |||||
kernel_size=conv_layer_parameter[1], padding=0), | |||||
nn.ReLU(), | |||||
nn.MaxPool1d(conv_layer_parameter[2])) | |||||
else: | |||||
conv_layer = nn.Sequential(nn.Conv1d(in_channels, | |||||
out_channels, | |||||
kernel_size=conv_layer_parameter[1], padding=0), | |||||
nn.ReLU()) | |||||
conv_layers.append(conv_layer) | |||||
self.conv_layers = nn.ModuleList(conv_layers) | |||||
input_shape = (args.batch_size, args.max_length, | |||||
args.number_of_characters + len(args.extra_characters)) | |||||
dimension = self._get_conv_output(input_shape) | |||||
print('dimension :', dimension) | |||||
fc_layer_parameter = self.config['model_parameters'][args.model_size]['fc'][0] | |||||
fc_layers = nn.ModuleList([ | |||||
nn.Sequential( | |||||
nn.Linear(dimension, fc_layer_parameter), nn.Dropout(0.5)), | |||||
nn.Sequential(nn.Linear(fc_layer_parameter, | |||||
fc_layer_parameter), nn.Dropout(0.5)), | |||||
nn.Linear(fc_layer_parameter, args.num_classes), | |||||
]) | |||||
self.fc_layers = fc_layers | |||||
if args.model_size == 'small': | |||||
self._create_weights(mean=0.0, std=0.05) | |||||
elif args.model_size == 'large': | |||||
self._create_weights(mean=0.0, std=0.02) | |||||
def _create_weights(self, mean=0.0, std=0.05): | |||||
for module in self.modules(): | |||||
if isinstance(module, nn.Conv1d) or isinstance(module, nn.Linear): | |||||
module.weight.data.normal_(mean, std) | |||||
def _get_conv_output(self, shape): | |||||
input = torch.rand(shape) | |||||
output = input.transpose(1, 2) | |||||
# forward pass through conv layers | |||||
for i in range(len(self.conv_layers)): | |||||
output = self.conv_layers[i](output) | |||||
output = output.view(output.size(0), -1) | |||||
n_size = output.size(1) | |||||
return n_size | |||||
def forward(self, chars): | |||||
input=self.embedding(chars) | |||||
output = input.transpose(1, 2) | |||||
# forward pass through conv layers | |||||
for i in range(len(self.conv_layers)): | |||||
output = self.conv_layers[i](output) | |||||
output = output.view(output.size(0), -1) | |||||
# forward pass through fc layers | |||||
for i in range(len(self.fc_layers)): | |||||
output = self.fc_layers[i](output) | |||||
return {C.OUTPUT: output} |
@@ -1 +1,106 @@ | |||||
# TODO | |||||
import torch | |||||
import torch.nn as nn | |||||
from fastNLP.modules.utils import get_embeddings | |||||
from fastNLP.core import Const as C | |||||
class DPCNN(nn.Module): | |||||
def __init__(self, init_embed, num_cls, n_filters=256, | |||||
kernel_size=3, n_layers=7, embed_dropout=0.1, cls_dropout=0.1): | |||||
super().__init__() | |||||
self.region_embed = RegionEmbedding( | |||||
init_embed, out_dim=n_filters, kernel_sizes=[1, 3, 5]) | |||||
embed_dim = self.region_embed.embedding_dim | |||||
self.conv_list = nn.ModuleList() | |||||
for i in range(n_layers): | |||||
self.conv_list.append(nn.Sequential( | |||||
nn.ReLU(), | |||||
nn.Conv1d(n_filters, n_filters, kernel_size, | |||||
padding=kernel_size//2), | |||||
nn.Conv1d(n_filters, n_filters, kernel_size, | |||||
padding=kernel_size//2), | |||||
)) | |||||
self.pool = nn.MaxPool1d(kernel_size=3, stride=2, padding=1) | |||||
self.embed_drop = nn.Dropout(embed_dropout) | |||||
self.classfier = nn.Sequential( | |||||
nn.Dropout(cls_dropout), | |||||
nn.Linear(n_filters, num_cls), | |||||
) | |||||
self.reset_parameters() | |||||
def reset_parameters(self): | |||||
for m in self.modules(): | |||||
if isinstance(m, (nn.Conv1d, nn.Conv2d, nn.Linear)): | |||||
nn.init.normal_(m.weight, mean=0, std=0.01) | |||||
if m.bias is not None: | |||||
nn.init.normal_(m.bias, mean=0, std=0.01) | |||||
def forward(self, words, seq_len=None): | |||||
words = words.long() | |||||
# get region embeddings | |||||
x = self.region_embed(words) | |||||
x = self.embed_drop(x) | |||||
# not pooling on first conv | |||||
x = self.conv_list[0](x) + x | |||||
for conv in self.conv_list[1:]: | |||||
x = self.pool(x) | |||||
x = conv(x) + x | |||||
# B, C, L => B, C | |||||
x, _ = torch.max(x, dim=2) | |||||
x = self.classfier(x) | |||||
return {C.OUTPUT: x} | |||||
def predict(self, words, seq_len=None): | |||||
x = self.forward(words, seq_len)[C.OUTPUT] | |||||
return {C.OUTPUT: torch.argmax(x, 1)} | |||||
class RegionEmbedding(nn.Module): | |||||
def __init__(self, init_embed, out_dim=300, kernel_sizes=None): | |||||
super().__init__() | |||||
if kernel_sizes is None: | |||||
kernel_sizes = [5, 9] | |||||
assert isinstance( | |||||
kernel_sizes, list), 'kernel_sizes should be List(int)' | |||||
self.embed = get_embeddings(init_embed) | |||||
try: | |||||
embed_dim = self.embed.embedding_dim | |||||
except Exception: | |||||
embed_dim = self.embed.embed_size | |||||
self.region_embeds = nn.ModuleList() | |||||
for ksz in kernel_sizes: | |||||
self.region_embeds.append(nn.Sequential( | |||||
nn.Conv1d(embed_dim, embed_dim, ksz, padding=ksz // 2), | |||||
)) | |||||
self.linears = nn.ModuleList([nn.Conv1d(embed_dim, out_dim, 1) | |||||
for _ in range(len(kernel_sizes))]) | |||||
self.embedding_dim = embed_dim | |||||
def forward(self, x): | |||||
x = self.embed(x) | |||||
x = x.transpose(1, 2) | |||||
# B, C, L | |||||
out = 0 | |||||
for conv, fc in zip(self.region_embeds, self.linears[1:]): | |||||
conv_i = conv(x) | |||||
out = out + fc(conv_i) | |||||
# B, C, L | |||||
return out | |||||
if __name__ == '__main__': | |||||
x = torch.randint(0, 10000, size=(5, 15), dtype=torch.long) | |||||
model = DPCNN((10000, 300), 20) | |||||
y = model(x) | |||||
print(y.size(), y.mean(1), y.std(1)) | |||||
@@ -0,0 +1,206 @@ | |||||
# 首先需要加入以下的路径到环境变量,因为当前只对内部测试开放,所以需要手动申明一下路径 | |||||
import os | |||||
os.environ['FASTNLP_BASE_URL'] = 'http://10.141.222.118:8888/file/download/' | |||||
os.environ['FASTNLP_CACHE_DIR'] = '/remote-home/hyan01/fastnlp_caches' | |||||
import sys | |||||
sys.path.append('../..') | |||||
from fastNLP.core.const import Const as C | |||||
import torch.nn as nn | |||||
from fastNLP.io.dataset_loader import SSTLoader | |||||
from data.yelpLoader import yelpLoader | |||||
from data.sstLoader import sst2Loader | |||||
from data.IMDBLoader import IMDBLoader | |||||
from model.char_cnn import CharacterLevelCNN | |||||
from fastNLP.core.vocabulary import Vocabulary | |||||
from fastNLP.models.cnn_text_classification import CNNText | |||||
from fastNLP.modules.encoder.embedding import CNNCharEmbedding,StaticEmbedding,StackEmbedding,LSTMCharEmbedding | |||||
from fastNLP import CrossEntropyLoss, AccuracyMetric | |||||
from fastNLP.core.trainer import Trainer | |||||
from torch.optim import SGD | |||||
from torch.autograd import Variable | |||||
import torch | |||||
from fastNLP import BucketSampler | |||||
##hyper | |||||
#todo 这里加入fastnlp的记录 | |||||
class Config(): | |||||
model_dir_or_name="en-base-uncased" | |||||
embedding_grad= False, | |||||
bert_embedding_larers= '4,-2,-1' | |||||
train_epoch= 50 | |||||
num_classes=2 | |||||
task= "IMDB" | |||||
#yelp_p | |||||
datapath = {"train": "/remote-home/ygwang/yelp_polarity/train.csv", | |||||
"test": "/remote-home/ygwang/yelp_polarity/test.csv"} | |||||
#IMDB | |||||
#datapath = {"train": "/remote-home/ygwang/IMDB_data/train.csv", | |||||
# "test": "/remote-home/ygwang/IMDB_data/test.csv"} | |||||
# sst | |||||
# datapath = {"train": "/remote-home/ygwang/workspace/GLUE/SST-2/train.tsv", | |||||
# "dev": "/remote-home/ygwang/workspace/GLUE/SST-2/dev.tsv"} | |||||
lr=0.01 | |||||
batch_size=128 | |||||
model_size="large" | |||||
number_of_characters=69 | |||||
extra_characters='' | |||||
max_length=1014 | |||||
char_cnn_config={ | |||||
"alphabet": { | |||||
"en": { | |||||
"lower": { | |||||
"alphabet": "abcdefghijklmnopqrstuvwxyz0123456789-,;.!?:'\"/\\|_@#$%^&*~`+-=<>()[]{}", | |||||
"number_of_characters": 69 | |||||
}, | |||||
"both": { | |||||
"alphabet": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-,;.!?:'\"/\\|_@#$%^&*~`+-=<>()[]{}", | |||||
"number_of_characters": 95 | |||||
} | |||||
} | |||||
}, | |||||
"model_parameters": { | |||||
"small": { | |||||
"conv": [ | |||||
#依次是channel,kennnel_size,maxpooling_size | |||||
[256,7,3], | |||||
[256,7,3], | |||||
[256,3,-1], | |||||
[256,3,-1], | |||||
[256,3,-1], | |||||
[256,3,3] | |||||
], | |||||
"fc": [1024,1024] | |||||
}, | |||||
"large":{ | |||||
"conv":[ | |||||
[1024, 7, 3], | |||||
[1024, 7, 3], | |||||
[1024, 3, -1], | |||||
[1024, 3, -1], | |||||
[1024, 3, -1], | |||||
[1024, 3, 3] | |||||
], | |||||
"fc": [2048,2048] | |||||
} | |||||
}, | |||||
"data": { | |||||
"text_column": "SentimentText", | |||||
"label_column": "Sentiment", | |||||
"max_length": 1014, | |||||
"num_of_classes": 2, | |||||
"encoding": None, | |||||
"chunksize": 50000, | |||||
"max_rows": 100000, | |||||
"preprocessing_steps": ["lower", "remove_hashtags", "remove_urls", "remove_user_mentions"] | |||||
}, | |||||
"training": { | |||||
"batch_size": 128, | |||||
"learning_rate": 0.01, | |||||
"epochs": 10, | |||||
"optimizer": "sgd" | |||||
} | |||||
} | |||||
ops=Config | |||||
##1.task相关信息:利用dataloader载入dataInfo | |||||
dataloader=sst2Loader() | |||||
dataloader=IMDBLoader() | |||||
#dataloader=yelpLoader(fine_grained=True) | |||||
datainfo=dataloader.process(ops.datapath,char_level_op=True) | |||||
char_vocab=ops.char_cnn_config["alphabet"]["en"]["lower"]["alphabet"] | |||||
ops.number_of_characters=len(char_vocab) | |||||
ops.embedding_dim=ops.number_of_characters | |||||
#chartoindex | |||||
def chartoindex(chars): | |||||
max_seq_len=ops.max_length | |||||
zero_index=len(char_vocab) | |||||
char_index_list=[] | |||||
for char in chars: | |||||
if char in char_vocab: | |||||
char_index_list.append(char_vocab.index(char)) | |||||
else: | |||||
#<unk>和<pad>均使用最后一个作为embbeding | |||||
char_index_list.append(zero_index) | |||||
if len(char_index_list) > max_seq_len: | |||||
char_index_list = char_index_list[:max_seq_len] | |||||
elif 0 < len(char_index_list) < max_seq_len: | |||||
char_index_list = char_index_list+[zero_index]*(max_seq_len-len(char_index_list)) | |||||
elif len(char_index_list) == 0: | |||||
char_index_list=[zero_index]*max_seq_len | |||||
return char_index_list | |||||
for dataset in datainfo.datasets.values(): | |||||
dataset.apply_field(chartoindex,field_name='chars',new_field_name='chars') | |||||
datainfo.datasets['train'].set_input('chars') | |||||
datainfo.datasets['test'].set_input('chars') | |||||
datainfo.datasets['train'].set_target('target') | |||||
datainfo.datasets['test'].set_target('target') | |||||
##2. 定义/组装模型,这里可以随意,就如果是fastNLP封装好的,类似CNNText就直接用初始化调用就好了,这里只是给出一个伪框架表示占位,在这里建立符合fastNLP输入输出规范的model | |||||
class ModelFactory(nn.Module): | |||||
""" | |||||
用于拼装embedding,encoder,decoder 以及设计forward过程 | |||||
:param embedding: embbeding model | |||||
:param encoder: encoder model | |||||
:param decoder: decoder model | |||||
""" | |||||
def __int__(self,embedding,encoder,decoder,**kwargs): | |||||
super(ModelFactory,self).__init__() | |||||
self.embedding=embedding | |||||
self.encoder=encoder | |||||
self.decoder=decoder | |||||
def forward(self,x): | |||||
return {C.OUTPUT:None} | |||||
## 2.或直接复用fastNLP的模型 | |||||
#vocab=datainfo.vocabs['words'] | |||||
vocab_label=datainfo.vocabs['target'] | |||||
''' | |||||
# emded_char=CNNCharEmbedding(vocab) | |||||
# embed_word = StaticEmbedding(vocab, model_dir_or_name='en-glove-6b-50', requires_grad=True) | |||||
# embedding=StackEmbedding([emded_char, embed_word]) | |||||
# cnn_char_embed = CNNCharEmbedding(vocab) | |||||
# lstm_char_embed = LSTMCharEmbedding(vocab) | |||||
# embedding = StackEmbedding([cnn_char_embed, lstm_char_embed]) | |||||
''' | |||||
#one-hot embedding | |||||
embedding_weight= Variable(torch.zeros(len(char_vocab)+1, len(char_vocab))) | |||||
for i in range(len(char_vocab)): | |||||
embedding_weight[i][i]=1 | |||||
embedding=nn.Embedding(num_embeddings=len(char_vocab)+1,embedding_dim=len(char_vocab),padding_idx=len(char_vocab),_weight=embedding_weight) | |||||
for para in embedding.parameters(): | |||||
para.requires_grad=False | |||||
#CNNText太过于简单 | |||||
#model=CNNText(init_embed=embedding, num_classes=ops.num_classes) | |||||
model=CharacterLevelCNN(ops,embedding) | |||||
## 3. 声明loss,metric,optimizer | |||||
loss=CrossEntropyLoss | |||||
metric=AccuracyMetric | |||||
optimizer= SGD([param for param in model.parameters() if param.requires_grad==True], lr=ops.lr) | |||||
## 4.定义train方法 | |||||
def train(model,datainfo,loss,metrics,optimizer,num_epochs=100): | |||||
trainer = Trainer(datainfo.datasets['train'], model, optimizer=optimizer, loss=loss(target='target'), | |||||
metrics=[metrics(target='target')], dev_data=datainfo.datasets['test'], device=0, check_code_level=-1, | |||||
n_epochs=num_epochs) | |||||
print(trainer.train()) | |||||
if __name__=="__main__": | |||||
#print(vocab_label) | |||||
#print(datainfo.datasets["train"]) | |||||
train(model,datainfo,loss,metric,optimizer,num_epochs=ops.train_epoch) | |||||
@@ -0,0 +1,125 @@ | |||||
# 首先需要加入以下的路径到环境变量,因为当前只对内部测试开放,所以需要手动申明一下路径 | |||||
import torch.cuda | |||||
from fastNLP.core.utils import cache_results | |||||
from torch.optim import SGD | |||||
from torch.optim.lr_scheduler import CosineAnnealingLR, LambdaLR | |||||
from fastNLP.core.trainer import Trainer | |||||
from fastNLP import CrossEntropyLoss, AccuracyMetric | |||||
from fastNLP.modules.encoder.embedding import StaticEmbedding, CNNCharEmbedding, StackEmbedding | |||||
from reproduction.text_classification.model.dpcnn import DPCNN | |||||
from data.yelpLoader import yelpLoader | |||||
import torch.nn as nn | |||||
from fastNLP.core import LRScheduler | |||||
from fastNLP.core.const import Const as C | |||||
from fastNLP.core.vocabulary import VocabularyOption | |||||
from utils.util_init import set_rng_seeds | |||||
import os | |||||
os.environ['FASTNLP_BASE_URL'] = 'http://10.141.222.118:8888/file/download/' | |||||
os.environ['FASTNLP_CACHE_DIR'] = '/remote-home/hyan01/fastnlp_caches' | |||||
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" | |||||
# hyper | |||||
class Config(): | |||||
seed = 12345 | |||||
model_dir_or_name = "dpcnn-yelp-p" | |||||
embedding_grad = True | |||||
train_epoch = 30 | |||||
batch_size = 100 | |||||
num_classes = 2 | |||||
task = "yelp_p" | |||||
#datadir = '/remote-home/yfshao/workdir/datasets/SST' | |||||
datadir = '/remote-home/yfshao/workdir/datasets/yelp_polarity' | |||||
#datafile = {"train": "train.txt", "dev": "dev.txt", "test": "test.txt"} | |||||
datafile = {"train": "train.csv", "test": "test.csv"} | |||||
lr = 1e-3 | |||||
src_vocab_op = VocabularyOption() | |||||
embed_dropout = 0.3 | |||||
cls_dropout = 0.1 | |||||
weight_decay = 1e-4 | |||||
def __init__(self): | |||||
self.datapath = {k: os.path.join(self.datadir, v) | |||||
for k, v in self.datafile.items()} | |||||
ops = Config() | |||||
set_rng_seeds(ops.seed) | |||||
print('RNG SEED: {}'.format(ops.seed)) | |||||
# 1.task相关信息:利用dataloader载入dataInfo | |||||
#datainfo=SSTLoader(fine_grained=True).process(paths=ops.datapath, train_ds=['train']) | |||||
@cache_results(ops.model_dir_or_name+'-data-cache') | |||||
def load_data(): | |||||
datainfo = yelpLoader(fine_grained=True, lower=True).process( | |||||
paths=ops.datapath, train_ds=['train'], src_vocab_op=ops.src_vocab_op) | |||||
for ds in datainfo.datasets.values(): | |||||
ds.apply_field(len, C.INPUT, C.INPUT_LEN) | |||||
ds.set_input(C.INPUT, C.INPUT_LEN) | |||||
ds.set_target(C.TARGET) | |||||
return datainfo | |||||
datainfo = load_data() | |||||
# 2.或直接复用fastNLP的模型 | |||||
vocab = datainfo.vocabs['words'] | |||||
# embedding = StackEmbedding([StaticEmbedding(vocab), CNNCharEmbedding(vocab, 100)]) | |||||
#embedding = StaticEmbedding(vocab) | |||||
embedding = StaticEmbedding( | |||||
vocab, model_dir_or_name='en-word2vec-300', requires_grad=ops.embedding_grad, | |||||
normalize=False | |||||
) | |||||
print(len(datainfo.datasets['train'])) | |||||
print(len(datainfo.datasets['test'])) | |||||
print(datainfo.datasets['train'][0]) | |||||
print(len(vocab)) | |||||
print(len(datainfo.vocabs['target'])) | |||||
model = DPCNN(init_embed=embedding, num_cls=ops.num_classes, | |||||
embed_dropout=ops.embed_dropout, cls_dropout=ops.cls_dropout) | |||||
print(model) | |||||
# 3. 声明loss,metric,optimizer | |||||
loss = CrossEntropyLoss(pred=C.OUTPUT, target=C.TARGET) | |||||
metric = AccuracyMetric(pred=C.OUTPUT, target=C.TARGET) | |||||
optimizer = SGD([param for param in model.parameters() if param.requires_grad == True], | |||||
lr=ops.lr, momentum=0.9, weight_decay=ops.weight_decay) | |||||
callbacks = [] | |||||
callbacks.append(LRScheduler(CosineAnnealingLR(optimizer, 5))) | |||||
# callbacks.append | |||||
# LRScheduler(LambdaLR(optimizer, lambda epoch: ops.lr if epoch < | |||||
# ops.train_epoch * 0.8 else ops.lr * 0.1)) | |||||
# ) | |||||
# callbacks.append( | |||||
# FitlogCallback(data=datainfo.datasets, verbose=1) | |||||
# ) | |||||
device = 'cuda:0' if torch.cuda.is_available() else 'cpu' | |||||
print(device) | |||||
# 4.定义train方法 | |||||
trainer = Trainer(datainfo.datasets['train'], model, optimizer=optimizer, loss=loss, | |||||
metrics=[metric], | |||||
dev_data=datainfo.datasets['test'], device=device, | |||||
check_code_level=-1, batch_size=ops.batch_size, callbacks=callbacks, | |||||
n_epochs=ops.train_epoch, num_workers=4) | |||||
if __name__ == "__main__": | |||||
print(trainer.train()) | |||||
@@ -0,0 +1,11 @@ | |||||
import numpy | |||||
import torch | |||||
import random | |||||
def set_rng_seeds(seed): | |||||
random.seed(seed) | |||||
numpy.random.seed(seed) | |||||
torch.random.manual_seed(seed) | |||||
torch.cuda.manual_seed_all(seed) | |||||
# print('RNG_SEED {}'.format(seed)) |