-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathgloss_data.py
266 lines (181 loc) · 10.9 KB
/
gloss_data.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
from torch.utils.data import Dataset
import json
import pandas as pd
import math
import torch
import random
import torch, os, gzip, pickle, json, numpy as np
from collections import OrderedDict
import re
# from charsplit import Splitter
import fasttext
from sklearn.metrics.pairwise import cosine_similarity
import itertools
import string
replacments = { "ueberschwemmung":"überschwemmung","sued":"süd","tagsueber":"tagsüber" , "fuenf":"fünf",
"ruegen" :"rügen" , "fruehling": "frühling" , "bluete": "blüte" , "tagsueber":"tagsüber",
"begruessen":"begrüßen", "fuehlen":"fühlen", "duenn":"dünn","kuehl":"kühl", "fuenf":"fünf","kueste":"küste",
"tschuess":"tschüss","trueb":"trüb","duesseldorf":"düsseldorf","ueberwiegen":"überwiegen","muessen":"müssen",
"gemuetlich":"gemütlich", "ueber":"über", "wuenschen":"wünschen", "ueberall":"überall", "zurueck":"zurück",
"thueringe":"thüringe", "luecke":"lücken", "schwuel":"schwül","tuerkei":"türkei", "wuerttemberg":"württemberg","voruebergehend":"vorübergehend","fuer":"für",
"thueringen":"thüringen", "schuetzen":"schützen","spueren":"spüren", "muenchen":"münchen","puenktlich":"pünktlich", "gluecken":"glücken",
"wuerz":"würz", "merkwuerdig":"merkwürdig" , "rücken":"rücken", "muenster":"münster", "aufbluehen":"aufblühen",
"gruen":"grün", "glück":"glück"}
class GlossData(Dataset):
def __init__(self, path:str ,subset , tokenizer, gloss_embeddings= None , map_ids = None, type="dec", use_embd = False, back_trans = False, model = None):
with gzip.open("./data/phoenix-2014t/phoenix-2014t_cleaned."+subset,'rb') as f:
data = pickle.load(f)
with gzip.open("./data/phoenix-2014t/phoenix-2014t_cleaned."+'test','rb') as f:
data_test = pickle.load(f)
with gzip.open("./data/phoenix-2014t/phoenix-2014t_cleaned."+'dev','rb') as f:
data_val = pickle.load(f)
splitter = Splitter()
self.list_of_tokens = []
self.gloss = []
self.translation = []
self.use_embd = use_embd
self.model = model
annotations = pd.read_csv(path+"/PHOENIX-2014-T."+ subset+ ".corpus.csv", sep="|")
self.gloss_f = annotations.orth.str.lower().tolist()
# self.translation = annotations.translation.tolist()
for i in range(len(data)):
self.gloss.append(data[i]['gloss'].lower())
self.translation.append(data[i]['text'])
if back_trans and subset =='train' :
back_trans_data = pd.read_csv("./res_nllb_text2gls.csv")
for i in range(len(back_trans_data)):
if isinstance(back_trans_data.iloc[i]['pred_gloss'] , str) :
self.gloss.append(back_trans_data.iloc[i]['pred_gloss'][:-1])
self.translation.append(back_trans_data.iloc[i]['translation'][:-1] + ' .')
back_trans_data_txt = pd.read_csv("./res_nllb_text2text.csv")
for i in range(len(back_trans_data_txt)):
self.gloss.append(back_trans_data_txt.iloc[i]['gloss'][:-1])
self.translation.append(back_trans_data_txt.iloc[i]['back_trans'])
all_words = self.translation[-1].split()
for sub_words in all_words:
if splitter.split_compound(sub_words)[0][0]> 0.90 :
self.translation[-1] = self.translation[-1].replace(sub_words ,splitter.split_compound(sub_words)[0][1].lower() + ' ' + splitter.split_compound(sub_words)[0][2].lower() )
self.type = type
self.gloss_embeddings = gloss_embeddings
self.map_ids = map_ids
self.tokenizer = tokenizer
self.subset =subset
self.X = []
print(type)
if "enc_dec" in type:
self.Y = []
self.ids_sim = set()
for idx, i in enumerate(self.gloss):
gl = i
tr = self.translation[idx]
self.X.append(gl)
self.Y.append(tr)
self.list_of_tokens.extend(self.tokenizer.encode(tr))
self.list_of_tokens = set(self.list_of_tokens)
self.list_of_tokens = np.array(list(self.list_of_tokens))
swapped_dict = dict(map(lambda item: (item[1], item[0]), self.tokenizer.get_vocab().items()))
numbers = 0
self.list_of_texts= []
# for tokens in self.list_of_tokens:
self.list_of_texts = self.tokenizer.batch_decode(self.list_of_tokens.reshape(-1,1))
model = fasttext.load_model('cc.de.300.bin')
embeddings = [model.get_sentence_vector(tokens) for tokens in self.list_of_texts]
self.similarity_matrix = cosine_similarity(embeddings)
else:
for idx, i in enumerate(self.gloss):
self.X.append("<startofstring> "+ i +" <trans>: "+self.translation[idx]+" <endofstring>")
print(self.X[0])
#torch.where(torch.all(tokenizer(self.X,max_length=110, truncation=True, padding="max_length" , return_tensors="pt")['attention_mask'], axis=1))
self.X_encoded = tokenizer(self.X, padding=True, return_tensors="pt")
self.input_ids = self.X_encoded['input_ids']
self.attention_mask = self.X_encoded['attention_mask']
def build_tokenizer(self,translation, gloss):
all_embds = {}
import string
train_vocabs = set()
for i in translation+gloss:
if i != i.translate(str.maketrans('', '', string.punctuation.replace("-", "").replace(".", ""))):
print(i)
train_vocabs.update(i.lower().split())
train_vocabs = list(train_vocabs)
initial_len = 0
# import pdb; pdb.set_trace()
with torch.no_grad():
full_embedding_weight = self.model.model.shared.weight
new_embeddings = torch.zeros((len(train_vocabs),full_embedding_weight.shape[1]), device=full_embedding_weight.device)
for i in train_vocabs:
gls_ids = self.tokenizer(i)['input_ids'][1:-1] # remove</s> <lang>
emb = []
for j in gls_ids:
emb.append(full_embedding_weight[j,:])
emb = torch.mean(torch.stack(emb, dim=0), dim=0)
new_embeddings[initial_len] = emb
initial_len += 1
self.tokenizer.add_tokens(train_vocabs)
self.model.model.resize_token_embeddings(len(self.tokenizer))
self.model.model.shared.weight.data[-len(train_vocabs):].copy_(new_embeddings)
def get_model(self):
return self.model,self.tokenizer
def __len__(self):
return len(self.gloss)
def return_sim(self):
return self.list_of_tokens, self.similarity_matrix, self.list_of_texts
def prepare_gloss_inputs(self, gloss):
gloss = gloss.split()
gloss = ["deu_Latn"] + gloss + ["</s>"]
input_emb = torch.zeros(len(gloss),2048)
for idx, gls in enumerate(gloss):
try:
input_emb[idx] = self.gloss_embeddings[gls]
except:
print("unrec gloss :", gls)
input_emb[idx] = self.gloss_embeddings['<unk>']
return input_emb*45.0
def prepare_txt_inputs(self, input_ids):
input_ids = input_ids.clone()
for idx, txt_id in enumerate(input_ids):
input_ids[idx] = self.map_ids[txt_id.item()]
input_ids = torch.cat((torch.tensor([6]),input_ids[:-1],torch.tensor([6])))
return input_ids.unsqueeze(0)
def __getitem__(self, idx):
self.Y_encoded = self.tokenizer(self.Y[idx], padding=True, return_tensors="pt")
self.target_ids = self.Y_encoded['input_ids']
self.target_attention_mask = self.Y_encoded['attention_mask']
if self.type== "enc_dec_nllb_bio":
self.X_encoded = self.tokenizer(self.X[idx], padding=True, return_tensors="pt")
self.input_ids = self.X_encoded['input_ids']
self.input_attention_mask = self.X_encoded['attention_mask']
self.input_ids[0][0] = 256204
self.target_ids[0][0] = 256042
if self.subset == 'train':
if not random.randrange(3):
return self.input_ids, self.target_ids , len(self.input_ids[0]) , len(self.target_ids[0])
return self.target_ids, self.input_ids , len(self.target_ids[0]) , len(self.input_ids[0])
if self.type== "enc_dec_nllb":
if self.use_embd:
input_enc = self.prepare_gloss_inputs(self.X[idx])
return self.target_ids, input_enc , len(self.target_ids[0]) , len(input_enc)
self.X_encoded = self.tokenizer(self.X[idx], padding=True, return_tensors="pt")
self.input_ids = self.X_encoded['input_ids']
self.input_attention_mask = self.X_encoded['attention_mask']
return self.target_ids, self.input_ids , len(self.target_ids[0]) , len(self.input_ids[0])
if self.type== "enc_dec_nllb_text_gls":
if self.use_embd:
input_enc = self.prepare_gloss_inputs(self.X[idx])
return self.target_ids, input_enc , len(self.target_ids[0]) , len(input_enc)
self.X_encoded = self.tokenizer(self.X[idx], padding=True, return_tensors="pt")
self.input_ids = self.X_encoded['input_ids']
self.input_attention_mask = self.X_encoded['attention_mask']
return self.input_ids, self.target_ids , len(self.input_ids[0]) , len(self.target_ids[0])
if self.type== "enc_dec_all":
self.X_encoded = self.tokenizer(self.X[idx], padding=True, return_tensors="pt")
self.input_ids = self.X_encoded['input_ids']
self.input_attention_mask = self.X_encoded['attention_mask']
self.target_ids = torch.cat((torch.tensor([250003]),self.target_ids[0])).unsqueeze(0)
return self.target_ids, self.input_ids , len(self.target_ids[0]) , len(self.input_ids[0])
if self.type== "enc_dec":
input_dec = self.prepare_txt_inputs(self.target_ids[0])
input_enc = self.prepare_gloss_inputs(self.gloss[idx])
return input_dec, input_enc , len(input_dec[0]) , len(input_enc)
else:
return (self.input_ids[idx], self.attention_mask[idx])