-
Notifications
You must be signed in to change notification settings - Fork 1
/
model2.py
170 lines (149 loc) · 9.52 KB
/
model2.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
import numpy as np
import torch
import torch.nn as nn
from preprocessing import tokenizer
from transformers import XLMTokenizer, XLMWithLMHeadModel, XLMModel
from utilities import model_utils
from Globals import *
batch_size = 32
dic = tokenizer.decoder
class xlmb2b(nn.Module, model_utils):
def __init__(self, dic = dic, d_model=1024, trfrmr_nlayers=4, pll_dat=True) :
super().__init__()
self.xlm = XLMModel.from_pretrained('xlm-mlm-ende-1024')
self.d_model = d_model
decoder_layer = torch.nn.TransformerDecoderLayer(self.d_model, nhead=8)
self.trnsfrmr_dcodr = torch.nn.TransformerDecoder(decoder_layer, num_layers=trfrmr_nlayers)
self.pll_data = pll_dat
self.max_tr_seq_len = 120
self.end_tok = 1 #Token id of end token
self.dic_tensor = torch.tensor([v for k,v in tokenizer.encoder.items()],device=device) #tensor with i_th token's id at i_th position
self.vocab_size = self.dic_tensor.shape[0]
self.final_linear = nn.Linear(self.d_model, self.vocab_size)
self.it_no = None
self.beam_size = 1
self.k = 1
self.m = 1
self.begin_prgrsiv_xlm_to_plt = True
self.begin_prgrsiv_real_to_pred = False
self.p = None #_______
self.beam_size = 1 #_______
def choose(self) :
'''Chooses final output beam for each sample using beam_size,
final_out,prev_probs'''
x = self.prev_probs.max(1, keepdim=True)[1] #batch_sizeX1Xbeam_size
s = torch.gather(self.prev_probs, dim=1, index=x) #batch_sizeX1Xbeam_size
y = s.max(2)[1] #batch_sizeX1
i = torch.tensor([i for i in range(y.shape[0])],device=device)
final_out = torch.stack(self.final_out).transpose(0,1)
final_out = final_out.reshape(self.beam_size,-1,final_out.shape[1])
return self.prev_probs[i,:,y.reshape(-1)], self.sr_tokens, final_out[y.reshape(-1),i.reshape(-1),:]
def update(self, update_m) :
if update_m :
self.m = self.m + 0.01 #______
else :
self.k = self.k + 0.01 #______
def get_prgrsiv_embdngs(self, dic, xlm_encoding) :
if self.begin_prgrsiv_xlm_to_plt :
plt_embdng = self.plt_embed(dic['input_ids'],dic['langs'], dic['position_ids'])
self.update(True)
return self.m*xlm_encoding+(1-self.m)*plt_embdng
return xlm_encoding
def filter_logits(self, logits) :
sorted_logits, sorted_indices = logits.sort(dim=-1, descending=True)
cum_probs = F.softmax(sorted_logits,dim=-1).cumsum(dim=-1)
sorted_indices_to_remove = cum_probs>=self.p
sorted_indices_to_remove[...,1:] = sorted_indices_to_remove[...,:-1]
sorted_indices_to_remove[...,0] = False #To ensure at least one is there in top-p
indices_to_remove = sorted_indices[sorted_indices_to_remove]
logits[indices_to_remove] = -float('Inf')
return logits
def k_nucleus_sample(self, logits, lang_ids, position_ids) :
filtered_logits = self.filter_logits(logits)
probs = F.softmax(filtered_logits, dim=-1)
token_ids = torch.multinomial(probs,self.beam_size)
probs_of_selected_tokens = F.softmax(torch.gather(probs,-1,token_ids),dim=-1) #2 softmax together. Can multiply by 1/k_sample in line 74, instead.
k_plt_embeds = self.plt_embed(token_ids, lang_ids, position_ids)
probabilistic_embeds = self.cycle_dims(self.cycle_dims(k_plt_embeds)*probs_of_selected_tokens,clockwise=False)
plt_embeds = torch.sum(probabilistic_embeds,dim=-2)
return plt_embeds
def get_prgrsiv_tr_embd(self, probs, prgrsiv_sr_embd, tr_embd, tr_dic) :
'''Converts tr_embd->k*tr_embd+(1-k)*plt_embdng_of_pred
using probs'''
if not self.begin_prgrsiv_real_to_pred :
return probs, prgrsiv_sr_embd, tr_embd
plt_embdng = self.k_nucleus_sample(probs, tr_dic['langs'], tr_dic['position_ids'])
self.update(False)
return probs, prgrsiv_sr_embd, self.k*(tr_embd)+(1-self.k)*plt_embdng
def forward(self, dat, already_embed = False) :
if self.pll_data :
inp = dat['X']
out = dat['Y']
if not already_embed :
sr_embd = self.xlm(**self.change_attn_for_xlm(inp))[0]
tr_embd = self.xlm(**self.change_attn_for_xlm(out))[0]
prgrsiv_sr_embd = self.get_prgrsiv_embdngs(inp, sr_embd)
prgrsiv_tr_embd = self.get_prgrsiv_embdngs(out, tr_embd)
else :
sr_embd = inp['input_ids']
prgrsiv_tr_embd = out['input_ids']
tr_len = int(out['lengths'].max())
tgt_mask = self.get_tgt_mask(tr_len)
trfrmr_out = self.trnsfrmr_dcodr(tgt=prgrsiv_tr_embd.transpose(0,1),
memory=sr_embd.transpose(0,1), tgt_mask=tgt_mask,
tgt_key_padding_mask=~(out['attention_mask'].bool()),
memory_key_padding_mask=~(inp['attention_mask'].bool())).transpose(0,1)
probs = self.apply_final_layer(trfrmr_out, out['attention_mask'].float())
out['attention_mask'] = out['attention_mask'].float()
if not already_embed :
return self.get_prgrsiv_tr_embd(probs, prgrsiv_sr_embd, tr_embd, out)
else :
return probs, sr_embd, prgrsiv_tr_embd
else :
inp = dat['X']
self.sr_tokens = inp['input_ids']
self.sr_embd = self.xlm(**self.change_attn_for_xlm(inp))[0].repeat_interleave(self.beam_size,0)
self.bs = inp['input_ids'].shape[0]*self.beam_size
self.max_tr_seq_len = self.beam_size*self.max_tr_seq_len
self.tgt_key_pad_mask = torch.zeros((self.bs, self.max_tr_seq_len),device=device)
self.mem_key_pad_mask = inp['attention_mask'].repeat_interleave(self.beam_size,0)
self.tgt_mask = self.get_tgt_mask(self.max_tr_seq_len,0)
self.tr_embd = torch.zeros((self.bs, self.max_tr_seq_len, self.d_model),dtype=torch.float64, device=device)
self.not_done_samples = torch.ones(self.bs, dtype=torch.bool, device=device)
self.it_no = 0 #if nth word of target sequence is being predicted,
self.final_out = [] #then iteration number(it_no) == n-1
self.lengs = torch.zeros((self.bs),device=device)
self.actual_bs = int(self.bs/self.beam_size)
self.prev_probs = torch.zeros((self.actual_bs,self.max_tr_seq_len+1,self.beam_size), device=device, dtype=torch.float64)
self.tgt_key_pad_mask[:,self.it_no] = torch.ones((self.bs), device=device)
self.just_now_completed_samples_mask = torch.zeros((self.bs), dtype=torch.bool, device=device)
self.seq_len_sr = inp['lengths'].max()
while True :
trfrmr_out = self.trnsfrmr_dcodr(tgt=self.tr_embd.transpose(0,1),
memory=self.sr_embd.transpose(0,1), tgt_mask=self.tgt_mask,
tgt_key_padding_mask=~(self.tgt_key_pad_mask.bool()),
memory_key_padding_mask=~(self.mem_key_pad_mask.bool())).transpose(0,1)
val, masky = self.apply_final_layer( trfrmr_out, self.tgt_key_pad_mask.float() )
trfrmr_out = torch.zeros((self.bs,self.vocab_size), device=device, dtype=torch.float64)
trfrmr_out[masky[:,self.it_no+1].bool()] = val
self.tgt_key_pad_mask = self.tgt_key_pad_mask.long()
dic_indices = self.reform(trfrmr_out)
dic_indices[~self.not_done_samples] = tokenizer.pad_token_id
output_at_it_no = torch.zeros((self.bs,1), dtype=torch.long, device=device)
output_at_it_no[self.not_done_samples] = self.dic_tensor[dic_indices].reshape(-1,1)[self.not_done_samples]
self.final_out.append(output_at_it_no)
self.tr_embd[self.not_done_samples,self.it_no+1,:] = self.embed_for_decoder(output_at_it_no[self.not_done_samples], inp['langs'][:,0]).squeeze(dim=1) #Adding next words embeddings to context for decoder
ind = output_at_it_no[self.not_done_samples]!=self.end_tok
ind=ind.reshape(-1)
new_done_samples_len = (self.not_done_samples==True).sum()-(ind==True).sum()
if new_done_samples_len!=0 :
self.calc_just_now_completed_samples_mask(ind)
self.lengs[self.just_now_completed_samples_mask] = self.it_no+1
self.mem_key_pad_mask[self.just_now_completed_samples_mask] = 0 #torch.zeros((new_done_samples_len, self.seq_len_sr)).long()
self.tgt_key_pad_mask[self.just_now_completed_samples_mask] = 0 #torch.zeros((new_done_samples_len, self.max_tr_seq_len)).long()
self.tgt_key_pad_mask[self.mask_fr_mask()] = 1
if self.not_done_samples.sum()==0 or self.it_no==self.max_tr_seq_len-2:
self.it_no = None
return self.choose()
self.it_no+=1
self.tgt_mask = self.get_tgt_mask(self.max_tr_seq_len, self.it_no)