File size: 22,783 Bytes
7d68ade |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 |
from numpy import asarray as np_asarray,int64 as np_int64
from json import load as json_load
import torch
import torch.nn as nn
from torch import device as Device
class Encoder(nn.Module):
def __init__(self, input_dim, embed_dim, hidden_dim ,
rnn_type = 'gru', layers = 1,
bidirectional =False,
dropout = 0, device = "cpu"):
super(Encoder, self).__init__()
self.input_dim = input_dim #src_vocab_sz
self.enc_embed_dim = embed_dim
self.enc_hidden_dim = hidden_dim
self.enc_rnn_type = rnn_type
self.enc_layers = layers
self.enc_directions = 2 if bidirectional else 1
self.device = device
self.embedding = nn.Embedding(self.input_dim, self.enc_embed_dim)
if self.enc_rnn_type == "gru":
self.enc_rnn = nn.GRU(input_size= self.enc_embed_dim,
hidden_size= self.enc_hidden_dim,
num_layers= self.enc_layers,
bidirectional= bidirectional)
elif self.enc_rnn_type == "lstm":
self.enc_rnn = nn.LSTM(input_size= self.enc_embed_dim,
hidden_size= self.enc_hidden_dim,
num_layers= self.enc_layers,
bidirectional= bidirectional)
else:
raise Exception("XlitError: unknown RNN type mentioned")
def forward(self, x, x_sz, hidden = None):
"""
x_sz: (batch_size, 1) - Unpadded sequence lengths used for pack_pad
"""
batch_sz = x.shape[0]
# x: batch_size, max_length, enc_embed_dim
x = self.embedding(x)
## pack the padded data
# x: max_length, batch_size, enc_embed_dim -> for pack_pad
x = x.permute(1,0,2)
x = nn.utils.rnn.pack_padded_sequence(x, x_sz, enforce_sorted=False) # unpad
# output: packed_size, batch_size, enc_embed_dim
# hidden: n_layer**num_directions, batch_size, hidden_dim | if LSTM (h_n, c_n)
output, hidden = self.enc_rnn(x) # gru returns hidden state of all timesteps as well as hidden state at last timestep
## pad the sequence to the max length in the batch
# output: max_length, batch_size, enc_emb_dim*directions)
output, _ = nn.utils.rnn.pad_packed_sequence(output)
# output: batch_size, max_length, hidden_dim
output = output.permute(1,0,2)
return output, hidden
def get_word_embedding(self, x):
"""
"""
x_sz = torch.tensor([len(x)])
x_ = torch.tensor(x).unsqueeze(0).to(dtype=torch.long)
# x: 1, max_length, enc_embed_dim
x = self.embedding(x_)
## pack the padded data
# x: max_length, 1, enc_embed_dim -> for pack_pad
x = x.permute(1,0,2)
x = nn.utils.rnn.pack_padded_sequence(x, x_sz, enforce_sorted=False) # unpad
# output: packed_size, 1, enc_embed_dim
# hidden: n_layer**num_directions, 1, hidden_dim | if LSTM (h_n, c_n)
output, hidden = self.enc_rnn(x) # gru returns hidden state of all timesteps as well as hidden state at last timestep
out_embed = hidden[0].squeeze()
return out_embed
class Decoder(nn.Module):
def __init__(self, output_dim, embed_dim, hidden_dim,
rnn_type = 'gru', layers = 1,
use_attention = True,
enc_outstate_dim = None, # enc_directions * enc_hidden_dim
dropout = 0, device = "cpu"):
super(Decoder, self).__init__()
self.output_dim = output_dim #tgt_vocab_sz
self.dec_hidden_dim = hidden_dim
self.dec_embed_dim = embed_dim
self.dec_rnn_type = rnn_type
self.dec_layers = layers
self.use_attention = use_attention
self.device = device
if self.use_attention:
self.enc_outstate_dim = enc_outstate_dim if enc_outstate_dim else hidden_dim
else:
self.enc_outstate_dim = 0
self.embedding = nn.Embedding(self.output_dim, self.dec_embed_dim)
if self.dec_rnn_type == 'gru':
self.dec_rnn = nn.GRU(input_size= self.dec_embed_dim + self.enc_outstate_dim, # to concat attention_output
hidden_size= self.dec_hidden_dim, # previous Hidden
num_layers= self.dec_layers,
batch_first = True )
elif self.dec_rnn_type == "lstm":
self.dec_rnn = nn.LSTM(input_size= self.dec_embed_dim + self.enc_outstate_dim, # to concat attention_output
hidden_size= self.dec_hidden_dim, # previous Hidden
num_layers= self.dec_layers,
batch_first = True )
else:
raise Exception("XlitError: unknown RNN type mentioned")
self.fc = nn.Sequential(
nn.Linear(self.dec_hidden_dim, self.dec_embed_dim), nn.LeakyReLU(),
# nn.Linear(self.dec_embed_dim, self.dec_embed_dim), nn.LeakyReLU(), # removing to reduce size
nn.Linear(self.dec_embed_dim, self.output_dim),
)
##----- Attention ----------
if self.use_attention:
self.W1 = nn.Linear( self.enc_outstate_dim, self.dec_hidden_dim)
self.W2 = nn.Linear( self.dec_hidden_dim, self.dec_hidden_dim)
self.V = nn.Linear( self.dec_hidden_dim, 1)
def attention(self, x, hidden, enc_output):
'''
x: (batch_size, 1, dec_embed_dim) -> after Embedding
enc_output: batch_size, max_length, enc_hidden_dim *num_directions
hidden: n_layers, batch_size, hidden_size | if LSTM (h_n, c_n)
'''
## perform addition to calculate the score
# hidden_with_time_axis: batch_size, 1, hidden_dim
## hidden_with_time_axis = hidden.permute(1, 0, 2) ## replaced with below 2lines
hidden_with_time_axis = torch.sum(hidden, axis=0) if self.dec_rnn_type != "lstm" \
else torch.sum(hidden[0], axis=0) # h_n
hidden_with_time_axis = hidden_with_time_axis.unsqueeze(1)
# score: batch_size, max_length, hidden_dim
score = torch.tanh(self.W1(enc_output) + self.W2(hidden_with_time_axis))
# attention_weights: batch_size, max_length, 1
# we get 1 at the last axis because we are applying score to self.V
attention_weights = torch.softmax(self.V(score), dim=1)
# context_vector shape after sum == (batch_size, hidden_dim)
context_vector = attention_weights * enc_output
context_vector = torch.sum(context_vector, dim=1)
# context_vector: batch_size, 1, hidden_dim
context_vector = context_vector.unsqueeze(1)
# attend_out (batch_size, 1, dec_embed_dim + hidden_size)
attend_out = torch.cat((context_vector, x), -1)
return attend_out, attention_weights
def forward(self, x, hidden, enc_output):
'''
x: (batch_size, 1)
enc_output: batch_size, max_length, dec_embed_dim
hidden: n_layer, batch_size, hidden_size | lstm: (h_n, c_n)
'''
if (hidden is None) and (self.use_attention is False):
raise Exception( "XlitError: No use of a decoder with No attention and No Hidden")
batch_sz = x.shape[0]
if hidden is None:
# hidden: n_layers, batch_size, hidden_dim
hid_for_att = torch.zeros((self.dec_layers, batch_sz,
self.dec_hidden_dim )).to(self.device)
elif self.dec_rnn_type == 'lstm':
hid_for_att = hidden[1] # c_n
# x (batch_size, 1, dec_embed_dim) -> after embedding
x = self.embedding(x)
if self.use_attention:
# x (batch_size, 1, dec_embed_dim + hidden_size) -> after attention
# aw: (batch_size, max_length, 1)
x, aw = self.attention( x, hidden, enc_output)
else:
x, aw = x, 0
# passing the concatenated vector to the GRU
# output: (batch_size, n_layers, hidden_size)
# hidden: n_layers, batch_size, hidden_size | if LSTM (h_n, c_n)
output, hidden = self.dec_rnn(x, hidden) if hidden is not None else self.dec_rnn(x)
# output :shp: (batch_size * 1, hidden_size)
output = output.view(-1, output.size(2))
# output :shp: (batch_size * 1, output_dim)
output = self.fc(output)
return output, hidden, aw
class Seq2Seq(nn.Module):
"""
Class dependency: Encoder, Decoder
"""
def __init__(self, encoder, decoder, pass_enc2dec_hid=False, dropout = 0, device = "cpu"):
super(Seq2Seq, self).__init__()
self.encoder = encoder
self.decoder = decoder
self.device = device
self.pass_enc2dec_hid = pass_enc2dec_hid
_force_en2dec_hid_conv = False
if self.pass_enc2dec_hid:
assert decoder.dec_hidden_dim == encoder.enc_hidden_dim, "Hidden Dimension of encoder and decoder must be same, or unset `pass_enc2dec_hid`"
if decoder.use_attention:
assert decoder.enc_outstate_dim == encoder.enc_directions*encoder.enc_hidden_dim,"Set `enc_out_dim` correctly in decoder"
assert self.pass_enc2dec_hid or decoder.use_attention, "No use of a decoder with No attention and No Hidden from Encoder"
self.use_conv_4_enc2dec_hid = False
if (
( self.pass_enc2dec_hid and
(encoder.enc_directions * encoder.enc_layers != decoder.dec_layers)
)
or _force_en2dec_hid_conv
):
if encoder.enc_rnn_type == "lstm" or encoder.enc_rnn_type == "lstm":
raise Exception("XlitError: conv for enc2dec_hid not implemented; Change the layer numbers appropriately")
self.use_conv_4_enc2dec_hid = True
self.enc_hid_1ax = encoder.enc_directions * encoder.enc_layers
self.dec_hid_1ax = decoder.dec_layers
self.e2d_hidden_conv = nn.Conv1d(self.enc_hid_1ax, self.dec_hid_1ax, 1)
def enc2dec_hidden(self, enc_hidden):
"""
enc_hidden: n_layer, batch_size, hidden_dim*num_directions
TODO: Implement the logic for LSTm bsed model
"""
# hidden: batch_size, enc_layer*num_directions, enc_hidden_dim
hidden = enc_hidden.permute(1,0,2).contiguous()
# hidden: batch_size, dec_layers, dec_hidden_dim -> [N,C,Tstep]
hidden = self.e2d_hidden_conv(hidden)
# hidden: dec_layers, batch_size , dec_hidden_dim
hidden_for_dec = hidden.permute(1,0,2).contiguous()
return hidden_for_dec
def active_beam_inference(self, src, beam_width=3, max_tgt_sz=50):
''' Search based decoding
src: (sequence_len)
'''
def _avg_score(p_tup):
""" Used for Sorting
TODO: Dividing by length of sequence power alpha as hyperparam
"""
return p_tup[0]
batch_size = 1
start_tok = src[0]
end_tok = src[-1]
src_sz = torch.tensor([len(src)])
src_ = src.unsqueeze(0)
# enc_output: (batch_size, padded_seq_length, enc_hidden_dim*num_direction)
# enc_hidden: (enc_layers*num_direction, batch_size, hidden_dim)
enc_output, enc_hidden = self.encoder(src_, src_sz)
if self.pass_enc2dec_hid:
# dec_hidden: dec_layers, batch_size , dec_hidden_dim
if self.use_conv_4_enc2dec_hid:
init_dec_hidden = self.enc2dec_hidden(enc_hidden)
else:
init_dec_hidden = enc_hidden
else:
# dec_hidden -> Will be initialized to zeros internally
init_dec_hidden = None
# top_pred[][0] = Σ-log_softmax
# top_pred[][1] = sequence torch.tensor shape: (1)
# top_pred[][2] = dec_hidden
top_pred_list = [ (0, start_tok.unsqueeze(0) , init_dec_hidden) ]
for t in range(max_tgt_sz):
cur_pred_list = []
for p_tup in top_pred_list:
if p_tup[1][-1] == end_tok:
cur_pred_list.append(p_tup)
continue
# dec_hidden: dec_layers, 1, hidden_dim
# dec_output: 1, output_dim
dec_output, dec_hidden, _ = self.decoder( x = p_tup[1][-1].view(1,1), #dec_input: (1,1)
hidden = p_tup[2],
enc_output = enc_output, )
## π{prob} = Σ{log(prob)} -> to prevent diminishing
# dec_output: (1, output_dim)
dec_output = nn.functional.log_softmax(dec_output, dim=1)
# pred_topk.values & pred_topk.indices: (1, beam_width)
pred_topk = torch.topk(dec_output, k=beam_width, dim=1)
for i in range(beam_width):
sig_logsmx_ = p_tup[0] + pred_topk.values[0][i]
# seq_tensor_ : (seq_len)
seq_tensor_ = torch.cat( (p_tup[1], pred_topk.indices[0][i].view(1)) )
cur_pred_list.append( (sig_logsmx_, seq_tensor_, dec_hidden) )
cur_pred_list.sort(key = _avg_score, reverse =True) # Maximized order
top_pred_list = cur_pred_list[:beam_width]
# check if end_tok of all topk
end_flags_ = [1 if t[1][-1] == end_tok else 0 for t in top_pred_list]
if beam_width == sum( end_flags_ ): break
pred_tnsr_list = [t[1] for t in top_pred_list ]
return pred_tnsr_list
class GlyphStrawboss():
def __init__(self, glyphs = 'en'):
""" list of letters in a language in unicode
lang: ISO Language code
glyphs: json file with script information
"""
if glyphs == 'en':
# Smallcase alone
self.glyphs = [chr(alpha) for alpha in range(97, 122+1)]
else:
self.dossier = json_load(open(glyphs, encoding='utf-8'))
self.glyphs = self.dossier["glyphs"]
self.numsym_map = self.dossier["numsym_map"]
self.char2idx = {}
self.idx2char = {}
self._create_index()
def _create_index(self):
self.char2idx['_'] = 0 #pad
self.char2idx['$'] = 1 #start
self.char2idx['#'] = 2 #end
self.char2idx['*'] = 3 #Mask
self.char2idx["'"] = 4 #apostrophe U+0027
self.char2idx['%'] = 5 #unused
self.char2idx['!'] = 6 #unused
# letter to index mapping
for idx, char in enumerate(self.glyphs):
self.char2idx[char] = idx + 7 # +7 token initially
# index to letter mapping
for char, idx in self.char2idx.items():
self.idx2char[idx] = char
def size(self):
return len(self.char2idx)
def word2xlitvec(self, word):
""" Converts given string of gyphs(word) to vector(numpy)
Also adds tokens for start and end
"""
try:
vec = [self.char2idx['$']] #start token
for i in list(word):
vec.append(self.char2idx[i])
vec.append(self.char2idx['#']) #end token
vec = np_asarray(vec, dtype=np_int64)
return vec
except Exception as error:
print("XlitError: In word:", word)
exit("Error Char not in Token: " + error)
def xlitvec2word(self, vector):
""" Converts vector(numpy) to string of glyphs(word)
"""
char_list = []
for i in vector:
char_list.append(self.idx2char[i])
word = "".join(char_list).replace('$','').replace('#','') # remove tokens
word = word.replace("_", "").replace('*','') # remove tokens
return word
class VocabSanitizer():
def __init__(self, data_file):
'''
data_file: path to file conatining vocabulary list
'''
self.vocab_set = set( json_load(open(data_file, encoding='utf-8')) )
def reposition(self, word_list):
'''Reorder Words in list
'''
new_list = []
temp_ = word_list.copy()
for v in word_list:
if v in self.vocab_set:
new_list.append(v)
temp_.remove(v)
new_list.extend(temp_)
return new_list
class XlitPiston():
"""
For handling prediction & post-processing of transliteration for a single language
Class dependency: Seq2Seq, GlyphStrawboss, VocabSanitizer
Global Variables: F_DIR
"""
def __init__(self, weight_path, tglyph_cfg_file,vocab_file,device:Device, iglyph_cfg_file = "en"):
self.device = device
self.in_glyph_obj = GlyphStrawboss(iglyph_cfg_file)
self.tgt_glyph_obj = GlyphStrawboss(glyphs = tglyph_cfg_file)
if vocab_file:
self.voc_sanitizer = VocabSanitizer(vocab_file)
else:
self.voc_sanitizer = None
self._numsym_set = set(json_load(open(tglyph_cfg_file, encoding='utf-8'))["numsym_map"].keys() )
self._inchar_set = set("abcdefghijklmnopqrstuvwxyz")
self._natscr_set = set().union(self.tgt_glyph_obj.glyphs,
sum(self.tgt_glyph_obj.numsym_map.values(),[]) )
## Model Config Static TODO: add defining in json support
input_dim = self.in_glyph_obj.size()
output_dim = self.tgt_glyph_obj.size()
enc_emb_dim = 300
dec_emb_dim = 300
enc_hidden_dim = 512
dec_hidden_dim = 512
rnn_type = "lstm"
enc2dec_hid = True
attention = True
enc_layers = 1
dec_layers = 2
m_dropout = 0
enc_bidirect = True
enc_outstate_dim = enc_hidden_dim * (2 if enc_bidirect else 1)
enc = Encoder( input_dim= input_dim, embed_dim = enc_emb_dim,
hidden_dim= enc_hidden_dim,
rnn_type = rnn_type, layers= enc_layers,
dropout= m_dropout, device = self.device,
bidirectional= enc_bidirect)
dec = Decoder( output_dim= output_dim, embed_dim = dec_emb_dim,
hidden_dim= dec_hidden_dim,
rnn_type = rnn_type, layers= dec_layers,
dropout= m_dropout,
use_attention = attention,
enc_outstate_dim= enc_outstate_dim,
device = self.device,)
self.model = Seq2Seq(enc, dec, pass_enc2dec_hid=enc2dec_hid, device=self.device)
self.model = self.model.to(self.device)
weights = torch.load( weight_path, map_location=torch.device(self.device))
self.model.load_state_dict(weights)
self.model.eval()
def character_model(self, word, beam_width = 1):
in_vec = torch.from_numpy(self.in_glyph_obj.word2xlitvec(word)).to(self.device)
## change to active or passive beam
p_out_list = self.model.active_beam_inference(in_vec, beam_width = beam_width)
p_result = [ self.tgt_glyph_obj.xlitvec2word(out.cpu().numpy()) for out in p_out_list]
if self.voc_sanitizer:
return self.voc_sanitizer.reposition(p_result)
#List type
return p_result
def numsym_model(self, seg):
''' tgt_glyph_obj.numsym_map[x] returns a list object
'''
if len(seg) == 1:
return [seg] + self.tgt_glyph_obj.numsym_map[seg]
a = [self.tgt_glyph_obj.numsym_map[n][0] for n in seg]
return [seg] + ["".join(a)]
def _word_segementer(self, sequence):
sequence = sequence.lower()
accepted = set().union(self._numsym_set, self._inchar_set, self._natscr_set)
# sequence = ''.join([i for i in sequence if i in accepted])
segment = []
idx = 0
seq_ = list(sequence)
while len(seq_):
# for Number-Symbol
temp = ""
while len(seq_) and seq_[0] in self._numsym_set:
temp += seq_[0]
seq_.pop(0)
if temp != "": segment.append(temp)
# for Target Chars
temp = ""
while len(seq_) and seq_[0] in self._natscr_set:
temp += seq_[0]
seq_.pop(0)
if temp != "": segment.append(temp)
# for Input-Roman Chars
temp = ""
while len(seq_) and seq_[0] in self._inchar_set:
temp += seq_[0]
seq_.pop(0)
if temp != "": segment.append(temp)
temp = ""
while len(seq_) and seq_[0] not in accepted:
temp += seq_[0]
seq_.pop(0)
if temp != "": segment.append(temp)
return segment
def inferencer(self, sequence, beam_width = 10):
seg = self._word_segementer(sequence[:120])
lit_seg = []
p = 0
while p < len(seg):
if seg[p][0] in self._natscr_set:
lit_seg.append([seg[p]])
p+=1
elif seg[p][0] in self._inchar_set:
lit_seg.append(self.character_model(seg[p], beam_width=beam_width))
p+=1
elif seg[p][0] in self._numsym_set: # num & punc
lit_seg.append(self.numsym_model(seg[p]))
p+=1
else:
lit_seg.append([ seg[p] ])
p+=1
## IF segment less/equal to 2 then return combinotorial,
## ELSE only return top1 of each result concatenated
if len(lit_seg) == 1:
final_result = lit_seg[0]
elif len(lit_seg) == 2:
final_result = [""]
for seg in lit_seg:
new_result = []
for s in seg:
for f in final_result:
new_result.append(f+s)
final_result = new_result
else:
new_result = []
for seg in lit_seg:
new_result.append(seg[0])
final_result = ["".join(new_result) ]
return final_result
|