| 1 | import os |
| 2 | from os.path import exists |
| 3 | import copy |
| 4 | import math |
| 5 | import time |
| 6 | from collections import Counter |
| 7 | |
| 8 | import spacy |
| 9 | import torch |
| 10 | import torch.nn as nn |
| 11 | from torch.nn.functional import log_softmax, pad |
| 12 | from torch.optim.lr_scheduler import LambdaLR |
| 13 | from torch.utils.data import DataLoader |
| 14 | from torch.utils.data.distributed import DistributedSampler |
| 15 | |
| 16 | |
| 17 | def attention(query, key, value, mask=None, dropout=None): |
| 18 | "Compute 'Scaled Dot Product Attention'" |
| 19 | d_k = query.size(-1) |
| 20 | scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k) |
| 21 | if mask is not None: |
| 22 | scores = scores.masked_fill(mask == 0, -1e9) |
| 23 | p_attn = scores.softmax(dim=-1) |
| 24 | if dropout is not None: |
| 25 | p_attn = dropout(p_attn) |
| 26 | return torch.matmul(p_attn, value), p_attn |
| 27 | |
| 28 | |
| 29 | def subsequent_mask(size): |
| 30 | "Mask out subsequent positions." |
| 31 | attn_shape = (1, size, size) |
| 32 | subsequent_mask = torch.triu(torch.ones(attn_shape), diagonal=1).type( |
| 33 | torch.uint8 |
| 34 | ) |
| 35 | return subsequent_mask == 0 |
| 36 | |
| 37 | |
| 38 | def clones(module, N): |
| 39 | "Produce N identical layers." |
| 40 | return nn.ModuleList([copy.deepcopy(module) for _ in range(N)]) |
| 41 | |
| 42 | |
| 43 | class MultiHeadedAttention(nn.Module): |
| 44 | def __init__(self, h, d_model, dropout=0.1): |
| 45 | "Take in model size and number of heads." |
| 46 | super(MultiHeadedAttention, self).__init__() |
| 47 | assert d_model % h == 0 |
| 48 | # We assume d_v always equals d_k |
| 49 | self.d_k = d_model // h |
| 50 | self.h = h |
| 51 | self.linears = clones(nn.Linear(d_model, d_model), 4) |
| 52 | self.attn = None |
| 53 | self.dropout = nn.Dropout(p=dropout) |
| 54 | |
| 55 | def forward(self, query, key, value, mask=None): |
| 56 | "Implements Figure 2" |
| 57 | if mask is not None: |
| 58 | # Same mask applied to all h heads. |
| 59 | mask = mask.unsqueeze(1) |
| 60 | nbatches = query.size(0) |
| 61 | |
| 62 | # 1) Do all the linear projections in batch from d_model => h x d_k |
| 63 | query, key, value = [ |
| 64 | lin(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2) |
| 65 | for lin, x in zip(self.linears, (query, key, value)) |
| 66 | ] |
| 67 | |
| 68 | # 2) Apply attention on all the projected vectors in batch. |
| 69 | x, self.attn = attention( |
| 70 | query, key, value, mask=mask, dropout=self.dropout |
| 71 | ) |
| 72 | |
| 73 | # 3) "Concat" using a view and apply a final linear. |
| 74 | x = ( |
| 75 | x.transpose(1, 2) |
| 76 | .contiguous() |
| 77 | .view(nbatches, -1, self.h * self.d_k) |
| 78 | ) |
| 79 | del query |
| 80 | del key |
| 81 | del value |
| 82 | return self.linears[-1](x) |
| 83 | |
| 84 | |
| 85 | class PositionwiseFeedForward(nn.Module): |
| 86 | "Implements FFN equation." |
| 87 | |
| 88 | def __init__(self, d_model, d_ff, dropout=0.1): |
| 89 | super(PositionwiseFeedForward, self).__init__() |
| 90 | self.w_1 = nn.Linear(d_model, d_ff) |
| 91 | self.w_2 = nn.Linear(d_ff, d_model) |
| 92 | self.dropout = nn.Dropout(dropout) |
| 93 | |
| 94 | def forward(self, x): |
| 95 | return self.w_2(self.dropout(self.w_1(x).relu())) |
| 96 | |
| 97 | |
| 98 | class Embeddings(nn.Module): |
| 99 | def __init__(self, d_model, vocab): |
| 100 | super(Embeddings, self).__init__() |
| 101 | self.lut = nn.Embedding(vocab, d_model) |
| 102 | self.d_model = d_model |
| 103 | |
| 104 | def forward(self, x): |
| 105 | return self.lut(x) * math.sqrt(self.d_model) |
| 106 | |
| 107 | |
| 108 | class PositionalEncoding(nn.Module): |
| 109 | "Implement the PE function." |
| 110 | |
| 111 | def __init__(self, d_model, dropout, max_len=5000): |
| 112 | super(PositionalEncoding, self).__init__() |
| 113 | self.dropout = nn.Dropout(p=dropout) |
| 114 | |
| 115 | # Compute the positional encodings once in log space. |
| 116 | pe = torch.zeros(max_len, d_model) |
| 117 | position = torch.arange(0, max_len).unsqueeze(1) |
| 118 | div_term = torch.exp( |
| 119 | torch.arange(0, d_model, 2) * -(math.log(10000.0) / d_model) |
| 120 | ) |
| 121 | pe[:, 0::2] = torch.sin(position * div_term) |
| 122 | pe[:, 1::2] = torch.cos(position * div_term) |
| 123 | pe = pe.unsqueeze(0) |
| 124 | self.register_buffer("pe", pe) |
| 125 | |
| 126 | def forward(self, x): |
| 127 | x = x + self.pe[:, : x.size(1)].requires_grad_(False) |
| 128 | return self.dropout(x) |
| 129 | |
| 130 | |
| 131 | class LayerNorm(nn.Module): |
| 132 | "Construct a layernorm module (See citation for details)." |
| 133 | |
| 134 | def __init__(self, features, eps=1e-6): |
| 135 | super(LayerNorm, self).__init__() |
| 136 | self.a_2 = nn.Parameter(torch.ones(features)) |
| 137 | self.b_2 = nn.Parameter(torch.zeros(features)) |
| 138 | self.eps = eps |
| 139 | |
| 140 | def forward(self, x): |
| 141 | mean = x.mean(-1, keepdim=True) |
| 142 | std = x.std(-1, keepdim=True) |
| 143 | return self.a_2 * (x - mean) / (std + self.eps) + self.b_2 |
| 144 | |
| 145 | |
| 146 | class SublayerConnection(nn.Module): |
| 147 | """ |
| 148 | A residual connection followed by a layer norm. |
| 149 | Note for code simplicity the norm is first as opposed to last. |
| 150 | """ |
| 151 | |
| 152 | def __init__(self, size, dropout): |
| 153 | super(SublayerConnection, self).__init__() |
| 154 | self.norm = LayerNorm(size) |
| 155 | self.dropout = nn.Dropout(dropout) |
| 156 | |
| 157 | def forward(self, x, sublayer): |
| 158 | "Apply residual connection to any sublayer with the same size." |
| 159 | return x + self.dropout(sublayer(self.norm(x))) |
| 160 | |
| 161 | |
| 162 | class EncoderLayer(nn.Module): |
| 163 | "Encoder is made up of self-attn and feed forward (defined below)" |
| 164 | |
| 165 | def __init__(self, size, self_attn, feed_forward, dropout): |
| 166 | super(EncoderLayer, self).__init__() |
| 167 | self.self_attn = self_attn |
| 168 | self.feed_forward = feed_forward |
| 169 | self.sublayer = clones(SublayerConnection(size, dropout), 2) |
| 170 | self.size = size |
| 171 | |
| 172 | def forward(self, x, mask): |
| 173 | "Follow Figure 1 (left) for connections." |
| 174 | x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask)) |
| 175 | return self.sublayer[1](x, self.feed_forward) |
| 176 | |
| 177 | |
| 178 | class Encoder(nn.Module): |
| 179 | "Core encoder is a stack of N layers" |
| 180 | |
| 181 | def __init__(self, layer, N): |
| 182 | super(Encoder, self).__init__() |
| 183 | self.layers = clones(layer, N) |
| 184 | self.norm = LayerNorm(layer.size) |
| 185 | |
| 186 | def forward(self, x, mask): |
| 187 | "Pass the input (and mask) through each layer in turn." |
| 188 | for layer in self.layers: |
| 189 | x = layer(x, mask) |
| 190 | return self.norm(x) |
| 191 | |
| 192 | |
| 193 | class DecoderLayer(nn.Module): |
| 194 | "Decoder is made of self-attn, src-attn, and feed forward (defined below)" |
| 195 | |
| 196 | def __init__(self, size, self_attn, src_attn, feed_forward, dropout): |
| 197 | super(DecoderLayer, self).__init__() |
| 198 | self.size = size |
| 199 | self.self_attn = self_attn |
| 200 | self.src_attn = src_attn |
| 201 | self.feed_forward = feed_forward |
| 202 | self.sublayer = clones(SublayerConnection(size, dropout), 3) |
| 203 | |
| 204 | def forward(self, x, memory, src_mask, tgt_mask): |
| 205 | "Follow Figure 1 (right) for connections." |
| 206 | m = memory |
| 207 | x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, tgt_mask)) |
| 208 | x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, src_mask)) |
| 209 | return self.sublayer[2](x, self.feed_forward) |
| 210 | |
| 211 | |
| 212 | class Decoder(nn.Module): |
| 213 | "Generic N layer decoder with masking." |
| 214 | |
| 215 | def __init__(self, layer, N): |
| 216 | super(Decoder, self).__init__() |
| 217 | self.layers = clones(layer, N) |
| 218 | self.norm = LayerNorm(layer.size) |
| 219 | |
| 220 | def forward(self, x, memory, src_mask, tgt_mask): |
| 221 | for layer in self.layers: |
| 222 | x = layer(x, memory, src_mask, tgt_mask) |
| 223 | return self.norm(x) |
| 224 | |
| 225 | |
| 226 | class EncoderDecoder(nn.Module): |
| 227 | """ |
| 228 | A standard Encoder-Decoder architecture. Base for this and many |
| 229 | other models. |
| 230 | """ |
| 231 | |
| 232 | def __init__(self, encoder, decoder, src_embed, tgt_embed, generator): |
| 233 | super(EncoderDecoder, self).__init__() |
| 234 | self.encoder = encoder |
| 235 | self.decoder = decoder |
| 236 | self.src_embed = src_embed |
| 237 | self.tgt_embed = tgt_embed |
| 238 | self.generator = generator |
| 239 | |
| 240 | def forward(self, src, tgt, src_mask, tgt_mask): |
| 241 | "Take in and process masked src and target sequences." |
| 242 | return self.decode(self.encode(src, src_mask), src_mask, tgt, tgt_mask) |
| 243 | |
| 244 | def encode(self, src, src_mask): |
| 245 | return self.encoder(self.src_embed(src), src_mask) |
| 246 | |
| 247 | def decode(self, memory, src_mask, tgt, tgt_mask): |
| 248 | return self.decoder(self.tgt_embed(tgt), memory, src_mask, tgt_mask) |
| 249 | |
| 250 | |
| 251 | class Generator(nn.Module): |
| 252 | "Define standard linear + softmax generation step." |
| 253 | |
| 254 | def __init__(self, d_model, vocab): |
| 255 | super(Generator, self).__init__() |
| 256 | self.proj = nn.Linear(d_model, vocab) |
| 257 | |
| 258 | def forward(self, x): |
| 259 | return log_softmax(self.proj(x), dim=-1) |
| 260 | |
| 261 | |
| 262 | def make_model( |
| 263 | src_vocab, tgt_vocab, N=6, d_model=512, d_ff=2048, h=8, dropout=0.1 |
| 264 | ): |
| 265 | "Helper: Construct a model from hyperparameters." |
| 266 | c = copy.deepcopy |
| 267 | attn = MultiHeadedAttention(h, d_model) |
| 268 | ff = PositionwiseFeedForward(d_model, d_ff, dropout) |
| 269 | position = PositionalEncoding(d_model, dropout) |
| 270 | model = EncoderDecoder( |
| 271 | Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N), |
| 272 | Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N), |
| 273 | nn.Sequential(Embeddings(d_model, src_vocab), c(position)), |
| 274 | nn.Sequential(Embeddings(d_model, tgt_vocab), c(position)), |
| 275 | Generator(d_model, tgt_vocab), |
| 276 | ) |
| 277 | |
| 278 | # This was important from their code. |
| 279 | # Initialize parameters with Glorot / fan_avg. |
| 280 | for p in model.parameters(): |
| 281 | if p.dim() > 1: |
| 282 | nn.init.xavier_uniform_(p) |
| 283 | return model |
| 284 | |
| 285 | |
| 286 | def greedy_decode(model, src, src_mask, max_len, start_symbol): |
| 287 | memory = model.encode(src, src_mask) |
| 288 | ys = torch.zeros(1, 1).fill_(start_symbol).type_as(src.data) |
| 289 | for i in range(max_len - 1): |
| 290 | out = model.decode( |
| 291 | memory, src_mask, ys, subsequent_mask(ys.size(1)).type_as(src.data) |
| 292 | ) |
| 293 | prob = model.generator(out[:, -1]) |
| 294 | _, next_word = torch.max(prob, dim=1) |
| 295 | next_word = next_word.data[0] |
| 296 | ys = torch.cat( |
| 297 | [ys, torch.zeros(1, 1).type_as(src.data).fill_(next_word)], dim=1 |
| 298 | ) |
| 299 | return ys |
| 300 | |
| 301 | |
| 302 | class Batch: |
| 303 | """Object for holding a batch of data with mask during training.""" |
| 304 | |
| 305 | def __init__(self, src, tgt=None, pad=2): # 2 = <blank> |
| 306 | self.src = src |
| 307 | self.src_mask = (src != pad).unsqueeze(-2) |
| 308 | if tgt is not None: |
| 309 | self.tgt = tgt[:, :-1] |
| 310 | self.tgt_y = tgt[:, 1:] |
| 311 | self.tgt_mask = self.make_std_mask(self.tgt, pad) |
| 312 | self.ntokens = (self.tgt_y != pad).data.sum() |
| 313 | |
| 314 | @staticmethod |
| 315 | def make_std_mask(tgt, pad): |
| 316 | "Create a mask to hide padding and future words." |
| 317 | tgt_mask = (tgt != pad).unsqueeze(-2) |
| 318 | tgt_mask = tgt_mask & subsequent_mask(tgt.size(-1)).type_as( |
| 319 | tgt_mask.data |
| 320 | ) |
| 321 | return tgt_mask |
| 322 | |
| 323 | |
| 324 | class LabelSmoothing(nn.Module): |
| 325 | "Implement label smoothing." |
| 326 | |
| 327 | def __init__(self, size, padding_idx, smoothing=0.0): |
| 328 | super(LabelSmoothing, self).__init__() |
| 329 | self.criterion = nn.KLDivLoss(reduction="sum") |
| 330 | self.padding_idx = padding_idx |
| 331 | self.confidence = 1.0 - smoothing |
| 332 | self.smoothing = smoothing |
| 333 | self.size = size |
| 334 | self.true_dist = None |
| 335 | |
| 336 | def forward(self, x, target): |
| 337 | assert x.size(1) == self.size |
| 338 | true_dist = x.data.clone() |
| 339 | true_dist.fill_(self.smoothing / (self.size - 2)) |
| 340 | true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence) |
| 341 | true_dist[:, self.padding_idx] = 0 |
| 342 | mask = torch.nonzero(target.data == self.padding_idx) |
| 343 | if mask.dim() > 0: |
| 344 | true_dist.index_fill_(0, mask.squeeze(), 0.0) |
| 345 | self.true_dist = true_dist |
| 346 | return self.criterion(x, true_dist.clone().detach()) |
| 347 | |
| 348 | |
| 349 | class SimpleLossCompute: |
| 350 | "A simple loss compute and train function." |
| 351 | |
| 352 | def __init__(self, generator, criterion): |
| 353 | self.generator = generator |
| 354 | self.criterion = criterion |
| 355 | |
| 356 | def __call__(self, x, y, norm): |
| 357 | x = self.generator(x) |
| 358 | sloss = ( |
| 359 | self.criterion( |
| 360 | x.contiguous().view(-1, x.size(-1)), y.contiguous().view(-1) |
| 361 | ) |
| 362 | / norm |
| 363 | ) |
| 364 | return sloss.data * norm, sloss |
| 365 | |
| 366 | |
| 367 | def rate(step, model_size, factor, warmup): |
| 368 | """ |
| 369 | we have to default the step to 1 for LambdaLR function |
| 370 | to avoid zero raising to negative power. |
| 371 | """ |
| 372 | if step == 0: |
| 373 | step = 1 |
| 374 | return factor * ( |
| 375 | model_size ** (-0.5) * min(step ** (-0.5), step * warmup ** (-1.5)) |
| 376 | ) |
| 377 | |
| 378 | |
| 379 | class TrainState: |
| 380 | """Track number of steps, examples, and tokens processed""" |
| 381 | |
| 382 | step: int = 0 # Steps in the current epoch |
| 383 | accum_step: int = 0 # Number of gradient accumulation steps |
| 384 | samples: int = 0 # total # of examples used |
| 385 | tokens: int = 0 # total # of tokens processed |
| 386 | |
| 387 | |
| 388 | def run_epoch( |
| 389 | data_iter, |
| 390 | model, |
| 391 | loss_compute, |
| 392 | optimizer, |
| 393 | scheduler, |
| 394 | mode="train", |
| 395 | accum_iter=1, |
| 396 | train_state=TrainState(), |
| 397 | ): |
| 398 | """Train a single epoch""" |
| 399 | start = time.time() |
| 400 | total_tokens = 0 |
| 401 | total_loss = 0 |
| 402 | tokens = 0 |
| 403 | n_accum = 0 |
| 404 | for i, batch in enumerate(data_iter): |
| 405 | out = model.forward( |
| 406 | batch.src, batch.tgt, batch.src_mask, batch.tgt_mask |
| 407 | ) |
| 408 | loss, loss_node = loss_compute(out, batch.tgt_y, batch.ntokens) |
| 409 | # loss_node = loss_node / accum_iter |
| 410 | if mode == "train" or mode == "train+log": |
| 411 | loss_node.backward() |
| 412 | train_state.step += 1 |
| 413 | train_state.samples += batch.src.shape[0] |
| 414 | train_state.tokens += batch.ntokens |
| 415 | if i % accum_iter == 0: |
| 416 | optimizer.step() |
| 417 | optimizer.zero_grad(set_to_none=True) |
| 418 | n_accum += 1 |
| 419 | train_state.accum_step += 1 |
| 420 | scheduler.step() |
| 421 | |
| 422 | total_loss += loss |
| 423 | total_tokens += batch.ntokens |
| 424 | tokens += batch.ntokens |
| 425 | if i % 40 == 1 and (mode == "train" or mode == "train+log"): |
| 426 | lr = optimizer.param_groups[0]["lr"] |
| 427 | elapsed = time.time() - start |
| 428 | print( |
| 429 | ( |
| 430 | "Epoch Step: %6d | Accumulation Step: %3d | Loss: %6.2f " |
| 431 | + "| Tokens / Sec: %7.1f | Learning Rate: %6.1e" |
| 432 | ) |
| 433 | % (i, n_accum, loss / batch.ntokens, tokens / elapsed, lr) |
| 434 | ) |
| 435 | start = time.time() |
| 436 | tokens = 0 |
| 437 | del loss |
| 438 | del loss_node |
| 439 | return total_loss / total_tokens, train_state |
| 440 | |
| 441 | |
| 442 | class DummyOptimizer(torch.optim.Optimizer): |
| 443 | def __init__(self): |
| 444 | self.param_groups = [{"lr": 0}] |
| 445 | None |
| 446 | |
| 447 | def step(self): |
| 448 | None |
| 449 | |
| 450 | def zero_grad(self, set_to_none=False): |
| 451 | None |
| 452 | |
| 453 | |
| 454 | class DummyScheduler: |
| 455 | def step(self): |
| 456 | None |
| 457 | |
| 458 | |
| 459 | def data_gen(V, batch_size, nbatches): |
| 460 | "Generate random data for a src-tgt copy task." |
| 461 | for i in range(nbatches): |
| 462 | data = torch.randint(1, V, size=(batch_size, 10)) |
| 463 | data[:, 0] = 1 |
| 464 | src = data.requires_grad_(False).clone().detach() |
| 465 | tgt = data.requires_grad_(False).clone().detach() |
| 466 | yield Batch(src, tgt, 0) |
| 467 | |
| 468 | |
| 469 | def example_simple_model(): |
| 470 | V = 11 |
| 471 | criterion = LabelSmoothing(size=V, padding_idx=0, smoothing=0.0) |
| 472 | model = make_model(V, V, N=2) |
| 473 | |
| 474 | optimizer = torch.optim.Adam( |
| 475 | model.parameters(), lr=0.5, betas=(0.9, 0.98), eps=1e-9 |
| 476 | ) |
| 477 | lr_scheduler = LambdaLR( |
| 478 | optimizer=optimizer, |
| 479 | lr_lambda=lambda step: rate( |
| 480 | step, model_size=model.src_embed[0].d_model, factor=1.0, warmup=400 |
| 481 | ), |
| 482 | ) |
| 483 | |
| 484 | batch_size = 80 |
| 485 | for epoch in range(20): |
| 486 | model.train() |
| 487 | run_epoch( |
| 488 | data_gen(V, batch_size, 20), |
| 489 | model, |
| 490 | SimpleLossCompute(model.generator, criterion), |
| 491 | optimizer, |
| 492 | lr_scheduler, |
| 493 | mode="train", |
| 494 | ) |
| 495 | model.eval() |
| 496 | run_epoch( |
| 497 | data_gen(V, batch_size, 5), |
| 498 | model, |
| 499 | SimpleLossCompute(model.generator, criterion), |
| 500 | DummyOptimizer(), |
| 501 | DummyScheduler(), |
| 502 | mode="eval", |
| 503 | )[0] |
| 504 | |
| 505 | model.eval() |
| 506 | src = torch.LongTensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]]) |
| 507 | max_len = src.shape[1] |
| 508 | src_mask = torch.ones(1, 1, max_len) |
| 509 | print(greedy_decode(model, src, src_mask, max_len=max_len, start_symbol=0)) |
| 510 | |
| 511 | |
| 512 | # torchtext 대신 (torchtext는 개발이 중단돼 최신 torch에서 import되지 않는다) |
| 513 | def Multi30k(language_pair=("de", "en")): |
| 514 | "torchtext.datasets.Multi30k와 같은 29,000 / 1,014 / 1,000 문장쌍" |
| 515 | from datasets import disable_progress_bars, load_dataset |
| 516 | |
| 517 | disable_progress_bars() |
| 518 | ds = load_dataset("bentrevett/multi30k", cache_dir=".data") |
| 519 | src, tgt = language_pair |
| 520 | return tuple( |
| 521 | list(zip(ds[split][src], ds[split][tgt])) |
| 522 | for split in ("train", "validation", "test") |
| 523 | ) |
| 524 | |
| 525 | |
| 526 | class Vocab: |
| 527 | "torchtext.vocab.Vocab 중 원본이 쓰는 부분" |
| 528 | |
| 529 | def __init__(self, itos): |
| 530 | self.itos = list(itos) |
| 531 | self.stoi = {t: i for i, t in enumerate(self.itos)} |
| 532 | self.default_index = None |
| 533 | |
| 534 | def __len__(self): |
| 535 | return len(self.itos) |
| 536 | |
| 537 | def __getitem__(self, token): |
| 538 | return self.stoi.get(token, self.default_index) |
| 539 | |
| 540 | def __call__(self, tokens): |
| 541 | return [self[t] for t in tokens] |
| 542 | |
| 543 | def set_default_index(self, index): |
| 544 | self.default_index = index |
| 545 | |
| 546 | def get_stoi(self): |
| 547 | return self.stoi |
| 548 | |
| 549 | def get_itos(self): |
| 550 | return self.itos |
| 551 | |
| 552 | |
| 553 | def build_vocab_from_iterator(iterator, min_freq=1, specials=None): |
| 554 | "specials 먼저, 나머지는 빈도 내림차순(같으면 사전순). torchtext 0.12와 같은 순서" |
| 555 | counter = Counter() |
| 556 | for tokens in iterator: |
| 557 | counter.update(tokens) |
| 558 | specials = specials or [] |
| 559 | for s in specials: |
| 560 | counter.pop(s, None) |
| 561 | ordered = sorted(counter.items(), key=lambda x: (-x[1], x[0])) |
| 562 | return Vocab(specials + [w for w, c in ordered if c >= min_freq]) |
| 563 | |
| 564 | |
| 565 | def to_map_style_dataset(iter_data): |
| 566 | "DistributedSampler가 len()을 요구해서 원본에 있던 함수" |
| 567 | return list(iter_data) |
| 568 | |
| 569 | |
| 570 | def load_tokenizers(): |
| 571 | |
| 572 | try: |
| 573 | spacy_de = spacy.load("de_core_news_sm") |
| 574 | except IOError: |
| 575 | os.system("python -m spacy download de_core_news_sm") |
| 576 | spacy_de = spacy.load("de_core_news_sm") |
| 577 | |
| 578 | try: |
| 579 | spacy_en = spacy.load("en_core_web_sm") |
| 580 | except IOError: |
| 581 | os.system("python -m spacy download en_core_web_sm") |
| 582 | spacy_en = spacy.load("en_core_web_sm") |
| 583 | |
| 584 | return spacy_de, spacy_en |
| 585 | |
| 586 | |
| 587 | def tokenize(text, tokenizer): |
| 588 | return [tok.text for tok in tokenizer.tokenizer(text)] |
| 589 | |
| 590 | |
| 591 | def yield_tokens(data_iter, tokenizer, index): |
| 592 | for from_to_tuple in data_iter: |
| 593 | yield tokenizer(from_to_tuple[index]) |
| 594 | |
| 595 | |
| 596 | def build_vocabulary(spacy_de, spacy_en): |
| 597 | def tokenize_de(text): |
| 598 | return tokenize(text, spacy_de) |
| 599 | |
| 600 | def tokenize_en(text): |
| 601 | return tokenize(text, spacy_en) |
| 602 | |
| 603 | print("Building German Vocabulary ...") |
| 604 | train, val, test = Multi30k(language_pair=("de", "en")) |
| 605 | vocab_src = build_vocab_from_iterator( |
| 606 | yield_tokens(train + val + test, tokenize_de, index=0), |
| 607 | min_freq=2, |
| 608 | specials=["<s>", "</s>", "<blank>", "<unk>"], |
| 609 | ) |
| 610 | |
| 611 | print("Building English Vocabulary ...") |
| 612 | train, val, test = Multi30k(language_pair=("de", "en")) |
| 613 | vocab_tgt = build_vocab_from_iterator( |
| 614 | yield_tokens(train + val + test, tokenize_en, index=1), |
| 615 | min_freq=2, |
| 616 | specials=["<s>", "</s>", "<blank>", "<unk>"], |
| 617 | ) |
| 618 | |
| 619 | vocab_src.set_default_index(vocab_src["<unk>"]) |
| 620 | vocab_tgt.set_default_index(vocab_tgt["<unk>"]) |
| 621 | |
| 622 | return vocab_src, vocab_tgt |
| 623 | |
| 624 | |
| 625 | def load_vocab(spacy_de, spacy_en): |
| 626 | if not exists("vocab.pt"): |
| 627 | vocab_src, vocab_tgt = build_vocabulary(spacy_de, spacy_en) |
| 628 | torch.save((vocab_src, vocab_tgt), "vocab.pt") |
| 629 | else: |
| 630 | vocab_src, vocab_tgt = torch.load("vocab.pt", weights_only=False) |
| 631 | print("Finished.\nVocabulary sizes:") |
| 632 | print(len(vocab_src)) |
| 633 | print(len(vocab_tgt)) |
| 634 | return vocab_src, vocab_tgt |
| 635 | |
| 636 | |
| 637 | def collate_batch( |
| 638 | batch, |
| 639 | src_pipeline, |
| 640 | tgt_pipeline, |
| 641 | src_vocab, |
| 642 | tgt_vocab, |
| 643 | device, |
| 644 | max_padding=128, |
| 645 | pad_id=2, |
| 646 | ): |
| 647 | bs_id = torch.tensor([0], device=device) # <s> token id |
| 648 | eos_id = torch.tensor([1], device=device) # </s> token id |
| 649 | src_list, tgt_list = [], [] |
| 650 | for (_src, _tgt) in batch: |
| 651 | processed_src = torch.cat( |
| 652 | [ |
| 653 | bs_id, |
| 654 | torch.tensor( |
| 655 | src_vocab(src_pipeline(_src)), |
| 656 | dtype=torch.int64, |
| 657 | device=device, |
| 658 | ), |
| 659 | eos_id, |
| 660 | ], |
| 661 | 0, |
| 662 | ) |
| 663 | processed_tgt = torch.cat( |
| 664 | [ |
| 665 | bs_id, |
| 666 | torch.tensor( |
| 667 | tgt_vocab(tgt_pipeline(_tgt)), |
| 668 | dtype=torch.int64, |
| 669 | device=device, |
| 670 | ), |
| 671 | eos_id, |
| 672 | ], |
| 673 | 0, |
| 674 | ) |
| 675 | src_list.append( |
| 676 | # warning - overwrites values for negative values of padding - len |
| 677 | pad( |
| 678 | processed_src, |
| 679 | ( |
| 680 | 0, |
| 681 | max_padding - len(processed_src), |
| 682 | ), |
| 683 | value=pad_id, |
| 684 | ) |
| 685 | ) |
| 686 | tgt_list.append( |
| 687 | pad( |
| 688 | processed_tgt, |
| 689 | (0, max_padding - len(processed_tgt)), |
| 690 | value=pad_id, |
| 691 | ) |
| 692 | ) |
| 693 | |
| 694 | src = torch.stack(src_list) |
| 695 | tgt = torch.stack(tgt_list) |
| 696 | return (src, tgt) |
| 697 | |
| 698 | |
| 699 | def create_dataloaders( |
| 700 | device, |
| 701 | vocab_src, |
| 702 | vocab_tgt, |
| 703 | spacy_de, |
| 704 | spacy_en, |
| 705 | batch_size=12000, |
| 706 | max_padding=128, |
| 707 | is_distributed=True, |
| 708 | ): |
| 709 | # def create_dataloaders(batch_size=12000): |
| 710 | def tokenize_de(text): |
| 711 | return tokenize(text, spacy_de) |
| 712 | |
| 713 | def tokenize_en(text): |
| 714 | return tokenize(text, spacy_en) |
| 715 | |
| 716 | def collate_fn(batch): |
| 717 | return collate_batch( |
| 718 | batch, |
| 719 | tokenize_de, |
| 720 | tokenize_en, |
| 721 | vocab_src, |
| 722 | vocab_tgt, |
| 723 | device, |
| 724 | max_padding=max_padding, |
| 725 | pad_id=vocab_src.get_stoi()["<blank>"], |
| 726 | ) |
| 727 | |
| 728 | train_iter, valid_iter, test_iter = Multi30k( |
| 729 | language_pair=("de", "en") |
| 730 | ) |
| 731 | |
| 732 | train_iter_map = to_map_style_dataset( |
| 733 | train_iter |
| 734 | ) # DistributedSampler needs a dataset len() |
| 735 | train_sampler = ( |
| 736 | DistributedSampler(train_iter_map) if is_distributed else None |
| 737 | ) |
| 738 | valid_iter_map = to_map_style_dataset(valid_iter) |
| 739 | valid_sampler = ( |
| 740 | DistributedSampler(valid_iter_map) if is_distributed else None |
| 741 | ) |
| 742 | |
| 743 | train_dataloader = DataLoader( |
| 744 | train_iter_map, |
| 745 | batch_size=batch_size, |
| 746 | shuffle=(train_sampler is None), |
| 747 | sampler=train_sampler, |
| 748 | collate_fn=collate_fn, |
| 749 | ) |
| 750 | valid_dataloader = DataLoader( |
| 751 | valid_iter_map, |
| 752 | batch_size=batch_size, |
| 753 | shuffle=(valid_sampler is None), |
| 754 | sampler=valid_sampler, |
| 755 | collate_fn=collate_fn, |
| 756 | ) |
| 757 | return train_dataloader, valid_dataloader |