| 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 altair as alt |
| 9 | import pandas as pd |
| 10 | import spacy |
| 11 | import torch |
| 12 | import torch.nn as nn |
| 13 | from torch.nn.functional import log_softmax, pad |
| 14 | from torch.optim.lr_scheduler import LambdaLR |
| 15 | from torch.utils.data import DataLoader |
| 16 | from torch.utils.data.distributed import DistributedSampler |
| 17 | import torch.distributed as dist |
| 18 | import torch.multiprocessing as mp |
| 19 | from torch.nn.parallel import DistributedDataParallel as DDP |
| 20 | import GPUtil |
| 21 | |
| 22 | |
| 23 | def attention(query, key, value, mask=None, dropout=None): |
| 24 | "Compute 'Scaled Dot Product Attention'" |
| 25 | d_k = query.size(-1) |
| 26 | scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k) |
| 27 | if mask is not None: |
| 28 | scores = scores.masked_fill(mask == 0, -1e9) |
| 29 | p_attn = scores.softmax(dim=-1) |
| 30 | if dropout is not None: |
| 31 | p_attn = dropout(p_attn) |
| 32 | return torch.matmul(p_attn, value), p_attn |
| 33 | |
| 34 | |
| 35 | def subsequent_mask(size): |
| 36 | "Mask out subsequent positions." |
| 37 | attn_shape = (1, size, size) |
| 38 | subsequent_mask = torch.triu(torch.ones(attn_shape), diagonal=1).type( |
| 39 | torch.uint8 |
| 40 | ) |
| 41 | return subsequent_mask == 0 |
| 42 | |
| 43 | |
| 44 | def clones(module, N): |
| 45 | "Produce N identical layers." |
| 46 | return nn.ModuleList([copy.deepcopy(module) for _ in range(N)]) |
| 47 | |
| 48 | |
| 49 | class MultiHeadedAttention(nn.Module): |
| 50 | def __init__(self, h, d_model, dropout=0.1): |
| 51 | "Take in model size and number of heads." |
| 52 | super(MultiHeadedAttention, self).__init__() |
| 53 | assert d_model % h == 0 |
| 54 | # We assume d_v always equals d_k |
| 55 | self.d_k = d_model // h |
| 56 | self.h = h |
| 57 | self.linears = clones(nn.Linear(d_model, d_model), 4) |
| 58 | self.attn = None |
| 59 | self.dropout = nn.Dropout(p=dropout) |
| 60 | |
| 61 | def forward(self, query, key, value, mask=None): |
| 62 | "Implements Figure 2" |
| 63 | if mask is not None: |
| 64 | # Same mask applied to all h heads. |
| 65 | mask = mask.unsqueeze(1) |
| 66 | nbatches = query.size(0) |
| 67 | |
| 68 | # 1) Do all the linear projections in batch from d_model => h x d_k |
| 69 | query, key, value = [ |
| 70 | lin(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2) |
| 71 | for lin, x in zip(self.linears, (query, key, value)) |
| 72 | ] |
| 73 | |
| 74 | # 2) Apply attention on all the projected vectors in batch. |
| 75 | x, self.attn = attention( |
| 76 | query, key, value, mask=mask, dropout=self.dropout |
| 77 | ) |
| 78 | |
| 79 | # 3) "Concat" using a view and apply a final linear. |
| 80 | x = ( |
| 81 | x.transpose(1, 2) |
| 82 | .contiguous() |
| 83 | .view(nbatches, -1, self.h * self.d_k) |
| 84 | ) |
| 85 | del query |
| 86 | del key |
| 87 | del value |
| 88 | return self.linears[-1](x) |
| 89 | |
| 90 | |
| 91 | class PositionwiseFeedForward(nn.Module): |
| 92 | "Implements FFN equation." |
| 93 | |
| 94 | def __init__(self, d_model, d_ff, dropout=0.1): |
| 95 | super(PositionwiseFeedForward, self).__init__() |
| 96 | self.w_1 = nn.Linear(d_model, d_ff) |
| 97 | self.w_2 = nn.Linear(d_ff, d_model) |
| 98 | self.dropout = nn.Dropout(dropout) |
| 99 | |
| 100 | def forward(self, x): |
| 101 | return self.w_2(self.dropout(self.w_1(x).relu())) |
| 102 | |
| 103 | |
| 104 | class Embeddings(nn.Module): |
| 105 | def __init__(self, d_model, vocab): |
| 106 | super(Embeddings, self).__init__() |
| 107 | self.lut = nn.Embedding(vocab, d_model) |
| 108 | self.d_model = d_model |
| 109 | |
| 110 | def forward(self, x): |
| 111 | return self.lut(x) * math.sqrt(self.d_model) |
| 112 | |
| 113 | |
| 114 | class PositionalEncoding(nn.Module): |
| 115 | "Implement the PE function." |
| 116 | |
| 117 | def __init__(self, d_model, dropout, max_len=5000): |
| 118 | super(PositionalEncoding, self).__init__() |
| 119 | self.dropout = nn.Dropout(p=dropout) |
| 120 | |
| 121 | # Compute the positional encodings once in log space. |
| 122 | pe = torch.zeros(max_len, d_model) |
| 123 | position = torch.arange(0, max_len).unsqueeze(1) |
| 124 | div_term = torch.exp( |
| 125 | torch.arange(0, d_model, 2) * -(math.log(10000.0) / d_model) |
| 126 | ) |
| 127 | pe[:, 0::2] = torch.sin(position * div_term) |
| 128 | pe[:, 1::2] = torch.cos(position * div_term) |
| 129 | pe = pe.unsqueeze(0) |
| 130 | self.register_buffer("pe", pe) |
| 131 | |
| 132 | def forward(self, x): |
| 133 | x = x + self.pe[:, : x.size(1)].requires_grad_(False) |
| 134 | return self.dropout(x) |
| 135 | |
| 136 | |
| 137 | class LayerNorm(nn.Module): |
| 138 | "Construct a layernorm module (See citation for details)." |
| 139 | |
| 140 | def __init__(self, features, eps=1e-6): |
| 141 | super(LayerNorm, self).__init__() |
| 142 | self.a_2 = nn.Parameter(torch.ones(features)) |
| 143 | self.b_2 = nn.Parameter(torch.zeros(features)) |
| 144 | self.eps = eps |
| 145 | |
| 146 | def forward(self, x): |
| 147 | mean = x.mean(-1, keepdim=True) |
| 148 | std = x.std(-1, keepdim=True) |
| 149 | return self.a_2 * (x - mean) / (std + self.eps) + self.b_2 |
| 150 | |
| 151 | |
| 152 | class SublayerConnection(nn.Module): |
| 153 | """ |
| 154 | A residual connection followed by a layer norm. |
| 155 | Note for code simplicity the norm is first as opposed to last. |
| 156 | """ |
| 157 | |
| 158 | def __init__(self, size, dropout): |
| 159 | super(SublayerConnection, self).__init__() |
| 160 | self.norm = LayerNorm(size) |
| 161 | self.dropout = nn.Dropout(dropout) |
| 162 | |
| 163 | def forward(self, x, sublayer): |
| 164 | "Apply residual connection to any sublayer with the same size." |
| 165 | return x + self.dropout(sublayer(self.norm(x))) |
| 166 | |
| 167 | |
| 168 | class EncoderLayer(nn.Module): |
| 169 | "Encoder is made up of self-attn and feed forward (defined below)" |
| 170 | |
| 171 | def __init__(self, size, self_attn, feed_forward, dropout): |
| 172 | super(EncoderLayer, self).__init__() |
| 173 | self.self_attn = self_attn |
| 174 | self.feed_forward = feed_forward |
| 175 | self.sublayer = clones(SublayerConnection(size, dropout), 2) |
| 176 | self.size = size |
| 177 | |
| 178 | def forward(self, x, mask): |
| 179 | "Follow Figure 1 (left) for connections." |
| 180 | x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask)) |
| 181 | return self.sublayer[1](x, self.feed_forward) |
| 182 | |
| 183 | |
| 184 | class Encoder(nn.Module): |
| 185 | "Core encoder is a stack of N layers" |
| 186 | |
| 187 | def __init__(self, layer, N): |
| 188 | super(Encoder, self).__init__() |
| 189 | self.layers = clones(layer, N) |
| 190 | self.norm = LayerNorm(layer.size) |
| 191 | |
| 192 | def forward(self, x, mask): |
| 193 | "Pass the input (and mask) through each layer in turn." |
| 194 | for layer in self.layers: |
| 195 | x = layer(x, mask) |
| 196 | return self.norm(x) |
| 197 | |
| 198 | |
| 199 | class DecoderLayer(nn.Module): |
| 200 | "Decoder is made of self-attn, src-attn, and feed forward (defined below)" |
| 201 | |
| 202 | def __init__(self, size, self_attn, src_attn, feed_forward, dropout): |
| 203 | super(DecoderLayer, self).__init__() |
| 204 | self.size = size |
| 205 | self.self_attn = self_attn |
| 206 | self.src_attn = src_attn |
| 207 | self.feed_forward = feed_forward |
| 208 | self.sublayer = clones(SublayerConnection(size, dropout), 3) |
| 209 | |
| 210 | def forward(self, x, memory, src_mask, tgt_mask): |
| 211 | "Follow Figure 1 (right) for connections." |
| 212 | m = memory |
| 213 | x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, tgt_mask)) |
| 214 | x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, src_mask)) |
| 215 | return self.sublayer[2](x, self.feed_forward) |
| 216 | |
| 217 | |
| 218 | class Decoder(nn.Module): |
| 219 | "Generic N layer decoder with masking." |
| 220 | |
| 221 | def __init__(self, layer, N): |
| 222 | super(Decoder, self).__init__() |
| 223 | self.layers = clones(layer, N) |
| 224 | self.norm = LayerNorm(layer.size) |
| 225 | |
| 226 | def forward(self, x, memory, src_mask, tgt_mask): |
| 227 | for layer in self.layers: |
| 228 | x = layer(x, memory, src_mask, tgt_mask) |
| 229 | return self.norm(x) |
| 230 | |
| 231 | |
| 232 | class EncoderDecoder(nn.Module): |
| 233 | """ |
| 234 | A standard Encoder-Decoder architecture. Base for this and many |
| 235 | other models. |
| 236 | """ |
| 237 | |
| 238 | def __init__(self, encoder, decoder, src_embed, tgt_embed, generator): |
| 239 | super(EncoderDecoder, self).__init__() |
| 240 | self.encoder = encoder |
| 241 | self.decoder = decoder |
| 242 | self.src_embed = src_embed |
| 243 | self.tgt_embed = tgt_embed |
| 244 | self.generator = generator |
| 245 | |
| 246 | def forward(self, src, tgt, src_mask, tgt_mask): |
| 247 | "Take in and process masked src and target sequences." |
| 248 | return self.decode(self.encode(src, src_mask), src_mask, tgt, tgt_mask) |
| 249 | |
| 250 | def encode(self, src, src_mask): |
| 251 | return self.encoder(self.src_embed(src), src_mask) |
| 252 | |
| 253 | def decode(self, memory, src_mask, tgt, tgt_mask): |
| 254 | return self.decoder(self.tgt_embed(tgt), memory, src_mask, tgt_mask) |
| 255 | |
| 256 | |
| 257 | class Generator(nn.Module): |
| 258 | "Define standard linear + softmax generation step." |
| 259 | |
| 260 | def __init__(self, d_model, vocab): |
| 261 | super(Generator, self).__init__() |
| 262 | self.proj = nn.Linear(d_model, vocab) |
| 263 | |
| 264 | def forward(self, x): |
| 265 | return log_softmax(self.proj(x), dim=-1) |
| 266 | |
| 267 | |
| 268 | def make_model( |
| 269 | src_vocab, tgt_vocab, N=6, d_model=512, d_ff=2048, h=8, dropout=0.1 |
| 270 | ): |
| 271 | "Helper: Construct a model from hyperparameters." |
| 272 | c = copy.deepcopy |
| 273 | attn = MultiHeadedAttention(h, d_model) |
| 274 | ff = PositionwiseFeedForward(d_model, d_ff, dropout) |
| 275 | position = PositionalEncoding(d_model, dropout) |
| 276 | model = EncoderDecoder( |
| 277 | Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N), |
| 278 | Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N), |
| 279 | nn.Sequential(Embeddings(d_model, src_vocab), c(position)), |
| 280 | nn.Sequential(Embeddings(d_model, tgt_vocab), c(position)), |
| 281 | Generator(d_model, tgt_vocab), |
| 282 | ) |
| 283 | |
| 284 | # This was important from their code. |
| 285 | # Initialize parameters with Glorot / fan_avg. |
| 286 | for p in model.parameters(): |
| 287 | if p.dim() > 1: |
| 288 | nn.init.xavier_uniform_(p) |
| 289 | return model |
| 290 | |
| 291 | |
| 292 | def greedy_decode(model, src, src_mask, max_len, start_symbol): |
| 293 | memory = model.encode(src, src_mask) |
| 294 | ys = torch.zeros(1, 1).fill_(start_symbol).type_as(src.data) |
| 295 | for i in range(max_len - 1): |
| 296 | out = model.decode( |
| 297 | memory, src_mask, ys, subsequent_mask(ys.size(1)).type_as(src.data) |
| 298 | ) |
| 299 | prob = model.generator(out[:, -1]) |
| 300 | _, next_word = torch.max(prob, dim=1) |
| 301 | next_word = next_word.data[0] |
| 302 | ys = torch.cat( |
| 303 | [ys, torch.zeros(1, 1).type_as(src.data).fill_(next_word)], dim=1 |
| 304 | ) |
| 305 | return ys |
| 306 | |
| 307 | |
| 308 | class Batch: |
| 309 | """Object for holding a batch of data with mask during training.""" |
| 310 | |
| 311 | def __init__(self, src, tgt=None, pad=2): # 2 = <blank> |
| 312 | self.src = src |
| 313 | self.src_mask = (src != pad).unsqueeze(-2) |
| 314 | if tgt is not None: |
| 315 | self.tgt = tgt[:, :-1] |
| 316 | self.tgt_y = tgt[:, 1:] |
| 317 | self.tgt_mask = self.make_std_mask(self.tgt, pad) |
| 318 | self.ntokens = (self.tgt_y != pad).data.sum() |
| 319 | |
| 320 | @staticmethod |
| 321 | def make_std_mask(tgt, pad): |
| 322 | "Create a mask to hide padding and future words." |
| 323 | tgt_mask = (tgt != pad).unsqueeze(-2) |
| 324 | tgt_mask = tgt_mask & subsequent_mask(tgt.size(-1)).type_as( |
| 325 | tgt_mask.data |
| 326 | ) |
| 327 | return tgt_mask |
| 328 | |
| 329 | |
| 330 | class LabelSmoothing(nn.Module): |
| 331 | "Implement label smoothing." |
| 332 | |
| 333 | def __init__(self, size, padding_idx, smoothing=0.0): |
| 334 | super(LabelSmoothing, self).__init__() |
| 335 | self.criterion = nn.KLDivLoss(reduction="sum") |
| 336 | self.padding_idx = padding_idx |
| 337 | self.confidence = 1.0 - smoothing |
| 338 | self.smoothing = smoothing |
| 339 | self.size = size |
| 340 | self.true_dist = None |
| 341 | |
| 342 | def forward(self, x, target): |
| 343 | assert x.size(1) == self.size |
| 344 | true_dist = x.data.clone() |
| 345 | true_dist.fill_(self.smoothing / (self.size - 2)) |
| 346 | true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence) |
| 347 | true_dist[:, self.padding_idx] = 0 |
| 348 | mask = torch.nonzero(target.data == self.padding_idx) |
| 349 | if mask.dim() > 0: |
| 350 | true_dist.index_fill_(0, mask.squeeze(), 0.0) |
| 351 | self.true_dist = true_dist |
| 352 | return self.criterion(x, true_dist.clone().detach()) |
| 353 | |
| 354 | |
| 355 | class SimpleLossCompute: |
| 356 | "A simple loss compute and train function." |
| 357 | |
| 358 | def __init__(self, generator, criterion): |
| 359 | self.generator = generator |
| 360 | self.criterion = criterion |
| 361 | |
| 362 | def __call__(self, x, y, norm): |
| 363 | x = self.generator(x) |
| 364 | sloss = ( |
| 365 | self.criterion( |
| 366 | x.contiguous().view(-1, x.size(-1)), y.contiguous().view(-1) |
| 367 | ) |
| 368 | / norm |
| 369 | ) |
| 370 | return sloss.data * norm, sloss |
| 371 | |
| 372 | |
| 373 | def rate(step, model_size, factor, warmup): |
| 374 | """ |
| 375 | we have to default the step to 1 for LambdaLR function |
| 376 | to avoid zero raising to negative power. |
| 377 | """ |
| 378 | if step == 0: |
| 379 | step = 1 |
| 380 | return factor * ( |
| 381 | model_size ** (-0.5) * min(step ** (-0.5), step * warmup ** (-1.5)) |
| 382 | ) |
| 383 | |
| 384 | |
| 385 | class TrainState: |
| 386 | """Track number of steps, examples, and tokens processed""" |
| 387 | |
| 388 | step: int = 0 # Steps in the current epoch |
| 389 | accum_step: int = 0 # Number of gradient accumulation steps |
| 390 | samples: int = 0 # total # of examples used |
| 391 | tokens: int = 0 # total # of tokens processed |
| 392 | |
| 393 | |
| 394 | def run_epoch( |
| 395 | data_iter, |
| 396 | model, |
| 397 | loss_compute, |
| 398 | optimizer, |
| 399 | scheduler, |
| 400 | mode="train", |
| 401 | accum_iter=1, |
| 402 | train_state=TrainState(), |
| 403 | ): |
| 404 | """Train a single epoch""" |
| 405 | start = time.time() |
| 406 | total_tokens = 0 |
| 407 | total_loss = 0 |
| 408 | tokens = 0 |
| 409 | n_accum = 0 |
| 410 | for i, batch in enumerate(data_iter): |
| 411 | out = model.forward( |
| 412 | batch.src, batch.tgt, batch.src_mask, batch.tgt_mask |
| 413 | ) |
| 414 | loss, loss_node = loss_compute(out, batch.tgt_y, batch.ntokens) |
| 415 | # loss_node = loss_node / accum_iter |
| 416 | if mode == "train" or mode == "train+log": |
| 417 | loss_node.backward() |
| 418 | train_state.step += 1 |
| 419 | train_state.samples += batch.src.shape[0] |
| 420 | train_state.tokens += batch.ntokens |
| 421 | if i % accum_iter == 0: |
| 422 | optimizer.step() |
| 423 | optimizer.zero_grad(set_to_none=True) |
| 424 | n_accum += 1 |
| 425 | train_state.accum_step += 1 |
| 426 | scheduler.step() |
| 427 | |
| 428 | total_loss += loss |
| 429 | total_tokens += batch.ntokens |
| 430 | tokens += batch.ntokens |
| 431 | if i % 40 == 1 and (mode == "train" or mode == "train+log"): |
| 432 | lr = optimizer.param_groups[0]["lr"] |
| 433 | elapsed = time.time() - start |
| 434 | print( |
| 435 | ( |
| 436 | "Epoch Step: %6d | Accumulation Step: %3d | Loss: %6.2f " |
| 437 | + "| Tokens / Sec: %7.1f | Learning Rate: %6.1e" |
| 438 | ) |
| 439 | % (i, n_accum, loss / batch.ntokens, tokens / elapsed, lr) |
| 440 | ) |
| 441 | start = time.time() |
| 442 | tokens = 0 |
| 443 | del loss |
| 444 | del loss_node |
| 445 | return total_loss / total_tokens, train_state |
| 446 | |
| 447 | |
| 448 | class DummyOptimizer(torch.optim.Optimizer): |
| 449 | def __init__(self): |
| 450 | self.param_groups = [{"lr": 0}] |
| 451 | None |
| 452 | |
| 453 | def step(self): |
| 454 | None |
| 455 | |
| 456 | def zero_grad(self, set_to_none=False): |
| 457 | None |
| 458 | |
| 459 | |
| 460 | class DummyScheduler: |
| 461 | def step(self): |
| 462 | None |
| 463 | |
| 464 | |
| 465 | def data_gen(V, batch_size, nbatches): |
| 466 | "Generate random data for a src-tgt copy task." |
| 467 | for i in range(nbatches): |
| 468 | data = torch.randint(1, V, size=(batch_size, 10)) |
| 469 | data[:, 0] = 1 |
| 470 | src = data.requires_grad_(False).clone().detach() |
| 471 | tgt = data.requires_grad_(False).clone().detach() |
| 472 | yield Batch(src, tgt, 0) |
| 473 | |
| 474 | |
| 475 | def example_simple_model(): |
| 476 | V = 11 |
| 477 | criterion = LabelSmoothing(size=V, padding_idx=0, smoothing=0.0) |
| 478 | model = make_model(V, V, N=2) |
| 479 | |
| 480 | optimizer = torch.optim.Adam( |
| 481 | model.parameters(), lr=0.5, betas=(0.9, 0.98), eps=1e-9 |
| 482 | ) |
| 483 | lr_scheduler = LambdaLR( |
| 484 | optimizer=optimizer, |
| 485 | lr_lambda=lambda step: rate( |
| 486 | step, model_size=model.src_embed[0].d_model, factor=1.0, warmup=400 |
| 487 | ), |
| 488 | ) |
| 489 | |
| 490 | batch_size = 80 |
| 491 | for epoch in range(20): |
| 492 | model.train() |
| 493 | run_epoch( |
| 494 | data_gen(V, batch_size, 20), |
| 495 | model, |
| 496 | SimpleLossCompute(model.generator, criterion), |
| 497 | optimizer, |
| 498 | lr_scheduler, |
| 499 | mode="train", |
| 500 | ) |
| 501 | model.eval() |
| 502 | run_epoch( |
| 503 | data_gen(V, batch_size, 5), |
| 504 | model, |
| 505 | SimpleLossCompute(model.generator, criterion), |
| 506 | DummyOptimizer(), |
| 507 | DummyScheduler(), |
| 508 | mode="eval", |
| 509 | )[0] |
| 510 | |
| 511 | model.eval() |
| 512 | src = torch.LongTensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]]) |
| 513 | max_len = src.shape[1] |
| 514 | src_mask = torch.ones(1, 1, max_len) |
| 515 | print(greedy_decode(model, src, src_mask, max_len=max_len, start_symbol=0)) |
| 516 | |
| 517 | |
| 518 | # torchtext 대신 (torchtext는 개발이 중단돼 최신 torch에서 import되지 않는다) |
| 519 | def Multi30k(language_pair=("de", "en")): |
| 520 | "torchtext.datasets.Multi30k와 같은 29,000 / 1,014 / 1,000 문장쌍" |
| 521 | from datasets import disable_progress_bars, load_dataset |
| 522 | |
| 523 | disable_progress_bars() |
| 524 | ds = load_dataset("bentrevett/multi30k", cache_dir=".data") |
| 525 | src, tgt = language_pair |
| 526 | return tuple( |
| 527 | list(zip(ds[split][src], ds[split][tgt])) |
| 528 | for split in ("train", "validation", "test") |
| 529 | ) |
| 530 | |
| 531 | |
| 532 | class Vocab: |
| 533 | "torchtext.vocab.Vocab 중 원본이 쓰는 부분" |
| 534 | |
| 535 | def __init__(self, itos): |
| 536 | self.itos = list(itos) |
| 537 | self.stoi = {t: i for i, t in enumerate(self.itos)} |
| 538 | self.default_index = None |
| 539 | |
| 540 | def __len__(self): |
| 541 | return len(self.itos) |
| 542 | |
| 543 | def __getitem__(self, token): |
| 544 | return self.stoi.get(token, self.default_index) |
| 545 | |
| 546 | def __call__(self, tokens): |
| 547 | return [self[t] for t in tokens] |
| 548 | |
| 549 | def set_default_index(self, index): |
| 550 | self.default_index = index |
| 551 | |
| 552 | def get_stoi(self): |
| 553 | return self.stoi |
| 554 | |
| 555 | def get_itos(self): |
| 556 | return self.itos |
| 557 | |
| 558 | |
| 559 | def build_vocab_from_iterator(iterator, min_freq=1, specials=None): |
| 560 | "specials 먼저, 나머지는 빈도 내림차순(같으면 사전순). torchtext 0.12와 같은 순서" |
| 561 | counter = Counter() |
| 562 | for tokens in iterator: |
| 563 | counter.update(tokens) |
| 564 | specials = specials or [] |
| 565 | for s in specials: |
| 566 | counter.pop(s, None) |
| 567 | ordered = sorted(counter.items(), key=lambda x: (-x[1], x[0])) |
| 568 | return Vocab(specials + [w for w, c in ordered if c >= min_freq]) |
| 569 | |
| 570 | |
| 571 | def to_map_style_dataset(iter_data): |
| 572 | "DistributedSampler가 len()을 요구해서 원본에 있던 함수" |
| 573 | return list(iter_data) |
| 574 | |
| 575 | |
| 576 | def load_tokenizers(): |
| 577 | |
| 578 | try: |
| 579 | spacy_de = spacy.load("de_core_news_sm") |
| 580 | except IOError: |
| 581 | os.system("python -m spacy download de_core_news_sm") |
| 582 | spacy_de = spacy.load("de_core_news_sm") |
| 583 | |
| 584 | try: |
| 585 | spacy_en = spacy.load("en_core_web_sm") |
| 586 | except IOError: |
| 587 | os.system("python -m spacy download en_core_web_sm") |
| 588 | spacy_en = spacy.load("en_core_web_sm") |
| 589 | |
| 590 | return spacy_de, spacy_en |
| 591 | |
| 592 | |
| 593 | def tokenize(text, tokenizer): |
| 594 | return [tok.text for tok in tokenizer.tokenizer(text)] |
| 595 | |
| 596 | |
| 597 | def yield_tokens(data_iter, tokenizer, index): |
| 598 | for from_to_tuple in data_iter: |
| 599 | yield tokenizer(from_to_tuple[index]) |
| 600 | |
| 601 | |
| 602 | def build_vocabulary(spacy_de, spacy_en): |
| 603 | def tokenize_de(text): |
| 604 | return tokenize(text, spacy_de) |
| 605 | |
| 606 | def tokenize_en(text): |
| 607 | return tokenize(text, spacy_en) |
| 608 | |
| 609 | print("Building German Vocabulary ...") |
| 610 | train, val, test = Multi30k(language_pair=("de", "en")) |
| 611 | vocab_src = build_vocab_from_iterator( |
| 612 | yield_tokens(train + val + test, tokenize_de, index=0), |
| 613 | min_freq=2, |
| 614 | specials=["<s>", "</s>", "<blank>", "<unk>"], |
| 615 | ) |
| 616 | |
| 617 | print("Building English Vocabulary ...") |
| 618 | train, val, test = Multi30k(language_pair=("de", "en")) |
| 619 | vocab_tgt = build_vocab_from_iterator( |
| 620 | yield_tokens(train + val + test, tokenize_en, index=1), |
| 621 | min_freq=2, |
| 622 | specials=["<s>", "</s>", "<blank>", "<unk>"], |
| 623 | ) |
| 624 | |
| 625 | vocab_src.set_default_index(vocab_src["<unk>"]) |
| 626 | vocab_tgt.set_default_index(vocab_tgt["<unk>"]) |
| 627 | |
| 628 | return vocab_src, vocab_tgt |
| 629 | |
| 630 | |
| 631 | def load_vocab(spacy_de, spacy_en): |
| 632 | if not exists("vocab.pt"): |
| 633 | vocab_src, vocab_tgt = build_vocabulary(spacy_de, spacy_en) |
| 634 | torch.save((vocab_src, vocab_tgt), "vocab.pt") |
| 635 | else: |
| 636 | vocab_src, vocab_tgt = torch.load("vocab.pt", weights_only=False) |
| 637 | print("Finished.\nVocabulary sizes:") |
| 638 | print(len(vocab_src)) |
| 639 | print(len(vocab_tgt)) |
| 640 | return vocab_src, vocab_tgt |
| 641 | |
| 642 | |
| 643 | def collate_batch( |
| 644 | batch, |
| 645 | src_pipeline, |
| 646 | tgt_pipeline, |
| 647 | src_vocab, |
| 648 | tgt_vocab, |
| 649 | device, |
| 650 | max_padding=128, |
| 651 | pad_id=2, |
| 652 | ): |
| 653 | bs_id = torch.tensor([0], device=device) # <s> token id |
| 654 | eos_id = torch.tensor([1], device=device) # </s> token id |
| 655 | src_list, tgt_list = [], [] |
| 656 | for (_src, _tgt) in batch: |
| 657 | processed_src = torch.cat( |
| 658 | [ |
| 659 | bs_id, |
| 660 | torch.tensor( |
| 661 | src_vocab(src_pipeline(_src)), |
| 662 | dtype=torch.int64, |
| 663 | device=device, |
| 664 | ), |
| 665 | eos_id, |
| 666 | ], |
| 667 | 0, |
| 668 | ) |
| 669 | processed_tgt = torch.cat( |
| 670 | [ |
| 671 | bs_id, |
| 672 | torch.tensor( |
| 673 | tgt_vocab(tgt_pipeline(_tgt)), |
| 674 | dtype=torch.int64, |
| 675 | device=device, |
| 676 | ), |
| 677 | eos_id, |
| 678 | ], |
| 679 | 0, |
| 680 | ) |
| 681 | src_list.append( |
| 682 | # warning - overwrites values for negative values of padding - len |
| 683 | pad( |
| 684 | processed_src, |
| 685 | ( |
| 686 | 0, |
| 687 | max_padding - len(processed_src), |
| 688 | ), |
| 689 | value=pad_id, |
| 690 | ) |
| 691 | ) |
| 692 | tgt_list.append( |
| 693 | pad( |
| 694 | processed_tgt, |
| 695 | (0, max_padding - len(processed_tgt)), |
| 696 | value=pad_id, |
| 697 | ) |
| 698 | ) |
| 699 | |
| 700 | src = torch.stack(src_list) |
| 701 | tgt = torch.stack(tgt_list) |
| 702 | return (src, tgt) |
| 703 | |
| 704 | |
| 705 | def create_dataloaders( |
| 706 | device, |
| 707 | vocab_src, |
| 708 | vocab_tgt, |
| 709 | spacy_de, |
| 710 | spacy_en, |
| 711 | batch_size=12000, |
| 712 | max_padding=128, |
| 713 | is_distributed=True, |
| 714 | ): |
| 715 | # def create_dataloaders(batch_size=12000): |
| 716 | def tokenize_de(text): |
| 717 | return tokenize(text, spacy_de) |
| 718 | |
| 719 | def tokenize_en(text): |
| 720 | return tokenize(text, spacy_en) |
| 721 | |
| 722 | def collate_fn(batch): |
| 723 | return collate_batch( |
| 724 | batch, |
| 725 | tokenize_de, |
| 726 | tokenize_en, |
| 727 | vocab_src, |
| 728 | vocab_tgt, |
| 729 | device, |
| 730 | max_padding=max_padding, |
| 731 | pad_id=vocab_src.get_stoi()["<blank>"], |
| 732 | ) |
| 733 | |
| 734 | train_iter, valid_iter, test_iter = Multi30k( |
| 735 | language_pair=("de", "en") |
| 736 | ) |
| 737 | |
| 738 | train_iter_map = to_map_style_dataset( |
| 739 | train_iter |
| 740 | ) # DistributedSampler needs a dataset len() |
| 741 | train_sampler = ( |
| 742 | DistributedSampler(train_iter_map) if is_distributed else None |
| 743 | ) |
| 744 | valid_iter_map = to_map_style_dataset(valid_iter) |
| 745 | valid_sampler = ( |
| 746 | DistributedSampler(valid_iter_map) if is_distributed else None |
| 747 | ) |
| 748 | |
| 749 | train_dataloader = DataLoader( |
| 750 | train_iter_map, |
| 751 | batch_size=batch_size, |
| 752 | shuffle=(train_sampler is None), |
| 753 | sampler=train_sampler, |
| 754 | collate_fn=collate_fn, |
| 755 | ) |
| 756 | valid_dataloader = DataLoader( |
| 757 | valid_iter_map, |
| 758 | batch_size=batch_size, |
| 759 | shuffle=(valid_sampler is None), |
| 760 | sampler=valid_sampler, |
| 761 | collate_fn=collate_fn, |
| 762 | ) |
| 763 | return train_dataloader, valid_dataloader |
| 764 | |
| 765 | |
| 766 | def train_worker( |
| 767 | gpu, |
| 768 | ngpus_per_node, |
| 769 | vocab_src, |
| 770 | vocab_tgt, |
| 771 | spacy_de, |
| 772 | spacy_en, |
| 773 | config, |
| 774 | is_distributed=False, |
| 775 | ): |
| 776 | print(f"Train worker process using GPU: {gpu} for training", flush=True) |
| 777 | torch.cuda.set_device(gpu) |
| 778 | |
| 779 | pad_idx = vocab_tgt["<blank>"] |
| 780 | d_model = 512 |
| 781 | model = make_model(len(vocab_src), len(vocab_tgt), N=6) |
| 782 | model.cuda(gpu) |
| 783 | module = model |
| 784 | is_main_process = True |
| 785 | if is_distributed: |
| 786 | dist.init_process_group( |
| 787 | "nccl", init_method="env://", rank=gpu, world_size=ngpus_per_node |
| 788 | ) |
| 789 | model = DDP(model, device_ids=[gpu]) |
| 790 | module = model.module |
| 791 | is_main_process = gpu == 0 |
| 792 | |
| 793 | criterion = LabelSmoothing( |
| 794 | size=len(vocab_tgt), padding_idx=pad_idx, smoothing=0.1 |
| 795 | ) |
| 796 | criterion.cuda(gpu) |
| 797 | |
| 798 | train_dataloader, valid_dataloader = create_dataloaders( |
| 799 | gpu, |
| 800 | vocab_src, |
| 801 | vocab_tgt, |
| 802 | spacy_de, |
| 803 | spacy_en, |
| 804 | batch_size=config["batch_size"] // ngpus_per_node, |
| 805 | max_padding=config["max_padding"], |
| 806 | is_distributed=is_distributed, |
| 807 | ) |
| 808 | |
| 809 | optimizer = torch.optim.Adam( |
| 810 | model.parameters(), lr=config["base_lr"], betas=(0.9, 0.98), eps=1e-9 |
| 811 | ) |
| 812 | lr_scheduler = LambdaLR( |
| 813 | optimizer=optimizer, |
| 814 | lr_lambda=lambda step: rate( |
| 815 | step, d_model, factor=1, warmup=config["warmup"] |
| 816 | ), |
| 817 | ) |
| 818 | train_state = TrainState() |
| 819 | |
| 820 | for epoch in range(config["num_epochs"]): |
| 821 | if is_distributed: |
| 822 | train_dataloader.sampler.set_epoch(epoch) |
| 823 | valid_dataloader.sampler.set_epoch(epoch) |
| 824 | |
| 825 | model.train() |
| 826 | print(f"[GPU{gpu}] Epoch {epoch} Training ====", flush=True) |
| 827 | _, train_state = run_epoch( |
| 828 | (Batch(b[0], b[1], pad_idx) for b in train_dataloader), |
| 829 | model, |
| 830 | SimpleLossCompute(module.generator, criterion), |
| 831 | optimizer, |
| 832 | lr_scheduler, |
| 833 | mode="train+log", |
| 834 | accum_iter=config["accum_iter"], |
| 835 | train_state=train_state, |
| 836 | ) |
| 837 | |
| 838 | GPUtil.showUtilization() |
| 839 | if is_main_process: |
| 840 | file_path = "%s%.2d.pt" % (config["file_prefix"], epoch) |
| 841 | torch.save(module.state_dict(), file_path) |
| 842 | torch.cuda.empty_cache() |
| 843 | |
| 844 | print(f"[GPU{gpu}] Epoch {epoch} Validation ====", flush=True) |
| 845 | model.eval() |
| 846 | sloss = run_epoch( |
| 847 | (Batch(b[0], b[1], pad_idx) for b in valid_dataloader), |
| 848 | model, |
| 849 | SimpleLossCompute(module.generator, criterion), |
| 850 | DummyOptimizer(), |
| 851 | DummyScheduler(), |
| 852 | mode="eval", |
| 853 | ) |
| 854 | print(sloss) |
| 855 | torch.cuda.empty_cache() |
| 856 | |
| 857 | if is_main_process: |
| 858 | file_path = "%sfinal.pt" % config["file_prefix"] |
| 859 | torch.save(module.state_dict(), file_path) |
| 860 | |
| 861 | |
| 862 | def train_distributed_model(vocab_src, vocab_tgt, spacy_de, spacy_en, config): |
| 863 | ngpus = torch.cuda.device_count() |
| 864 | os.environ["MASTER_ADDR"] = "localhost" |
| 865 | os.environ["MASTER_PORT"] = "12356" |
| 866 | print(f"Number of GPUs detected: {ngpus}") |
| 867 | print("Spawning training processes ...") |
| 868 | mp.spawn( |
| 869 | train_worker, |
| 870 | nprocs=ngpus, |
| 871 | args=(ngpus, vocab_src, vocab_tgt, spacy_de, spacy_en, config, True), |
| 872 | ) |
| 873 | |
| 874 | |
| 875 | def train_model(vocab_src, vocab_tgt, spacy_de, spacy_en, config): |
| 876 | if config["distributed"]: |
| 877 | train_distributed_model( |
| 878 | vocab_src, vocab_tgt, spacy_de, spacy_en, config |
| 879 | ) |
| 880 | else: |
| 881 | train_worker( |
| 882 | 0, 1, vocab_src, vocab_tgt, spacy_de, spacy_en, config, False |
| 883 | ) |
| 884 | |
| 885 | |
| 886 | def load_trained_model(vocab_src, vocab_tgt, spacy_de, spacy_en): |
| 887 | config = { |
| 888 | "batch_size": 32, |
| 889 | "distributed": False, |
| 890 | "num_epochs": 8, |
| 891 | "accum_iter": 10, |
| 892 | "base_lr": 1.0, |
| 893 | "max_padding": 72, |
| 894 | "warmup": 3000, |
| 895 | "file_prefix": "multi30k_model_", |
| 896 | } |
| 897 | model_path = "multi30k_model_final.pt" |
| 898 | if not exists(model_path): |
| 899 | train_model(vocab_src, vocab_tgt, spacy_de, spacy_en, config) |
| 900 | |
| 901 | model = make_model(len(vocab_src), len(vocab_tgt), N=6) |
| 902 | model.load_state_dict(torch.load("multi30k_model_final.pt")) |
| 903 | return model |
| 904 | |
| 905 | |
| 906 | def check_outputs( |
| 907 | valid_dataloader, |
| 908 | model, |
| 909 | vocab_src, |
| 910 | vocab_tgt, |
| 911 | n_examples=15, |
| 912 | pad_idx=2, |
| 913 | eos_string="</s>", |
| 914 | ): |
| 915 | results = [()] * n_examples |
| 916 | for idx in range(n_examples): |
| 917 | print("\nExample %d ========\n" % idx) |
| 918 | b = next(iter(valid_dataloader)) |
| 919 | rb = Batch(b[0], b[1], pad_idx) |
| 920 | greedy_decode(model, rb.src, rb.src_mask, 64, 0)[0] |
| 921 | |
| 922 | src_tokens = [ |
| 923 | vocab_src.get_itos()[x] for x in rb.src[0] if x != pad_idx |
| 924 | ] |
| 925 | tgt_tokens = [ |
| 926 | vocab_tgt.get_itos()[x] for x in rb.tgt[0] if x != pad_idx |
| 927 | ] |
| 928 | |
| 929 | print( |
| 930 | "Source Text (Input) : " |
| 931 | + " ".join(src_tokens).replace("\n", "") |
| 932 | ) |
| 933 | print( |
| 934 | "Target Text (Ground Truth) : " |
| 935 | + " ".join(tgt_tokens).replace("\n", "") |
| 936 | ) |
| 937 | model_out = greedy_decode(model, rb.src, rb.src_mask, 72, 0)[0] |
| 938 | model_txt = ( |
| 939 | " ".join( |
| 940 | [vocab_tgt.get_itos()[x] for x in model_out if x != pad_idx] |
| 941 | ).split(eos_string, 1)[0] |
| 942 | + eos_string |
| 943 | ) |
| 944 | print("Model Output : " + model_txt.replace("\n", "")) |
| 945 | results[idx] = (rb, src_tokens, tgt_tokens, model_out, model_txt) |
| 946 | return results |
| 947 | |
| 948 | |
| 949 | def mtx2df(m, max_row, max_col, row_tokens, col_tokens): |
| 950 | "convert a dense matrix to a data frame with row and column indices" |
| 951 | return pd.DataFrame( |
| 952 | [ |
| 953 | ( |
| 954 | r, |
| 955 | c, |
| 956 | float(m[r, c]), |
| 957 | "%.3d %s" |
| 958 | % (r, row_tokens[r] if len(row_tokens) > r else "<blank>"), |
| 959 | "%.3d %s" |
| 960 | % (c, col_tokens[c] if len(col_tokens) > c else "<blank>"), |
| 961 | ) |
| 962 | for r in range(m.shape[0]) |
| 963 | for c in range(m.shape[1]) |
| 964 | if r < max_row and c < max_col |
| 965 | ], |
| 966 | # if float(m[r,c]) != 0 and r < max_row and c < max_col], |
| 967 | columns=["row", "column", "value", "row_token", "col_token"], |
| 968 | ) |
| 969 | |
| 970 | |
| 971 | def attn_map(attn, layer, head, row_tokens, col_tokens, max_dim=30): |
| 972 | df = mtx2df( |
| 973 | attn[0, head].data, |
| 974 | max_dim, |
| 975 | max_dim, |
| 976 | row_tokens, |
| 977 | col_tokens, |
| 978 | ) |
| 979 | return ( |
| 980 | alt.Chart(data=df) |
| 981 | .mark_rect() |
| 982 | .encode( |
| 983 | x=alt.X("col_token", axis=alt.Axis(title="")), |
| 984 | y=alt.Y("row_token", axis=alt.Axis(title="")), |
| 985 | color="value", |
| 986 | tooltip=["row", "column", "value", "row_token", "col_token"], |
| 987 | ) |
| 988 | .properties(height=400, width=400) |
| 989 | .interactive() |
| 990 | ) |
| 991 | |
| 992 | |
| 993 | def get_encoder(model, layer): |
| 994 | return model.encoder.layers[layer].self_attn.attn |
| 995 | |
| 996 | |
| 997 | def get_decoder_self(model, layer): |
| 998 | return model.decoder.layers[layer].self_attn.attn |
| 999 | |
| 1000 | |
| 1001 | def get_decoder_src(model, layer): |
| 1002 | return model.decoder.layers[layer].src_attn.attn |
| 1003 | |
| 1004 | |
| 1005 | def visualize_layer(model, layer, getter_fn, ntokens, row_tokens, col_tokens): |
| 1006 | # ntokens = last_example[0].ntokens |
| 1007 | attn = getter_fn(model, layer) |
| 1008 | n_heads = attn.shape[1] |
| 1009 | charts = [ |
| 1010 | attn_map( |
| 1011 | attn, |
| 1012 | 0, |
| 1013 | h, |
| 1014 | row_tokens=row_tokens, |
| 1015 | col_tokens=col_tokens, |
| 1016 | max_dim=ntokens, |
| 1017 | ) |
| 1018 | for h in range(n_heads) |
| 1019 | ] |
| 1020 | assert n_heads == 8 |
| 1021 | return alt.vconcat( |
| 1022 | charts[0] |
| 1023 | # | charts[1] |
| 1024 | | charts[2] |
| 1025 | # | charts[3] |
| 1026 | | charts[4] |
| 1027 | # | charts[5] |
| 1028 | | charts[6] |
| 1029 | # | charts[7] |
| 1030 | # layer + 1 due to 0-indexing |
| 1031 | ).properties(title="Layer %d" % (layer + 1)) |