STEP 15 / 18

진짜 데이터: 토큰화와 사전

원본: 3부: 실전 예제 · 데이터 불러오기

3부에서는 Multi30k 독일어→영어 번역을 학습한다. 이미지 설명 문장 약 3만 쌍이라 논문의 WMT(450만 쌍)보다 훨씬 작지만, 전체 흐름을 보기에는 충분하다.

load_tokenizers: spaCy의 독일어·영어 모델로 문장을 단어 단위로 자른다.
build_vocabulary: train+valid+test 전체에서 두 번 이상 나온 단어로 사전을 만든다. 맨 앞 네 칸은 특수 토큰이다: <s>(0, 시작), </s>(1, 끝), <blank>(2, 패딩), <unk>(3, 모르는 단어). 그래서 앞 단계 코드들이 pad=2, start_symbol=0을 기본값으로 쓴다.
load_vocab: 만든 사전을 vocab.pt에 저장해 두고 다음부터는 불러온다.

원본은 torchtext의 Multi30k, build_vocab_from_iterator, to_map_style_dataset을 쓴다. torchtext는 개발이 중단돼 최신 torch에서 import조차 되지 않는다. 그래서 같은 이름·같은 동작의 대체품을 파일 안에 넣었다(diff의 # torchtext 대신 부분). 데이터는 같은 분할을 Hugging Face Hub의 bentrevett/multi30k에서 받는다. torch 2.6부터 torch.load의 기본값이 바뀌어서 사전을 불러올 때 weights_only=False가 필요하다.

실행 결과에서 볼 것

29,000 / 1,014 / 1,000 문장쌍.
사전 크기: 독일어 8,316, 영어 6,384. 원본이 torchtext로 만든 것과 같은 순서다.
사전은 특수 토큰 4개 다음에 가장 흔한 토큰부터 온다. 독일어는 ., Ein, einem, 영어는 a, ., A 순서다. 대소문자를 구분하므로 a와 A가 다른 단어다.

코드 변화

+129 −0 · 파일 전체 632줄

이전 단계의 transformer.py와 비교했다. 초록 줄이 새로 생긴 줄, 빨간 줄이 사라진 줄이다.

GitHub에서 보기 ↗

실행 결과

실행한 코드 demo.py 보기

1	from transformer import Multi30k, load_tokenizers, load_vocab, tokenize
2
3	spacy_de, spacy_en = load_tokenizers()
4
5	train, val, test = Multi30k(language_pair=("de", "en"))
6	print(f"Multi30k: train {len(train):,} / valid {len(val):,} / test {len(test):,} 문장쌍")
7	print()
8	de, en = train[0]
9	print("첫 문장쌍")
10	print(" de:", de)
11	print(" en:", en)
12	print(" de 토큰:", tokenize(de, spacy_de))
13	print(" en 토큰:", tokenize(en, spacy_en))
14	print()
15
16	vocab_src, vocab_tgt = load_vocab(spacy_de, spacy_en)
17	print()
18	print("독일어 사전 앞 20개:", vocab_src.get_itos()[:20])
19	print("영어 사전 앞 20개 :", vocab_tgt.get_itos()[:20])
20	print()
21	tokens = tokenize(en, spacy_en)
22	print("영어 문장 → id:", vocab_tgt(tokens))
23	print("사전에 없는 단어 'Transformerology' →", vocab_tgt["Transformerology"], "(<unk>)")

Multi30k: train 29,000 / valid 1,014 / test 1,000 문장쌍

첫 문장쌍
  de: Zwei junge weiße Männer sind im Freien in der Nähe vieler Büsche.
  en: Two young, White males are outside near many bushes.
  de 토큰: ['Zwei', 'junge', 'weiße', 'Männer', 'sind', 'im', 'Freien', 'in', 'der', 'Nähe', 'vieler', 'Büsche', '.']
  en 토큰: ['Two', 'young', ',', 'White', 'males', 'are', 'outside', 'near', 'many', 'bushes', '.']

Building German Vocabulary ...
Building English Vocabulary ...
Finished.
Vocabulary sizes:
8316
6384

독일어 사전 앞 20개: ['<s>', '</s>', '<blank>', '<unk>', '.', 'Ein', 'einem', 'in', 'und', ',', 'mit', 'auf', 'Mann', 'einer', 'Eine', 'ein', 'der', 'Frau', 'eine', 'die']
영어 사전 앞 20개 : ['<s>', '</s>', '<blank>', '<unk>', 'a', '.', 'A', 'in', 'the', 'on', 'is', 'and', 'man', 'of', 'with', ',', 'woman', 'are', 'to', 'Two']

영어 문장 → id: [19, 25, 15, 1138, 822, 17, 58, 84, 335, 1359, 5]
사전에 없는 단어 'Transformerology' → 3 (<unk>)

	1	+	import os
	2	+	from os.path import exists
1	3		import copy
2	4		import math
3	5		import time
	6	+	from collections import Counter
4	7
	8	+	import spacy
5	9		import torch
6	10		import torch.nn as nn
7	11		from torch.nn.functional import log_softmax

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