STEP 11 / 18

학습용 배치: 한 칸 밀기와 마스크

원본: 배치와 마스킹

여기서부터 2부, 학습이다. 첫 부품은 문장쌍 배치를 학습에 필요한 형태로 바꾸는 Batch 클래스다.

teacher forcing: 학습할 때는 디코더에 정답 문장을 통째로 넣고, 모든 위치에서 동시에 "다음 토큰"을 맞히게 한다. 그러려면 정답을 한 칸 밀어야 한다.

tgt = 정답에서 마지막 토큰을 뺀 것 → 디코더 입력
tgt_y = 정답에서 첫 토큰을 뺀 것 → 맞혀야 할 답

i번째 위치는 tgt[0..i]를 보고 tgt_y[i]를 맞힌다. 이때 i번째 위치가 tgt[i+1..]를 보면 답을 베끼는 셈이므로 미래 마스크가 필요하다. make_std_mask는 패딩 마스크와 subsequent_mask를 AND로 합친다.

ntokens는 패딩이 아닌 정답 토큰 수다. 손실을 이 값으로 나눠 토큰당 평균 손실을 만든다.

실행 결과에서 볼 것

tgt와 tgt_y가 한 칸 어긋나 있다.
tgt_mask는 아래 삼각 행렬인데, 패딩 열(마지막 열)이 추가로 0이다.

코드 변화

+22 −0 · 파일 전체 313줄

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

GitHub에서 보기 ↗

실행 결과

실행한 코드 demo.py 보기

1	import torch
2
3	from transformer import Batch
4
5	# 0 = <s>, 1 = </s>, 2 = <blank>(패딩)
6	src = torch.tensor([[0, 11, 12, 13, 1, 2]])
7	tgt = torch.tensor([[0, 21, 22, 1, 2, 2]])
8	b = Batch(src, tgt, pad=2)
9
10	print("src ", b.src.tolist())
11	print("src_mask ", b.src_mask.int().tolist(), " (패딩만 가림)")
12	print()
13	print("tgt ", tgt.tolist())
14	print("b.tgt ", b.tgt.tolist(), " ← 디코더 입력: 마지막 토큰을 뺀 것")
15	print("b.tgt_y ", b.tgt_y.tolist(), " ← 정답: 첫 토큰을 뺀 것 (한 칸 앞으로 민 것)")
16	print()
17	print("i번째 입력을 보고 i번째 정답을 맞힌다:")
18	for i, (a, y) in enumerate(zip(b.tgt[0].tolist(), b.tgt_y[0].tolist())):
19	print(f" 위치 {i}: {b.tgt[0, : i + 1].tolist()} 까지 보고 → {y}")
20	print()
21	print("tgt_mask = 미래 가리기 AND 패딩 가리기")
22	print(b.tgt_mask[0].int())
23	print("ntokens (패딩이 아닌 정답 수):", b.ntokens.item())

src       [[0, 11, 12, 13, 1, 2]]
src_mask  [[[1, 1, 1, 1, 1, 0]]]  (패딩만 가림)

tgt       [[0, 21, 22, 1, 2, 2]]
b.tgt     [[0, 21, 22, 1, 2]]  ← 디코더 입력: 마지막 토큰을 뺀 것
b.tgt_y   [[21, 22, 1, 2, 2]]  ← 정답: 첫 토큰을 뺀 것 (한 칸 앞으로 민 것)

i번째 입력을 보고 i번째 정답을 맞힌다:
  위치 0: [0] 까지 보고 → 21
  위치 1: [0, 21] 까지 보고 → 22
  위치 2: [0, 21, 22] 까지 보고 → 1
  위치 3: [0, 21, 22, 1] 까지 보고 → 2
  위치 4: [0, 21, 22, 1, 2] 까지 보고 → 2

tgt_mask = 미래 가리기 AND 패딩 가리기
tensor([[1, 0, 0, 0, 0],
        [1, 1, 0, 0, 0],
        [1, 1, 1, 0, 0],
        [1, 1, 1, 1, 0],
        [1, 1, 1, 1, 0]], dtype=torch.int32)
ntokens (패딩이 아닌 정답 수): 3


1	1		import copy
2	2		import math
3	3
4	4		import torch
5	5		import torch.nn as nn
6	6		from torch.nn.functional import log_softmax
7	7
8	8
9	9		def attention(query, key, value, mask=None, dropout=None):
10	10		"Compute 'Scaled Dot Product Attention'"
11	11		d_k = query.size(-1)
12	12		scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
13	13		if mask is not None:
14	14		scores = scores.masked_fill(mask == 0, -1e9)
15	15		p_attn = scores.softmax(dim=-1)
16	16		if dropout is not None:
17	17		p_attn = dropout(p_attn)
18	18		return torch.matmul(p_attn, value), p_attn
19	19
20	20
21	21		def subsequent_mask(size):
22	22		"Mask out subsequent positions."
23	23		attn_shape = (1, size, size)
24	24		subsequent_mask = torch.triu(torch.ones(attn_shape), diagonal=1).type(
25	25		torch.uint8
26	26		)
27	27		return subsequent_mask == 0
28	28
29	29
30	30		def clones(module, N):
31	31		"Produce N identical layers."
32	32		return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])
33	33
34	34
35	35		class MultiHeadedAttention(nn.Module):
36	36		def __init__(self, h, d_model, dropout=0.1):
37	37		"Take in model size and number of heads."
38	38		super(MultiHeadedAttention, self).__init__()
39	39		assert d_model % h == 0
40	40		# We assume d_v always equals d_k
41	41		self.d_k = d_model // h
42	42		self.h = h
43	43		self.linears = clones(nn.Linear(d_model, d_model), 4)
44	44		self.attn = None
45	45		self.dropout = nn.Dropout(p=dropout)
46	46
47	47		def forward(self, query, key, value, mask=None):
48	48		"Implements Figure 2"
49	49		if mask is not None:
50	50		# Same mask applied to all h heads.
51	51		mask = mask.unsqueeze(1)
52	52		nbatches = query.size(0)
53	53
54	54		# 1) Do all the linear projections in batch from d_model => h x d_k
55	55		query, key, value = [
56	56		lin(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2)
57	57		for lin, x in zip(self.linears, (query, key, value))
58	58		]
59	59
60	60		# 2) Apply attention on all the projected vectors in batch.
61	61		x, self.attn = attention(
62	62		query, key, value, mask=mask, dropout=self.dropout
63	63		)
64	64
65	65		# 3) "Concat" using a view and apply a final linear.
66	66		x = (
67	67		x.transpose(1, 2)
68	68		.contiguous()
69	69		.view(nbatches, -1, self.h * self.d_k)
70	70		)
71	71		del query
72	72		del key
73	73		del value
74	74		return self.linears[-1](x)
75	75
76	76
77	77		class PositionwiseFeedForward(nn.Module):
78	78		"Implements FFN equation."
79	79
80	80		def __init__(self, d_model, d_ff, dropout=0.1):
81	81		super(PositionwiseFeedForward, self).__init__()
82	82		self.w_1 = nn.Linear(d_model, d_ff)
83	83		self.w_2 = nn.Linear(d_ff, d_model)
84	84		self.dropout = nn.Dropout(dropout)
85	85
86	86		def forward(self, x):
87	87		return self.w_2(self.dropout(self.w_1(x).relu()))
88	88
89	89
90	90		class Embeddings(nn.Module):
91	91		def __init__(self, d_model, vocab):
92	92		super(Embeddings, self).__init__()
93	93		self.lut = nn.Embedding(vocab, d_model)
94	94		self.d_model = d_model
95	95
96	96		def forward(self, x):
97	97		return self.lut(x) * math.sqrt(self.d_model)
98	98
99	99
100	100		class PositionalEncoding(nn.Module):
101	101		"Implement the PE function."
102	102
103	103		def __init__(self, d_model, dropout, max_len=5000):
104	104		super(PositionalEncoding, self).__init__()
105	105		self.dropout = nn.Dropout(p=dropout)
106	106
107	107		# Compute the positional encodings once in log space.
108	108		pe = torch.zeros(max_len, d_model)
109	109		position = torch.arange(0, max_len).unsqueeze(1)
110	110		div_term = torch.exp(
111	111		torch.arange(0, d_model, 2) * -(math.log(10000.0) / d_model)
112	112		)
113	113		pe[:, 0::2] = torch.sin(position * div_term)
114	114		pe[:, 1::2] = torch.cos(position * div_term)
115	115		pe = pe.unsqueeze(0)
116	116		self.register_buffer("pe", pe)
117	117
118	118		def forward(self, x):
119	119		x = x + self.pe[:, : x.size(1)].requires_grad_(False)
120	120		return self.dropout(x)
121	121
122	122
123	123		class LayerNorm(nn.Module):
124	124		"Construct a layernorm module (See citation for details)."
125	125
126	126		def __init__(self, features, eps=1e-6):
127	127		super(LayerNorm, self).__init__()
128	128		self.a_2 = nn.Parameter(torch.ones(features))
129	129		self.b_2 = nn.Parameter(torch.zeros(features))
130	130		self.eps = eps
131	131
132	132		def forward(self, x):
133	133		mean = x.mean(-1, keepdim=True)
134	134		std = x.std(-1, keepdim=True)
135	135		return self.a_2 * (x - mean) / (std + self.eps) + self.b_2
136	136
137	137
138	138		class SublayerConnection(nn.Module):
139	139		"""
140	140		A residual connection followed by a layer norm.
141	141		Note for code simplicity the norm is first as opposed to last.
142	142		"""
143	143
144	144		def __init__(self, size, dropout):
145	145		super(SublayerConnection, self).__init__()
146	146		self.norm = LayerNorm(size)
147	147		self.dropout = nn.Dropout(dropout)
148	148
149	149		def forward(self, x, sublayer):
150	150		"Apply residual connection to any sublayer with the same size."
151	151		return x + self.dropout(sublayer(self.norm(x)))
152	152
153	153
154	154		class EncoderLayer(nn.Module):
155	155		"Encoder is made up of self-attn and feed forward (defined below)"
156	156
157	157		def __init__(self, size, self_attn, feed_forward, dropout):
158	158		super(EncoderLayer, self).__init__()
159	159		self.self_attn = self_attn
160	160		self.feed_forward = feed_forward
161	161		self.sublayer = clones(SublayerConnection(size, dropout), 2)
162	162		self.size = size
163	163
164	164		def forward(self, x, mask):
165	165		"Follow Figure 1 (left) for connections."
166	166		x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))
167	167		return self.sublayer[1](x, self.feed_forward)
168	168
169	169
170	170		class Encoder(nn.Module):
171	171		"Core encoder is a stack of N layers"
172	172
173	173		def __init__(self, layer, N):
174	174		super(Encoder, self).__init__()
175	175		self.layers = clones(layer, N)
176	176		self.norm = LayerNorm(layer.size)
177	177
178	178		def forward(self, x, mask):
179	179		"Pass the input (and mask) through each layer in turn."
180	180		for layer in self.layers:
181	181		x = layer(x, mask)
182	182		return self.norm(x)
183	183
184	184
185	185		class DecoderLayer(nn.Module):
186	186		"Decoder is made of self-attn, src-attn, and feed forward (defined below)"
187	187
188	188		def __init__(self, size, self_attn, src_attn, feed_forward, dropout):
189	189		super(DecoderLayer, self).__init__()
190	190		self.size = size
191	191		self.self_attn = self_attn
192	192		self.src_attn = src_attn
193	193		self.feed_forward = feed_forward
194	194		self.sublayer = clones(SublayerConnection(size, dropout), 3)
195	195
196	196		def forward(self, x, memory, src_mask, tgt_mask):
197	197		"Follow Figure 1 (right) for connections."
198	198		m = memory
199	199		x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, tgt_mask))
200	200		x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, src_mask))
201	201		return self.sublayer[2](x, self.feed_forward)
202	202
203	203
204	204		class Decoder(nn.Module):
205	205		"Generic N layer decoder with masking."
206	206
207	207		def __init__(self, layer, N):
208	208		super(Decoder, self).__init__()
209	209		self.layers = clones(layer, N)
210	210		self.norm = LayerNorm(layer.size)
211	211
212	212		def forward(self, x, memory, src_mask, tgt_mask):
213	213		for layer in self.layers:
214	214		x = layer(x, memory, src_mask, tgt_mask)
215	215		return self.norm(x)
216	216
217	217
218	218		class EncoderDecoder(nn.Module):
219	219		"""
220	220		A standard Encoder-Decoder architecture. Base for this and many
221	221		other models.
222	222		"""
223	223
224	224		def __init__(self, encoder, decoder, src_embed, tgt_embed, generator):
225	225		super(EncoderDecoder, self).__init__()
226	226		self.encoder = encoder
227	227		self.decoder = decoder
228	228		self.src_embed = src_embed
229	229		self.tgt_embed = tgt_embed
230	230		self.generator = generator
231	231
232	232		def forward(self, src, tgt, src_mask, tgt_mask):
233	233		"Take in and process masked src and target sequences."
234	234		return self.decode(self.encode(src, src_mask), src_mask, tgt, tgt_mask)
235	235
236	236		def encode(self, src, src_mask):
237	237		return self.encoder(self.src_embed(src), src_mask)
238	238
239	239		def decode(self, memory, src_mask, tgt, tgt_mask):
240	240		return self.decoder(self.tgt_embed(tgt), memory, src_mask, tgt_mask)
241	241
242	242
243	243		class Generator(nn.Module):
244	244		"Define standard linear + softmax generation step."
245	245
246	246		def __init__(self, d_model, vocab):
247	247		super(Generator, self).__init__()
248	248		self.proj = nn.Linear(d_model, vocab)
249	249
250	250		def forward(self, x):
251	251		return log_softmax(self.proj(x), dim=-1)
252	252
253	253
254	254		def make_model(
255	255		src_vocab, tgt_vocab, N=6, d_model=512, d_ff=2048, h=8, dropout=0.1
256	256		):
257	257		"Helper: Construct a model from hyperparameters."
258	258		c = copy.deepcopy
259	259		attn = MultiHeadedAttention(h, d_model)
260	260		ff = PositionwiseFeedForward(d_model, d_ff, dropout)
261	261		position = PositionalEncoding(d_model, dropout)
262	262		model = EncoderDecoder(
263	263		Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N),
264	264		Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N),
265	265		nn.Sequential(Embeddings(d_model, src_vocab), c(position)),
266	266		nn.Sequential(Embeddings(d_model, tgt_vocab), c(position)),
267	267		Generator(d_model, tgt_vocab),
268	268		)
269	269
270	270		# This was important from their code.
271	271		# Initialize parameters with Glorot / fan_avg.
272	272		for p in model.parameters():
273	273		if p.dim() > 1:
274	274		nn.init.xavier_uniform_(p)
275	275		return model
276	276
277	277
278	278		def greedy_decode(model, src, src_mask, max_len, start_symbol):
279	279		memory = model.encode(src, src_mask)
280	280		ys = torch.zeros(1, 1).fill_(start_symbol).type_as(src.data)
281	281		for i in range(max_len - 1):
282	282		out = model.decode(
283	283		memory, src_mask, ys, subsequent_mask(ys.size(1)).type_as(src.data)
284	284		)
285	285		prob = model.generator(out[:, -1])
286	286		_, next_word = torch.max(prob, dim=1)
287	287		next_word = next_word.data[0]
288	288		ys = torch.cat(
289	289		[ys, torch.zeros(1, 1).type_as(src.data).fill_(next_word)], dim=1
290	290		)
291	291		return ys
	292	+
	293	+
	294	+	class Batch:
	295	+	"""Object for holding a batch of data with mask during training."""
	296	+
	297	+	def __init__(self, src, tgt=None, pad=2): # 2 = <blank>
	298	+	self.src = src
	299	+	self.src_mask = (src != pad).unsqueeze(-2)
	300	+	if tgt is not None:
	301	+	self.tgt = tgt[:, :-1]
	302	+	self.tgt_y = tgt[:, 1:]
	303	+	self.tgt_mask = self.make_std_mask(self.tgt, pad)
	304	+	self.ntokens = (self.tgt_y != pad).data.sum()
	305	+
	306	+	@staticmethod
	307	+	def make_std_mask(tgt, pad):
	308	+	"Create a mask to hide padding and future words."
	309	+	tgt_mask = (tgt != pad).unsqueeze(-2)
	310	+	tgt_mask = tgt_mask & subsequent_mask(tgt.size(-1)).type_as(
	311	+	tgt_mask.data
	312	+	)
	313	+	return tgt_mask