transformer_model.py 15.8 KB
Newer Older
Y
Yu Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from functools import partial
import numpy as np

18
import paddle
Y
Yu Yang 已提交
19 20 21 22 23
import paddle.fluid as fluid
import paddle.fluid.layers as layers

pos_enc_param_names = (
    "src_pos_enc_table",
24 25
    "trg_pos_enc_table",
)
Y
Yu Yang 已提交
26

W
Wu Yi 已提交
27
batch_size = 2
Y
Yu Yang 已提交
28 29 30 31 32 33


def position_encoding_init(n_position, d_pos_vec):
    """
    Generate the initial values for the sinusoid position encoding table.
    """
34 35 36 37 38 39 40 41 42 43 44
    position_enc = np.array(
        [
            [
                pos / np.power(10000, 2 * (j // 2) / d_pos_vec)
                for j in range(d_pos_vec)
            ]
            if pos != 0
            else np.zeros(d_pos_vec)
            for pos in range(n_position)
        ]
    )
Y
Yu Yang 已提交
45 46 47 48 49
    position_enc[1:, 0::2] = np.sin(position_enc[1:, 0::2])  # dim 2i
    position_enc[1:, 1::2] = np.cos(position_enc[1:, 1::2])  # dim 2i+1
    return position_enc.astype("float32")


50 51 52 53 54 55 56 57 58 59 60
def multi_head_attention(
    queries,
    keys,
    values,
    attn_bias,
    d_key,
    d_value,
    d_model,
    n_head=1,
    dropout_rate=0.0,
):
Y
Yu Yang 已提交
61 62 63 64 65 66 67
    """
    Multi-Head Attention. Note that attn_bias is added to the logit before
    computing softmax activiation to mask certain selected positions so that
    they will not considered in attention weights.
    """
    if not (len(queries.shape) == len(keys.shape) == len(values.shape) == 3):
        raise ValueError(
68 69
            "Inputs: queries, keys and values should all be 3-D tensors."
        )
Y
Yu Yang 已提交
70 71 72 73 74

    def __compute_qkv(queries, keys, values, n_head, d_key, d_value):
        """
        Add linear projection to queries, keys, and values.
        """
75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103
        q = layers.fc(
            input=queries,
            size=d_key * n_head,
            param_attr=fluid.initializer.Xavier(
                uniform=False, fan_in=d_model * d_key, fan_out=n_head * d_key
            ),
            bias_attr=False,
            num_flatten_dims=2,
        )
        k = layers.fc(
            input=keys,
            size=d_key * n_head,
            param_attr=fluid.initializer.Xavier(
                uniform=False, fan_in=d_model * d_key, fan_out=n_head * d_key
            ),
            bias_attr=False,
            num_flatten_dims=2,
        )
        v = layers.fc(
            input=values,
            size=d_value * n_head,
            param_attr=fluid.initializer.Xavier(
                uniform=False,
                fan_in=d_model * d_value,
                fan_out=n_head * d_value,
            ),
            bias_attr=False,
            num_flatten_dims=2,
        )
Y
Yu Yang 已提交
104 105 106 107
        return q, k, v

    def __split_heads(x, n_head):
        """
T
tianshuo78520a 已提交
108
        Reshape the last dimension of input tensor x so that it becomes two
Y
Yu Yang 已提交
109 110 111 112 113 114 115 116 117
        dimensions and then transpose. Specifically, input a tensor with shape
        [bs, max_sequence_length, n_head * hidden_dim] then output a tensor
        with shape [bs, n_head, max_sequence_length, hidden_dim].
        """
        if n_head == 1:
            return x

        hidden_size = x.shape[-1]
        # FIXME(guosheng): Decouple the program desc with batch_size.
118
        reshaped = paddle.reshape(
119 120
            x=x, shape=[batch_size, -1, n_head, hidden_size // n_head]
        )
Y
Yu Yang 已提交
121

T
tianshuo78520a 已提交
122
        # permute the dimensions into:
Y
Yu Yang 已提交
123 124 125 126 127
        # [batch_size, n_head, max_sequence_len, hidden_size_per_head]
        return layers.transpose(x=reshaped, perm=[0, 2, 1, 3])

    def __combine_heads(x):
        """
T
tianshuo78520a 已提交
128
        Transpose and then reshape the last two dimensions of input tensor x
Y
Yu Yang 已提交
129 130
        so that it becomes one dimension, which is reverse to __split_heads.
        """
131 132
        if len(x.shape) == 3:
            return x
Y
Yu Yang 已提交
133 134 135 136 137
        if len(x.shape) != 4:
            raise ValueError("Input(x) should be a 4-D Tensor.")

        trans_x = layers.transpose(x, perm=[0, 2, 1, 3])
        # FIXME(guosheng): Decouple the program desc with batch_size.
138
        return paddle.reshape(
Y
Yu Yang 已提交
139
            x=trans_x,
140
            shape=list(
141 142 143
                map(int, [batch_size, -1, trans_x.shape[2] * trans_x.shape[3]])
            ),
        )
Y
Yu Yang 已提交
144 145 146 147 148 149 150 151 152 153 154

    def scaled_dot_product_attention(q, k, v, attn_bias, d_model, dropout_rate):
        """
        Scaled Dot-Product Attention
        """

        # FIXME(guosheng): Optimize the shape in reshape_op or softmax_op.

        # The current implementation of softmax_op only supports 2D tensor,
        # consequently it cannot be directly used here.
        # If to use the reshape_op, Besides, the shape of product inferred in
翟飞跃 已提交
155
        # compile-time is not the actual shape in run-time. It can't be used
Y
Yu Yang 已提交
156 157 158 159
        # to set the attribute of reshape_op.
        # So, here define the softmax for temporary solution.

        def __softmax(x, eps=1e-9):
160
            exp_out = paddle.exp(x=x)
Y
Yu Yang 已提交
161 162 163 164 165 166 167
            sum_out = layers.reduce_sum(exp_out, dim=-1, keep_dim=False)
            return layers.elementwise_div(x=exp_out, y=sum_out, axis=0)

        scaled_q = layers.scale(x=q, scale=d_model**-0.5)
        product = layers.matmul(x=scaled_q, y=k, transpose_y=True)
        weights = __softmax(layers.elementwise_add(x=product, y=attn_bias))
        if dropout_rate:
168 169 170
            weights = layers.dropout(
                weights, dropout_prob=dropout_rate, is_test=False
            )
Y
Yu Yang 已提交
171 172 173 174 175 176 177 178 179
        out = layers.matmul(weights, v)
        return out

    q, k, v = __compute_qkv(queries, keys, values, n_head, d_key, d_value)

    q = __split_heads(q, n_head)
    k = __split_heads(k, n_head)
    v = __split_heads(v, n_head)

180 181 182
    ctx_multiheads = scaled_dot_product_attention(
        q, k, v, attn_bias, d_model, dropout_rate
    )
Y
Yu Yang 已提交
183 184 185 186

    out = __combine_heads(ctx_multiheads)

    # Project back to the model size.
187 188 189 190 191 192 193
    proj_out = layers.fc(
        input=out,
        size=d_model,
        param_attr=fluid.initializer.Xavier(uniform=False),
        bias_attr=False,
        num_flatten_dims=2,
    )
Y
Yu Yang 已提交
194 195 196 197 198 199 200 201 202
    return proj_out


def positionwise_feed_forward(x, d_inner_hid, d_hid):
    """
    Position-wise Feed-Forward Networks.
    This module consists of two linear transformations with a ReLU activation
    in between, which is applied to each position separately and identically.
    """
203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219
    hidden = layers.fc(
        input=x,
        size=d_inner_hid,
        num_flatten_dims=2,
        param_attr=fluid.initializer.Uniform(
            low=-(d_hid**-0.5), high=(d_hid**-0.5)
        ),
        act="relu",
    )
    out = layers.fc(
        input=hidden,
        size=d_hid,
        num_flatten_dims=2,
        param_attr=fluid.initializer.Uniform(
            low=-(d_inner_hid**-0.5), high=(d_inner_hid**-0.5)
        ),
    )
Y
Yu Yang 已提交
220 221 222
    return out


223
def pre_post_process_layer(prev_out, out, process_cmd, dropout=0.0):
Y
Yu Yang 已提交
224 225 226 227 228 229 230 231 232 233 234
    """
    Add residual connection, layer normalization and droput to the out tensor
    optionally according to the value of process_cmd.

    This will be used before or after multi-head attention and position-wise
    feed-forward networks.
    """
    for cmd in process_cmd:
        if cmd == "a":  # add residual connection
            out = out + prev_out if prev_out else out
        elif cmd == "n":  # add layer normalization
235 236 237 238 239 240
            out = layers.layer_norm(
                out,
                begin_norm_axis=len(out.shape) - 1,
                param_attr=fluid.initializer.Constant(1.0),
                bias_attr=fluid.initializer.Constant(0.0),
            )
Y
Yu Yang 已提交
241 242 243 244 245 246 247 248 249 250
        elif cmd == "d":  # add dropout
            if dropout:
                out = layers.dropout(out, dropout_prob=dropout, is_test=False)
    return out


pre_process_layer = partial(pre_post_process_layer, None)
post_process_layer = pre_post_process_layer


251 252 253 254 255 256 257 258 259 260 261
def prepare_encoder(
    src_word,
    src_pos,
    src_vocab_size,
    src_emb_dim,
    src_pad_idx,
    src_max_len,
    dropout=0.0,
    pos_pad_idx=0,
    pos_enc_param_name=None,
):
Y
Yu Yang 已提交
262 263 264 265 266 267
    """Add word embeddings and position encodings.
    The output tensor has a shape of:
    [batch_size, max_src_length_in_batch, d_model].

    This module is used at the bottom of the encoder stacks.
    """
268 269 270 271 272 273
    src_word_emb = layers.embedding(
        src_word,
        size=[src_vocab_size, src_emb_dim],
        padding_idx=src_pad_idx,
        param_attr=fluid.initializer.Normal(0.0, 1.0),
    )
Y
Yu Yang 已提交
274 275 276 277
    src_pos_enc = layers.embedding(
        src_pos,
        size=[src_max_len, src_emb_dim],
        padding_idx=pos_pad_idx,
278 279
        param_attr=fluid.ParamAttr(name=pos_enc_param_name, trainable=False),
    )
C
chengduo 已提交
280
    src_pos_enc.stop_gradient = True
Y
Yu Yang 已提交
281 282 283
    enc_input = src_word_emb + src_pos_enc

    # FIXME(guosheng): Decouple the program desc with batch_size.
284
    enc_input = paddle.reshape(x=enc_input, shape=[batch_size, -1, src_emb_dim])
285 286 287 288 289
    return (
        layers.dropout(enc_input, dropout_prob=dropout, is_test=False)
        if dropout
        else enc_input
    )
Y
Yu Yang 已提交
290 291


292 293 294 295 296 297
prepare_encoder = partial(
    prepare_encoder, pos_enc_param_name=pos_enc_param_names[0]
)
prepare_decoder = partial(
    prepare_encoder, pos_enc_param_name=pos_enc_param_names[1]
)
Y
Yu Yang 已提交
298 299


300 301 302 303 304 305 306 307 308 309
def encoder_layer(
    enc_input,
    attn_bias,
    n_head,
    d_key,
    d_value,
    d_model,
    d_inner_hid,
    dropout_rate=0.0,
):
Y
Yu Yang 已提交
310 311 312 313 314 315 316
    """The encoder layers that can be stacked to form a deep encoder.

    This module consits of a multi-head (self) attention followed by
    position-wise feed-forward networks and both the two components companied
    with the post_process_layer to add residual connection, layer normalization
    and droput.
    """
317 318 319 320 321 322 323 324 325 326 327 328 329 330
    attn_output = multi_head_attention(
        enc_input,
        enc_input,
        enc_input,
        attn_bias,
        d_key,
        d_value,
        d_model,
        n_head,
        dropout_rate,
    )
    attn_output = post_process_layer(
        enc_input, attn_output, "dan", dropout_rate
    )
Y
Yu Yang 已提交
331 332 333 334
    ffd_output = positionwise_feed_forward(attn_output, d_inner_hid, d_model)
    return post_process_layer(attn_output, ffd_output, "dan", dropout_rate)


335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352
def encoder(
    enc_input,
    attn_bias,
    n_layer,
    n_head,
    d_key,
    d_value,
    d_model,
    d_inner_hid,
    dropout_rate=0.0,
):
    """
    The encoder is composed of a stack of identical layers returned by calling
    encoder_layer.
    """
    for i in range(n_layer):
        enc_output = encoder_layer(
            enc_input,
Y
Yu Yang 已提交
353 354 355 356 357 358
            attn_bias,
            n_head,
            d_key,
            d_value,
            d_model,
            d_inner_hid,
359 360
            dropout_rate,
        )
Y
Yu Yang 已提交
361 362 363 364
        enc_input = enc_output
    return enc_output


365 366 367 368 369 370 371 372 373 374 375 376 377
def decoder_layer(
    dec_input,
    enc_output,
    slf_attn_bias,
    dec_enc_attn_bias,
    n_head,
    d_key,
    d_value,
    d_model,
    d_inner_hid,
    dropout_rate=0.0,
):
    """The layer to be stacked in decoder part.
Y
Yu Yang 已提交
378 379 380 381 382 383 384 385 386 387 388 389 390

    The structure of this module is similar to that in the encoder part except
    a multi-head attention is added to implement encoder-decoder attention.
    """
    slf_attn_output = multi_head_attention(
        dec_input,
        dec_input,
        dec_input,
        slf_attn_bias,
        d_key,
        d_value,
        d_model,
        n_head,
391 392
        dropout_rate,
    )
Y
Yu Yang 已提交
393 394 395 396
    slf_attn_output = post_process_layer(
        dec_input,
        slf_attn_output,
        "dan",  # residual connection + dropout + layer normalization
397 398
        dropout_rate,
    )
Y
Yu Yang 已提交
399 400 401 402 403 404 405 406 407
    enc_attn_output = multi_head_attention(
        slf_attn_output,
        enc_output,
        enc_output,
        dec_enc_attn_bias,
        d_key,
        d_value,
        d_model,
        n_head,
408 409
        dropout_rate,
    )
Y
Yu Yang 已提交
410 411 412 413
    enc_attn_output = post_process_layer(
        slf_attn_output,
        enc_attn_output,
        "dan",  # residual connection + dropout + layer normalization
414 415
        dropout_rate,
    )
Y
Yu Yang 已提交
416 417 418
    ffd_output = positionwise_feed_forward(
        enc_attn_output,
        d_inner_hid,
419 420
        d_model,
    )
Y
Yu Yang 已提交
421 422 423 424
    dec_output = post_process_layer(
        enc_attn_output,
        ffd_output,
        "dan",  # residual connection + dropout + layer normalization
425 426
        dropout_rate,
    )
Y
Yu Yang 已提交
427 428 429
    return dec_output


430 431 432 433 434 435 436 437 438 439 440 441 442
def decoder(
    dec_input,
    enc_output,
    dec_slf_attn_bias,
    dec_enc_attn_bias,
    n_layer,
    n_head,
    d_key,
    d_value,
    d_model,
    d_inner_hid,
    dropout_rate=0.0,
):
Y
Yu Yang 已提交
443 444 445 446 447 448 449 450 451 452 453 454 455 456
    """
    The decoder is composed of a stack of identical decoder_layer layers.
    """
    for i in range(n_layer):
        dec_output = decoder_layer(
            dec_input,
            enc_output,
            dec_slf_attn_bias,
            dec_enc_attn_bias,
            n_head,
            d_key,
            d_value,
            d_model,
            d_inner_hid,
457 458
            dropout_rate,
        )
Y
Yu Yang 已提交
459 460 461 462
        dec_input = dec_output
    return dec_output


463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502
def build_inputs(max_length, n_head):
    names = [
        'src_word',
        'src_pos',
        'trg_word',
        'trg_pos',
        'src_slf_attn_bias',
        'trg_slf_attn_bias',
        'trg_src_attn_bias',
        'gold',
        'weights',
    ]

    shapes = [
        [batch_size * max_length, 1],
        [batch_size * max_length, 1],
        [batch_size * max_length, 1],
        [batch_size * max_length, 1],
        [batch_size, n_head, max_length, max_length],
        [batch_size, n_head, max_length, max_length],
        [batch_size, n_head, max_length, max_length],
        [batch_size * max_length, 1],
        [batch_size * max_length, 1],
    ]

    dtypes = [
        'int64',
        'int64',
        'int64',
        'int64',
        'float32',
        'float32',
        'float32',
        'int64',
        'float32',
    ]

    all_inputs = []
    for name, shape, dtype in zip(names, shapes, dtypes):
        all_inputs.append(
503 504 505 506
            fluid.layers.data(
                name=name, shape=shape, dtype=dtype, append_batch_size=False
            )
        )
507 508 509
    return all_inputs


Y
Yu Yang 已提交
510
def transformer(
511 512 513 514 515 516 517 518 519 520 521 522 523 524
    src_vocab_size,
    trg_vocab_size,
    max_length,
    n_layer,
    n_head,
    d_key,
    d_value,
    d_model,
    d_inner_hid,
    dropout_rate,
    src_pad_idx,
    trg_pad_idx,
    pos_pad_idx,
):
525

526 527 528 529 530 531 532 533 534 535 536
    (
        src_word,
        src_pos,
        trg_word,
        trg_pos,
        src_slf_attn_bias,
        trg_slf_attn_bias,
        trg_src_attn_bias,
        gold,
        weights,
    ) = build_inputs(max_length, n_head)
Y
Yu Yang 已提交
537 538 539 540 541 542 543 544

    enc_input = prepare_encoder(
        src_word,
        src_pos,
        src_vocab_size,
        d_model,
        src_pad_idx,
        max_length,
545 546
        dropout_rate,
    )
Y
Yu Yang 已提交
547 548 549 550 551 552 553 554 555
    enc_output = encoder(
        enc_input,
        src_slf_attn_bias,
        n_layer,
        n_head,
        d_key,
        d_value,
        d_model,
        d_inner_hid,
556 557
        dropout_rate,
    )
Y
Yu Yang 已提交
558 559 560 561 562 563 564 565

    dec_input = prepare_decoder(
        trg_word,
        trg_pos,
        trg_vocab_size,
        d_model,
        trg_pad_idx,
        max_length,
566 567
        dropout_rate,
    )
Y
Yu Yang 已提交
568 569 570 571 572 573 574 575 576 577 578
    dec_output = decoder(
        dec_input,
        enc_output,
        trg_slf_attn_bias,
        trg_src_attn_bias,
        n_layer,
        n_head,
        d_key,
        d_value,
        d_model,
        d_inner_hid,
579 580
        dropout_rate,
    )
Y
Yu Yang 已提交
581 582 583

    # TODO(guosheng): Share the weight matrix between the embedding layers and
    # the pre-softmax linear transformation.
584
    predict = paddle.reshape(
585 586 587 588 589 590 591 592 593
        x=layers.fc(
            input=dec_output,
            size=trg_vocab_size,
            param_attr=fluid.initializer.Xavier(uniform=False),
            bias_attr=False,
            num_flatten_dims=2,
        ),
        shape=[-1, trg_vocab_size],
    )
594
    predict = paddle.nn.functional.softmax(predict)
Y
Yu Yang 已提交
595 596 597 598

    cost = layers.cross_entropy(input=predict, label=gold)
    weighted_cost = cost * weights
    return layers.reduce_sum(weighted_cost)