File size: 24,077 Bytes
b1c0f8d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
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
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
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
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
#!/usr/bin/env python
# coding=utf-8
# Copyright 2022 Vladislav Lialin and Namrata Shivagunde 
#
# 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.
import os
import json
from collections import namedtuple

import torch
import torch.nn as nn
import torch.nn.functional as F

from transformer_mt.modeling_attention import MultiHeadAttention
from transformer_mt.utils import pad


Hypothesis = namedtuple("Hypothesis", ["value", "score"])


class TransformerEncoderLayer(nn.Module):
    def __init__(self, hidden, num_heads, fcn_hidden, dropout=0.0, causal=False):
        super().__init__()

        self.self_attention = MultiHeadAttention(
            input_size=hidden,
            hidden=hidden,
            num_heads=num_heads,
            causal=causal,
        )
        self.att_layer_norm = nn.LayerNorm(hidden)

        self.fcn = nn.Sequential(
            nn.Linear(hidden, fcn_hidden),
            nn.ReLU(),
            nn.Linear(fcn_hidden, hidden),
        )
        self.fcn_layer_norm = nn.LayerNorm(hidden)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x, key_padding_mask=None):
        """Self-Attention -> residual -> LayerNorm -> FCN -> residual -> LayerNorm
        
        Args:
            x: FloatTensor[batch_size, seq_len, input_size]
        
        Returns:
            FloatTensor[batch_size, seq_len, hidden]
        """
#         print('calling encode', key_padding_mask.shape)
        residual = x
        x = self.self_attention(x, key_padding_mask=key_padding_mask)
        x = self.att_layer_norm(x + residual)

        residual = x
        x = self.fcn(x)
        x = self.dropout(x)
        x = self.fcn_layer_norm(x + residual)
        
       
        return x


class TransformerDecoderLayer(nn.Module):
    def __init__(self, hidden, num_heads, fcn_hidden, dropout=0.0):
        super().__init__()

        # Task 2.1 (1 point)
        # Create layers needed for Transformer Decoder Layer
        # 1. Create self.self_attention layer using MultiHeadAttention
        # 2. Create self.cross_attention layer using MultiHeadAttention
        # 2a. Which one of self_attention or cross_attention should have causal=True? Set it there.
        # 3. Create self.att_layer_norm, self.cross_att_layer_norm, and self.fcn_layer_norm layers using LayerNorm
        # 4. Create self.fcn network using nn.Sequential, nn.ReLU and nn.Linear
        # 5. Create self.dropout layer using nn.Dropout
        # YOUR CODE STARTS HERE  (our implementation is about 5-8 lines) 

        self.self_attention = MultiHeadAttention(
            input_size=hidden,
            hidden=hidden,
            num_heads=num_heads,
            causal=True,
        )
        
        self.cross_attention = MultiHeadAttention(
            input_size=hidden,
            hidden=hidden,
            num_heads=num_heads,
            causal=False,
        )

        self.self_att_layer_norm = nn.LayerNorm(hidden)
        self.cross_att_layer_norm = nn.LayerNorm(hidden)

        self.fcn = nn.Sequential(
            nn.Linear(hidden, fcn_hidden),
            nn.ReLU(),
            nn.Linear(fcn_hidden, hidden),
        )
        self.fcn_layer_norm = nn.LayerNorm(hidden)
        self.dropout = nn.Dropout(dropout)
        
        # YOUR CODE ENDS HERE 
    
    def forward(self, decoder_hidden_states, encoder_hidden_states, key_padding_mask=None):
        """Transformer Decoder Layer

        Args:
            decoder_hidden_states: FloatTensor[batch_size, query_seq_len, hidden]
            encoder_hidden_states: FloatTensor[batch_size, kv_seq_len, hidden]
            key_padding_mask: ByteTensor[batch_size, kv_seq_len] with 1 for padded tokens and 0 for regular tokens

        Returns:
            FloatTensor[batch_size, query_seq_len, hidden]
        """

        # Task 2.2 (1 point)
        # Implement Transformer decoder block
        # Remember that transformer decoder block is composed of:
        # 1. Self-Attention
        # 2. Residual connection
        # 3. LayerNorm
        # 4. Cross-Attention
        # 5. Residual connection
        # 6. LayerNorm
        # 7. Fully-Connected Layer
        # 8. Dropout
        # 9. Residual connection
        # 10. LayerNorm
        # Note : Please write shape of the tensor for each line of code
        # YOUR CODE STARTS HERE (our implementation is about 10 lines)
#         print('calling decode', "decoder hidden states:",decoder_hidden_states.shape, 'encoder_hidden_states:',encoder_hidden_states.shape, "key_oadding:",key_padding_mask.shape)
        residual_1 = decoder_hidden_states
#         print("calling_self attention for decoder")
        out = self.self_attention(decoder_hidden_states, key_padding_mask=None)
        out = self.self_att_layer_norm(residual_1 + out)
        residual_2 = out
#         print("calling_cross attention for decoder")
        out = self.cross_attention(q = out, kv = encoder_hidden_states, key_padding_mask = key_padding_mask)
#         print("out after cross", out.shape)
#         print('----')
        out = self.cross_att_layer_norm(out+residual_2)
        out = self.fcn(out)
        out = self.dropout(out)
        residual_3 = out
        out = self.fcn_layer_norm(out+residual_3)
        

        ##YOUR CODE ENDS HERE##
        return out


class TransfomerEncoderDecoderModel(nn.Module):
    def __init__(
        self,
        *,
        num_layers,
        hidden,
        num_heads,
        fcn_hidden,
        max_seq_len,
        src_vocab_size,
        tgt_vocab_size,
        dropout=0.1,
    ):
        """A minimal implementation of Transformer Encoder Decoder Model
        
        Args:
            num_layer: number of layers for encoder and decoder (in total, model will have 2 * num_layers layers)
            hidden : embedding size and hidden size of attentions
            fcn_hidden: hidden size of fully-connected networks inside transformer layers
            vocab_size: size of vocabulary
            max_seq_len: maximum length of input, target sequence whichever is higher number
            src_vocab_size : source voacb size
            tgt_vocab_size : target voab size
        """
        super().__init__()
        self.src_vocab_size = src_vocab_size
        self.tgt_vocab_size = tgt_vocab_size
        self.num_layers = num_layers
        self.hidden = hidden
        self.num_heads = num_heads
        self.fcn_hidden = fcn_hidden
        self.dropout_rate = dropout
        self.max_seq_len = max_seq_len

        # Task 2.3 (1 point)
        # 1. Create encoder, decoder and positional embedding layer
        # Use nn.Embedding for that and make sure to include source and target vocabulary size
        # 2. Create a linear layer out_proj that will project contextualized representations
        # of size hidden to your target vocabulary size.
        # 3. Create a dropout layer
        # YOUR CODE STARTS HERE (our implementation is about 5 lines)

        self.encoder_embeddings = nn.Embedding(self.src_vocab_size, self.hidden)
        self.decoder_embeddings = nn.Embedding(self.tgt_vocab_size, self.hidden)
        self.positional_emb = nn.Embedding(self.max_seq_len, self.hidden)

        self.out_proj = nn.Linear(self.hidden, self.tgt_vocab_size)
        
        self.dropout = nn.Dropout(self.dropout_rate)
        # YOUR CODE ENDS HERE

        # Task 2.4 (1 point)
        # 1. Create a list of encoder Layers
        # 2. Create a list of decoder Layers
        #
        # Note that you need to wrap it with nn.ModuleList,
        # so that the parameters of the layers would be counted as the paramertes of the model
        # https://pytorch.org/docs/stable/generated/torch.nn.ModuleList.html
        # Read more about ModuleList here:
        # https://github.com/FrancescoSaverioZuppichini/Pytorch-how-and-when-to-use-Module-Sequential-ModuleList-and-ModuleDict
        # You can use for-loop of python list comprehension to create the list of layers
        #
        # YOUR CODE STARTS HERE (our implementation is 3-6 lines)
        self.encoder_layers = nn.ModuleList([TransformerEncoderLayer(hidden = self.hidden,
                                                                    num_heads = self.num_heads,
                                                                    fcn_hidden = self.fcn_hidden,
                                                                    dropout=self.dropout_rate
                                                                    )   
                                             for _ in range(self.num_layers)
                                            ])
        
        self.decoder_layers = nn.ModuleList([TransformerDecoderLayer(hidden = self.hidden,
                                                                    num_heads = self.num_heads,
                                                                    fcn_hidden = self.fcn_hidden,
                                                                    dropout=self.dropout_rate
                                                                    )   
                                             for _ in range(self.num_layers)
                                            ])
        
        # YOUR CODE ENDS HERE

    def _add_positions(self, sequence_tensor):
        """Adds positional embeddings to the input tensor.
        Args:
            sequence_tensor: FloatTensor[batch_size, seq_len, hidden]
        """
        seq_len = sequence_tensor.shape[1]
        positions = torch.arange(seq_len, device=sequence_tensor.device)
        positional_emb = self.positional_emb(positions)
        output = sequence_tensor + positional_emb
        return output

    def forward(
        self,
        input_ids=None,
        encoder_hidden_states=None,
        decoder_input_ids=None,
        key_padding_mask=None,
    ):
        """
        input_ids -> encoder_emb -> encoder -> 
                                                -->  decoder(encoder_output, decoder_emb) -> logits
        decoder_input_ids -> decoder_emb ---->

        Model accepts either input_ids or encoder_hidden_states.
        The former is used for training, the latter is used for inference, because during inference
        we don't have the target sequence and want to forward the decoder multiple times.
        To make the inference more efficient, we can only compute encoder output once and reuse it
        for all decoder steps.

        Meaning during training you should forward the model like this:
            model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)

        but during inference (generating translation) you should forward the model like this:
            model(encoder_hidden_states=encoder_hidden_states, decoder_input_ids=decoder_input_ids)

        Args:
            input_ids (LongTensor): Encoder input sequence of size (batch_size, seq_len)
            encoder_hidden_states (FloatTensor): Encoder hidden states of size (batch_size, seq_len, hidden)
            decoder_input_ids (LongTensor) : Decoder input sequence of size (batch_size, out_seq_len)
            key_padding_mask (ByteTensor): Mask of size (batch_size, seq_len) where 1 means that the token is padding

        Return:
            logits (FloatTensor): Logits for output sequence of size (batch_size, out_seq_len, dec_vocab_size)

        """
        if input_ids is None and encoder_hidden_states is None:
            raise ValueError("You should provide either input_ids or encoder_hidden_states")

        if encoder_hidden_states is None:
            encoder_hidden_states = self._encode(input_ids, key_padding_mask)

        logits = self._decode(encoder_hidden_states, decoder_input_ids, key_padding_mask)
#         print("Targte vocab size", decoder_input_ids.shape)
#         print("logits---------", logits.shape)

        return logits

    def _encode(self, input_ids, key_padding_mask):
        # Task 2.5 (2 points)
        # 1. Get source embeddings using self.encoder_embeddings
        # 2. Add positional embedding to encoder embeddings using _add_positions
        # 3. Pass source embeddings through the encoder layers, name them encoder_hidden_states
        # 3a. Remember to use key_padding_mask to mask out padding tokens
        # YOUR CODE STARTS HERE
        encoder_hidden_states = self.encoder_embeddings(input_ids)
        encoder_hidden_states = self._add_positions(encoder_hidden_states)
        for l in self.encoder_layers:
            encoder_hidden_states = l(encoder_hidden_states, key_padding_mask = key_padding_mask)

        # YOUR CODE ENDS HERE

        return encoder_hidden_states
    
    def _decode(self, encoder_hidden_states, decoder_input_ids, key_padding_mask):
        # TASK 2.6 (2 points)
        # 1. Get decoder embeddings using self.decoder_embeddings
        # 2. Add positional embedding to target embeddings using _add_positions
        # 3.Use decoder embeddings and encoder_hidden_states for the decoder input
        # (please use keyword arguments instead of positional arguments to minimize a chance of a bug)
        # 3a. Remember to use key_padding_mask to mask out padding tokens for the encoder inputs
        # 4. use self.out_proj to get output logits, a.k.a log-probabilies of the next translation tokens
        # YOUR CODE STARTS HERE
        decoder_embedding =  self.decoder_embeddings(decoder_input_ids)
        decoder_embedding = self._add_positions(decoder_embedding)
#         print("decoder_Embedding", decoder_embedding.shape)
        for l in self.decoder_layers:
            decoder_embedding = l(decoder_hidden_states = decoder_embedding, encoder_hidden_states=encoder_hidden_states, key_padding_mask = key_padding_mask)

        logits = self.out_proj(decoder_embedding)
        ## YOUR CODE ENDS HERE
        return logits

    ##############################################################################
    # Don't worry about any of the code below this line, but feel free to take a look
    # if you are interested in generation or model saving/loading.
    ##############################################################################
    @torch.inference_mode()
    def generate(
        self,
        input_ids,
        *,
        bos_token_id,
        eos_token_id,
        pad_token_id=None,
        key_padding_mask=None,
        max_length=50,
        beam_size=5,
        kind="beam_search",
    ):
        """
        Generate a translation given an input sequence.

        Args:
            input_ids (LongTensor): Encoder input sequence of size (batch_size, seq_len)
            bos_token_id (int): Beginning of sentence token id
            eos_token_id (int): End of sentence token id
            pad_token_id (int): Padding token id, required if doing beam search
            key_padding_mask (ByteTensor): Mask of size (batch_size, seq_len) where 1 means that the token is padding
            max_length (int): Maximum length of the generated sequence
            beam_size (int): Beam size for beam search
            kind (str): Can be either "greedy" or "beam_search"

        Return:
            decoded_ids (LongTensor): Decoder output sequence of size (batch_size, seq_len)
        """
        if kind not in ["greedy", "beam_search"]:
            raise ValueError("Unknown kind of generation: {}".format(kind))
        if kind == "beam_search" and pad_token_id is None:
            raise ValueError("Beam search requires a pad_token_id to be provided")

        if kind == "greedy":
            return self._generate_greedy(
                input_ids=input_ids,
                bos_token_id=bos_token_id,
                eos_token_id=eos_token_id,
                key_padding_mask=key_padding_mask,
                max_length=max_length,
            )
        
        # beam search only supports batch size 1
        beam_search_generations = []
        for i in range(input_ids.size(0)):
            _input_ids = input_ids[i].unsqueeze(0)
            _key_padding_mask = key_padding_mask[i].unsqueeze(0) if key_padding_mask is not None else None

            generated = self._generate_beam_search(
                input_ids=_input_ids,
                bos_token_id=bos_token_id,
                eos_token_id=eos_token_id,
                key_padding_mask=_key_padding_mask,
                max_length=max_length,
                beam_size=beam_size,
            )

            beam_search_generations.append(generated[0].detach().cpu().tolist())
        
        return pad(beam_search_generations, pad_id=eos_token_id)

    @torch.inference_mode()
    def _generate_greedy(
        self,
        input_ids,
        *,
        bos_token_id,
        eos_token_id,
        key_padding_mask=None,
        max_length=50,
    ):
        """
        Greedy generation of translation. Selects most likely word on every step.

        Args:
            input_ids (LongTensor): Encoder input sequence of size (batch_size, seq_len)
            max_length (int): Maximum length of the generated sequence
            bos_token_id (int): Beginning of sentence token id
            eos_token_id (int): End of sequence token id

        Return:
            translation (LongTensor): Decoder output sequence of size (batch_size, out_seq_len)
                where out_seq_len <= max_length
        """
        encoder_hidden_states = self._encode(input_ids, key_padding_mask)

        decoder_input_ids = torch.full((input_ids.shape[0], 1), bos_token_id, dtype=torch.long, device=input_ids.device)
        translation = torch.zeros((input_ids.shape[0], 0), dtype=torch.long, device=input_ids.device)

        eos_flags = torch.zeros((input_ids.shape[0],), dtype=torch.uint8, device=input_ids.device)

        for _ in range(max_length):
            logits = self._decode(encoder_hidden_states, decoder_input_ids, key_padding_mask)
            logits = logits[:, -1, :]

            next_token_id = torch.argmax(logits, dim=-1)

            decoder_input_ids = torch.cat((decoder_input_ids, next_token_id.unsqueeze(1)), dim=1)
            translation = torch.cat((translation, next_token_id.unsqueeze(1)), dim=1)

            eos_flags |= (next_token_id == eos_token_id)

            if eos_flags.all():
                break

        return translation

    @torch.inference_mode()
    def _generate_beam_search(
        self,
        input_ids,
        *,
        bos_token_id,
        eos_token_id,
        key_padding_mask=None,
        beam_size=5,
        max_length=50,
    ):
        """
        Beam search generation of translation.
        Heavily inspired by https://github.com/pcyin/pytorch_basic_nmt

        Args:
            input_ids (LongTensor): Encoder input sequence of size (batch_size, seq_len)
            max_length (int): Maximum length of the generated sequence
            bos_token_id (int): Beginning of sentence token id
            eos_token_id (int): End of sequence token id

        Return:
            translation (LongTensor): Decoder output sequence of size (batch_size, out_seq_len)
                where out_seq_len <= max_length
        """
        assert len(input_ids) == 1, "Beam search is only supported for a single input sequence"
        encoder_hidden_states = self._encode(input_ids, key_padding_mask)
        device = input_ids.device

        hypotheses = [[bos_token_id]]
        hyp_scores = torch.zeros(len(hypotheses), dtype=torch.float, device=device)
        completed_hypotheses = []

        for _ in range(max_length):
            if len(completed_hypotheses) >= beam_size:
                break

            hyp_num = len(hypotheses)
            expanded_encoder_hidden_states = encoder_hidden_states.expand(
                hyp_num,
                encoder_hidden_states.size(1),
                encoder_hidden_states.size(2),
            )

            # [batch_size*hyp_num=1*hyp_num, seq_len, hidden]
            hypotheses_tensor = torch.tensor(hypotheses, dtype=torch.int64, device=device)
            logits = self._decode(expanded_encoder_hidden_states, hypotheses_tensor, key_padding_mask)
            logits = logits[:, -1, :]  # [vocab_size]

            log_p_t = F.log_softmax(logits, dim=-1)
            live_hyp_num = beam_size - len(completed_hypotheses)

            # [hyp_num] -> [1, hyp_num] -> [hyp_num, vocab_size] -> [hyp_num * vocab_size]
            new_hyp_scores = (hyp_scores.unsqueeze(1).expand_as(log_p_t) + log_p_t).view(-1)
            # [live_hyp_num], [live_hyp_num]
            # for indices, the values range from 0 to hyp_num * vocab_size
            top_new_hyp_scores, top_new_hyp_pos = torch.topk(new_hyp_scores, k=live_hyp_num)

            # hypotheses ids in hyp_scores tensor [hyp_num,]
            prev_hyp_ids = torch.div(top_new_hyp_pos, self.tgt_vocab_size, rounding_mode='floor')

            # ids of the next words for each hypothesis
            token_ids = top_new_hyp_pos % self.tgt_vocab_size

            new_hypotheses = []
            new_hyp_scores = []

            # iterate live_hyp_num times
            for prev_hyp_id, hyp_token_id, cand_new_hyp_score in zip(prev_hyp_ids, token_ids, top_new_hyp_scores):
                prev_hyp_id = prev_hyp_id.item()
                hyp_token_id = hyp_token_id.item()
                cand_new_hyp_score = cand_new_hyp_score.item()

                new_hyp_sent = hypotheses[prev_hyp_id] + [hyp_token_id]
                if hyp_token_id == eos_token_id:
                    completed_hypotheses.append(Hypothesis(value=new_hyp_sent[1:-1], score=cand_new_hyp_score))
                else:
                    new_hypotheses.append(new_hyp_sent)
                    new_hyp_scores.append(cand_new_hyp_score)

            if len(completed_hypotheses) == beam_size:
                break

            hypotheses = new_hypotheses
            hyp_scores = torch.tensor(new_hyp_scores, dtype=torch.float, device=device)

        if len(completed_hypotheses) == 0:
            completed_hypotheses.append(Hypothesis(value=hypotheses[0][1:], score=hyp_scores[0].item()))
        
        completed_hypotheses.sort(key=lambda hyp: hyp.score, reverse=True)
        return torch.LongTensor(completed_hypotheses[0].value).unsqueeze(0)

    def save_pretrained(self, save_path):
        """Save the model weights to a directory

        Args:
            save_path: directory to save the model
        """
        config = {
            "num_layers": self.num_layers,
            "hidden": self.hidden,
            "num_heads": self.num_heads,
            "fcn_hidden": self.fcn_hidden,
            "src_vocab_size": self.src_vocab_size,
            "tgt_vocab_size": self.tgt_vocab_size,
            "max_seq_len": self.max_seq_len,
            "dropout": self.dropout_rate,
        }

        with open(os.path.join(save_path, "model_config.json"), "w") as f:
           json.dump(config, f)

        state_dict = self.state_dict()
        torch.save(state_dict, os.path.join(save_path, "model.pt"))
    
    @classmethod
    def from_pretrained(cls, save_path, map_location=None):
        """Load the model weights from a directory

        Args:
            save_path: directory to load the model
        """
        if map_location is None and not torch.cuda.is_available():
            map_location = "cpu"

        with open(os.path.join(save_path, "model_config.json"), "r") as f:
            config = json.load(f)

        model = cls(**config)
        state_dict = torch.load(os.path.join(save_path, "model.pt"), map_location=map_location)
        model.load_state_dict(state_dict)
        return model