File size: 5,348 Bytes
23bd7af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# coding=utf-8
# Copyright (c) 2020, NVIDIA CORPORATION.  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.

"""Multiple choice model."""

import torch

from megatron import get_args, print_rank_last
from megatron import mpu
from megatron.model.enums import AttnMaskType
from megatron.model.bert_model import bert_extended_attention_mask, bert_position_ids
from megatron.model.language_model import get_language_model
from megatron.model.utils import get_linear_layer
from megatron.model.utils import init_method_normal
from megatron.model.utils import scaled_init_method_normal
from .module import MegatronModule


class MultipleChoice(MegatronModule):

    def __init__(self,
                 num_tokentypes=2,
                 pre_process=True,
                 post_process=True):
        super(MultipleChoice, self).__init__(share_word_embeddings=False)
        args = get_args()

        init_method = init_method_normal(args.init_method_std)
        self.pre_process = pre_process
        self.post_process = post_process

        self.language_model, self._language_model_key = get_language_model(
            num_tokentypes=num_tokentypes,
            add_pooler=True,
            encoder_attn_mask_type=AttnMaskType.padding,
            init_method=init_method,
            scaled_init_method=scaled_init_method_normal(args.init_method_std,
                                                         args.num_layers),
            pre_process=self.pre_process,
            post_process=self.post_process)

        # Multi-choice head.
        if self.post_process:
            self.multichoice_dropout = torch.nn.Dropout(args.hidden_dropout)
            self.multichoice_head = get_linear_layer(args.hidden_size, 1,
                                                     init_method)
            self._multichoice_head_key = 'multichoice_head'

    def set_input_tensor(self, input_tensor):
        """See megatron.model.transformer.set_input_tensor()"""
        self.language_model.set_input_tensor(input_tensor)

    def forward(self, model_input, attention_mask, tokentype_ids=None):

        # [batch, choices, sequence] --> [batch * choices, sequence] -->
        #    transformer --> [batch, choices] --> softmax

        # Ensure the shape is [batch-size, choices, sequence]
        assert len(attention_mask.shape) == 3
        num_choices = attention_mask.shape[1]

        # Reshape and treat choice dimension the same as batch.
        attention_mask = attention_mask.view(-1, attention_mask.size(-1))
        extended_attention_mask = bert_extended_attention_mask(attention_mask)

        input_ids = model_input
        # Do the same as attention_mask for input_ids, tokentype_ids
        assert len(input_ids.shape) == 3
        assert len(tokentype_ids.shape) == 3
        input_ids = input_ids.view(-1, input_ids.size(-1))
        tokentype_ids = tokentype_ids.view(-1, tokentype_ids.size(-1))
        position_ids = bert_position_ids(input_ids)

        lm_output = self.language_model(
            input_ids,
            position_ids,
            extended_attention_mask,
            tokentype_ids=tokentype_ids
        )
        if self.post_process:
            _, pooled_output = lm_output
            multichoice_output = self.multichoice_dropout(pooled_output)
            multichoice_logits = self.multichoice_head(multichoice_output)

            # Reshape back to separate choices.
            multichoice_logits = multichoice_logits.view(-1, num_choices)

            return multichoice_logits
        return lm_output

    def state_dict_for_save_checkpoint(self, destination=None, prefix='',
                                       keep_vars=False):
        """For easy load when model is combined with other heads,
        add an extra key."""

        state_dict_ = {}
        state_dict_[self._language_model_key] \
            = self.language_model.state_dict_for_save_checkpoint(
                destination, prefix, keep_vars)
        if self.post_process:
            state_dict_[self._multichoice_head_key] \
                = self.multichoice_head.state_dict(
                    destination, prefix, keep_vars)
        return state_dict_

    def load_state_dict(self, state_dict, strict=True):
        """Customized load."""

        self.language_model.load_state_dict(
            state_dict[self._language_model_key], strict=strict)
        if self.post_process:
            if self._multichoice_head_key in state_dict:
                self.multichoice_head.load_state_dict(
                    state_dict[self._multichoice_head_key], strict=strict)
            else:
                print_rank_last('***WARNING*** could not find {} in the checkpoint, '
                                'initializing to random'.format(
                                    self._multichoice_head_key))