RMLM/victims.py at main · Raibows/RMLM · GitHub

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
import os
import torch
import torch.nn as nn
from tools import tools_json_load
from transformers import BertModel
from tools import tools_get_logger
from config import config_dataset


def build_model(model_name, model_config, config:config_dataset, load_path, rank,):
    device = torch.device(rank)
    if not config.using_bert_vocab and model_config['pretrained_wv_path']:
        model_config['pretrained_wv_path'] = f"{model_config['pretrained_wv_path']}.{config.name}"

    if rank == 0:
        tools_get_logger('model').info(model_config)

    if model_name == 'lstm':
        assert model_config['bid'] == False
        model = LSTM(**model_config)
    elif model_name == 'bilstm':
        assert model_config['bid'] == True
        model = LSTM(**model_config)
    elif model_name == 'wordcnn':
        model = WordCNN(**model_config)
    elif 'bert' in model_name.lower():
        model = BERT(**model_config)
    else:
        raise NotImplementedError(f'not support {model_name}')
    if load_path:
        if load_path == 'config':
            load_path = config.get_load_path(model_name)
        load_dict = torch.load(load_path, map_location=device)
        model_dict = model.state_dict()
        states = {k: v for k, v in load_dict.items() if k in model_dict}
        not_loaded = [k for k in load_dict.keys() if k not in model_dict]
        need_loaded_missing = [k for k in model_dict.keys() if k not in load_dict]
        model_dict.update(states)
        if rank == 0:
            tools_get_logger('model').info(f'loading model from {load_path} but keys below are not loaded\n{not_loaded}\nneed load below keys but missing\n{need_loaded_missing}')
        model.load_state_dict(model_dict)
    if torch.cuda.is_available():
        model.to(device)
    model.eval()

    return model

def build_embedding_layer(pretrained_wv_path, vocab_size, embed_size, update_wv):
    if isinstance(pretrained_wv_path, str):
        assert os.path.exists(pretrained_wv_path), f"make sure {pretrained_wv_path} exists"
        wv = torch.tensor(tools_json_load(pretrained_wv_path))
        layer = nn.Embedding.from_pretrained(wv)
    else:
        layer = nn.Embedding(vocab_size, embed_size)
    layer.weight.requires_grad = update_wv
    return layer

class LSTM(nn.Module):
    def __init__(self, n_classes, vocab_size, embed_size, layer_num, hidden_size, bid=False, pretrained_wv_path=None, update_wv=True, dropout=0.5):
        super(LSTM, self).__init__()
        self.hidden_size = hidden_size
        self.bid = bid
        self.embedding_layer = build_embedding_layer(pretrained_wv_path, vocab_size, embed_size, update_wv)
        self.encoder = nn.LSTM(embed_size, hidden_size, layer_num, bidirectional=self.bid, batch_first=True)
        self.output_size = hidden_size * 2 if self.bid else hidden_size
        self.dropout = nn.Dropout(p=dropout)
        self.mlp = nn.Linear(self.output_size, n_classes)

    def encode_pack(self, embeds, X, X_len):
        sort = torch.sort(X_len, descending=True)
        sort_len, idx_sort = sort.values, sort.indices
        idx_reverse = torch.argsort(idx_sort)
        embeds = embeds.index_select(0, idx_sort)
        packed = nn.utils.rnn.pack_padded_sequence(embeds, sort_len.cpu(), batch_first=True)
        outs, (h, c) = self.encoder.forward(packed)
        outs, _ = nn.utils.rnn.pad_packed_sequence(outs, padding_value=0.0, batch_first=True)
        outs = outs.index_select(0, idx_reverse)
        outs = outs[torch.arange(X.size(0)), X_len - 1, :]
        return outs

    def forward(self, *inputs, return_sen_embed=False, gumbel=False):
        X, _, _, X_len = inputs
        if gumbel:
            embeds = X @ self.embedding_layer.weight
        else:
            embeds = self.embedding_layer.forward(X)  # [batch, sen_len, embed_size]
        embeds = self.dropout(embeds)
        outs = self.encode_pack(embeds, X, X_len)
        logits = self.mlp.forward(self.dropout(outs))
        if return_sen_embed:
            return logits, outs
        return logits


class WordCNN(nn.Module):
    def __init__(self, n_classes, vocab_size, embed_size, kernel_size:list, channel_size:list, mode, pretrained_wv_path=None, dropout=0.5):
        super(WordCNN, self).__init__()
        self.mode = mode
        if mode is None:
            self.embedding_layer = build_embedding_layer(None, vocab_size, embed_size, True)
        elif self.mode == 'static':
            assert pretrained_wv_path is not None
            self.embedding_layer = build_embedding_layer(pretrained_wv_path, vocab_size, embed_size, False)
        elif self.mode == 'update':
            assert pretrained_wv_path is not None
            self.embedding_layer = build_embedding_layer(pretrained_wv_path, vocab_size, embed_size, True)
        elif self.mode == 'dynamic':
            assert pretrained_wv_path is not None
            self.static_embedding_layer = build_embedding_layer(pretrained_wv_path, vocab_size, embed_size, False)
            self.update_embedding_layer = build_embedding_layer(pretrained_wv_path, vocab_size, embed_size, True)
            embed_size *= 2
        else:
            raise NotImplementedError('you have to choose mode from [None, static, update, dynamic]')
        self.embed_size = embed_size
        self.convs = nn.ModuleList()
        for c, k in zip(channel_size, kernel_size):
            self.convs.append(
                nn.Conv1d(in_channels=self.embed_size,
                          out_channels=c,
                          kernel_size=k)
            )
        self.pool = nn.AdaptiveMaxPool1d(output_size=1)
        self.relu = nn.ReLU()
        self.dropout = nn.Dropout(p=dropout)
        self.mlp = nn.Linear(in_features=sum(channel_size), out_features=n_classes)

    def __forward_embedding_layer(self, X, gumbel=False):
        if self.mode == 'dynamic':
            static_embeds = self.static_embedding_layer.forward(X) if not gumbel else X @ self.static_embedding_layer.weight
            update_embeds = self.update_embedding_layer.forward(X) if not gumbel else X @ self.update_embedding_layer.weight
            embeds = torch.cat([static_embeds, update_embeds], dim=2)
        else:
            embeds = self.embedding_layer.forward(X) if not gumbel else X @ self.embedding_layer.weight

        return self.dropout(embeds)

    def forward(self, *inputs, gumbel=False):
        X, _, _, X_len = inputs
        embeds = self.__forward_embedding_layer(X, gumbel)
        embeds = embeds.permute(0, 2, 1) # [batch, embed_size, sen_len]
        outs = torch.cat([self.pool(self.relu(conv(embeds))).squeeze(-1) for conv in self.convs], dim=1)
        logits = self.mlp(self.dropout(outs))

        return logits

class BERT(nn.Module):
    def __init__(self, n_classes, bert_name, dropout, **kwargs):
        super(BERT, self).__init__()
        self.bert = BertModel.from_pretrained(bert_name)
        self.dropout = nn.Dropout(p=dropout)
        self.mlp = nn.Linear(768, n_classes)

    def forward(self, *inputs, gumbel=False):
        X, types, masks, X_len = inputs
        if gumbel:
            inputs_embeds = X @ self.bert.embeddings.word_embeddings.weight
            cls = self.bert.forward(None, masks, types, inputs_embeds=inputs_embeds, return_dict=True)['last_hidden_state'][:, 0, :]
        else:
            cls = self.bert.forward(X, masks, types, return_dict=True)['last_hidden_state'][:, 0, :]
        logits = self.mlp.forward(self.dropout(cls))

        return logits


if __name__ == '__main__':
    pass