ICPC2023MoCoCS/model.py at main · FDUDSDE/ICPC2023MoCoCS · 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
168
169
170
import torch.nn as nn
import torch
from torch.nn import CrossEntropyLoss


class Model(nn.Module):
    def __init__(self, encoder, encoder_momentum, args):
        super(Model, self).__init__()
        self.K = 8192  # queue size
        self.m = 0.999  # momentum coefficient
        self.T = 0.05  # temperature coefficient
        self.args = args
        self.encoder = encoder
        self.encoder_momentum = encoder_momentum
        for param, param_m in zip(self.encoder.parameters(), self.encoder_momentum.parameters()):
            param_m.data.copy_(param.data)  # initialize
            param_m.requires_grad = False  # not update by gradient
        self.register_buffer("queue_code", torch.randn(768, self.K))
        self.register_buffer("queue_nl", torch.randn(768, self.K))
        self.register_buffer("queue_ast", torch.randn(768, self.K))
        self.register_buffer("queue_dfg", torch.randn(768, self.K))
        self.queue_code = nn.functional.normalize(self.queue_code, dim=0)
        self.queue_nl = nn.functional.normalize(self.queue_nl, dim=0)
        self.queue_ast = nn.functional.normalize(self.queue_ast, dim=0)
        self.queue_dfg = nn.functional.normalize(self.queue_dfg, dim=0)
        self.register_buffer("queue_ptr_code", torch.zeros(1, dtype=torch.long))
        self.register_buffer("queue_ptr_nl", torch.zeros(1, dtype=torch.long))
        self.register_buffer("queue_ptr_ast", torch.zeros(1, dtype=torch.long))
        self.register_buffer("queue_ptr_dfg", torch.zeros(1, dtype=torch.long))

    @torch.no_grad()
    def _momentum_update_key_encoder(self):
        """
        Momentum update encoder
        """
        for param, param_m in zip(self.encoder.parameters(), self.encoder_momentum.parameters()):
            param_m.data = param_m.data * self.m + param.data * (1. - self.m)

    @torch.no_grad()
    def _dequeue_and_enqueue_nl(self, nl_vec):
        """
        NL Queue dequeue and enqueue
        """
        nl_size = nl_vec.shape[0]
        ptr = int(self.queue_ptr_nl)
        if nl_size == self.args.train_batch_size:
            # replace the keys at ptr (dequeue and enqueue)
            self.queue_nl[:, ptr:ptr + nl_size] = nl_vec.T
            ptr = (ptr + nl_size) % self.K  # move pointer
            self.queue_ptr_nl[0] = ptr

    @torch.no_grad()
    def _dequeue_and_enqueue_code(self, code_vec):
        """
        PL Queue dequeue and enqueue
        """
        code_size = code_vec.shape[0]
        ptr = int(self.queue_ptr_code)
        if code_size == self.args.train_batch_size:
            # replace the keys at ptr (dequeue and enqueue)
            self.queue_code[:, ptr:ptr + code_size] = code_vec.T
            ptr = (ptr + code_size) % self.K  # move pointer
            self.queue_ptr_code[0] = ptr

    @torch.no_grad()
    def _dequeue_and_enqueue_ast(self, ast_vec):
        """
        AST Queue dequeue and enqueue
        """
        ast_size = ast_vec.shape[0]
        ptr = int(self.queue_ptr_ast)
        if ast_size == self.args.train_batch_size:
            # replace the keys at ptr (dequeue and enqueue)
            self.queue_ast[:, ptr:ptr + ast_size] = ast_vec.T
            ptr = (ptr + ast_size) % self.K  # move pointer
            self.queue_ptr_ast[0] = ptr

    @torch.no_grad()
    def _dequeue_and_enqueue_dfg(self, dfg_vec):
        """
        DFG Queue dequeue and enqueue
        """
        dfg_size = dfg_vec.shape[0]
        ptr = int(self.queue_ptr_dfg)
        if dfg_size == self.args.train_batch_size:
            # replace the keys at ptr (dequeue and enqueue)
            self.queue_dfg[:, ptr:ptr + dfg_size] = dfg_vec.T
            ptr = (ptr + dfg_size) % self.K  # move pointer
            self.queue_ptr_dfg[0] = ptr

    def forward(self, code_inputs=None, nl_inputs=None, ast_inputs=None, dfg_inputs=None):
        if code_inputs is not None and nl_inputs is not None and ast_inputs is not None:
            outputs_code = self.encoder(code_inputs, attention_mask=code_inputs.ne(1))[0]
            code_vec = (outputs_code*code_inputs.ne(1)[:, :, None]).sum(1)/code_inputs.ne(1).sum(-1)[:, None]
            code_vec = torch.nn.functional.normalize(code_vec, p=2, dim=1)
            outputs_nl = self.encoder(nl_inputs, attention_mask=nl_inputs.ne(1))[0]
            nl_vec = (outputs_nl*nl_inputs.ne(1)[:, :, None]).sum(1)/nl_inputs.ne(1).sum(-1)[:, None]
            nl_vec = torch.nn.functional.normalize(nl_vec, p=2, dim=1)

            with torch.no_grad():
                self._momentum_update_key_encoder()
                outputs_code_m = self.encoder_momentum(code_inputs, attention_mask=code_inputs.ne(1))[0]
                code_vec_m = (outputs_code_m*code_inputs.ne(1)[:, :, None]).sum(1)/code_inputs.ne(1).sum(-1)[:, None]
                code_vec_m = torch.nn.functional.normalize(code_vec_m, p=2, dim=1)

                outputs_nl_m = self.encoder_momentum(nl_inputs, attention_mask=nl_inputs.ne(1))[0]
                nl_vec_m = (outputs_nl_m*nl_inputs.ne(1)[:, :, None]).sum(1)/nl_inputs.ne(1).sum(-1)[:, None]
                nl_vec_m = torch.nn.functional.normalize(nl_vec_m, p=2, dim=1)

                outputs_ast_m = self.encoder_momentum(ast_inputs, attention_mask=ast_inputs.ne(1))[0]
                ast_vec_m = (outputs_ast_m*ast_inputs.ne(1)[:, :, None]).sum(1)/ast_inputs.ne(1).sum(-1)[:, None]
                ast_vec_m = torch.nn.functional.normalize(ast_vec_m, p=2, dim=1)

                outputs_dfg_m = self.encoder_momentum(dfg_inputs, attention_mask=dfg_inputs.ne(1))[0]
                dfg_vec_m = (outputs_dfg_m*dfg_inputs.ne(1)[:, :, None]).sum(1)/dfg_inputs.ne(1).sum(-1)[:, None]
                dfg_vec_m = torch.nn.functional.normalize(dfg_vec_m, p=2, dim=1)

            l_pos_code = torch.einsum('nc,nc->n', [code_vec, nl_vec_m]).unsqueeze(-1)
            l_pos_nl = torch.einsum('nc,nc->n', [nl_vec, code_vec_m]).unsqueeze(-1)
            l_pos_code_ast = torch.einsum('nc,nc->n', [code_vec, ast_vec_m]).unsqueeze(-1)
            l_pos_nl_ast = torch.einsum('nc,nc->n', [nl_vec, ast_vec_m]).unsqueeze(-1)
            l_pos_code_dfg = torch.einsum('nc,nc->n', [code_vec, dfg_vec_m]).unsqueeze(-1)
            l_pos_nl_dfg = torch.einsum('nc,nc->n', [nl_vec, dfg_vec_m]).unsqueeze(-1)

            l_neg_code = torch.einsum('nc,ck->nk', [code_vec, self.queue_nl.clone().detach()])
            l_neg_nl = torch.einsum('nc,ck->nk', [nl_vec, self.queue_code.clone().detach()])
            l_neg_code_ast = torch.einsum('nc,ck->nk', [code_vec, self.queue_ast.clone().detach()])
            l_neg_nl_ast = torch.einsum('nc,ck->nk', [nl_vec, self.queue_ast.clone().detach()])
            l_neg_code_dfg = torch.einsum('nc,ck->nk', [code_vec, self.queue_dfg.clone().detach()])
            l_neg_nl_dfg = torch.einsum('nc,ck->nk', [nl_vec, self.queue_dfg.clone().detach()])

            logits_code = torch.cat([l_pos_code, l_neg_code], dim=1)
            logits_nl = torch.cat([l_pos_nl, l_neg_nl], dim=1)
            logits_code_ast = torch.cat([l_pos_code_ast, l_neg_code_ast], dim=1)
            logits_code_dfg = torch.cat([l_pos_code_dfg, l_neg_code_dfg], dim=1)
            logits_nl_ast = torch.cat([l_pos_nl_ast, l_neg_nl_ast], dim=1)
            logits_nl_dfg = torch.cat([l_pos_nl_dfg, l_neg_nl_dfg], dim=1)
            logits_code /= self.T
            logits_nl /= self.T
            logits_code_ast /= self.T
            logits_nl_ast /= self.T
            logits_code_dfg /= self.T
            logits_nl_dfg /= self.T

            loss_fct = CrossEntropyLoss()
            loss_code = loss_fct(logits_code, torch.zeros(logits_code.size(0), dtype=torch.long).cuda())
            loss_nl = loss_fct(logits_nl, torch.zeros(logits_nl.size(0), dtype=torch.long).cuda())
            loss_code_ast = loss_fct(logits_code_ast, torch.zeros(logits_code_ast.size(0), dtype=torch.long).cuda())
            loss_nl_ast = loss_fct(logits_nl_ast, torch.zeros(logits_nl_ast.size(0), dtype=torch.long).cuda())
            loss_nl_dfg = loss_fct(logits_nl_dfg, torch.zeros(logits_nl_dfg.size(0), dtype=torch.long).cuda())
            loss_code_dfg = loss_fct(logits_code_dfg, torch.zeros(logits_code_dfg.size(0), dtype=torch.long).cuda())
            loss = loss_code + loss_nl + (loss_code_ast + loss_nl_ast)*0.01 + (loss_nl_dfg + loss_code_dfg)*0.01

            self._dequeue_and_enqueue_code(code_vec_m)
            self._dequeue_and_enqueue_nl(nl_vec_m)
            self._dequeue_and_enqueue_ast(ast_vec_m)
            self._dequeue_and_enqueue_dfg(dfg_vec_m)

            return loss, code_vec, nl_vec

        if code_inputs is not None:
            outputs = self.encoder(code_inputs, attention_mask=code_inputs.ne(1))[0]
            code_vec = (outputs*code_inputs.ne(1)[:, :, None]).sum(1)/code_inputs.ne(1).sum(-1)[:, None]
            code_vec = torch.nn.functional.normalize(code_vec, p=2, dim=1)
            return code_vec
        else:
            outputs = self.encoder(nl_inputs, attention_mask=nl_inputs.ne(1))[0]
            nl_vec = (outputs*nl_inputs.ne(1)[:, :, None]).sum(1)/nl_inputs.ne(1).sum(-1)[:, None]
            nl_vec = torch.nn.functional.normalize(nl_vec, p=2, dim=1)
            return nl_vec