tree2code/lib_treeautoencoder.py at master · zomux/tree2code · 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
#!/usr/bin/env python
# -*- coding: utf-8 -*-

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function


import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import dgl
import dgl.init

from lib_treelstm import TreeLSTMCell
from lib_treelstm import DecoderTreeLSTMCell
from lib_treelstm_dataloader import BilingualTreeDataLoader
from lib_semhash import SemanticHashing

from nmtlab.modules.transformer_modules import TransformerEmbedding
from nmtlab.modules.transformer_modules import TransformerEncoderLayer

class TreeAutoEncoder(nn.Module):

    def __init__(self, dataset, hidden_size=256, code_bits=5, without_source=False, dropout_ratio=0.1):
        super(TreeAutoEncoder, self).__init__()
        assert isinstance(dataset, BilingualTreeDataLoader)
        self.hidden_size = hidden_size
        self._vocab_size = dataset.src_vocab().size()
        self._label_size = dataset.label_vocab().size()
        self._code_bits = code_bits
        self._without_source = without_source

        # Encoder
        self.src_embed_layer = TransformerEmbedding(self._vocab_size, self.hidden_size, dropout_ratio=dropout_ratio)
        self.encoder_norm = nn.LayerNorm(self.hidden_size)
        self.encoder_layers = nn.ModuleList()
        ff_size = hidden_size * 4
        for _ in range(3):
            layer = TransformerEncoderLayer(self.hidden_size, ff_size, n_att_head=8, dropout_ratio=dropout_ratio)
            self.encoder_layers.append(layer)

        self.label_embed_layer = nn.Embedding(self._label_size, self.hidden_size)
        self.enc_cell = TreeLSTMCell(hidden_size, hidden_size)
        self.dec_cell = DecoderTreeLSTMCell(hidden_size, hidden_size)
        self.dropout = nn.Dropout(dropout_ratio)
        self.logit_nn = nn.Linear(self.hidden_size, self._label_size)
        if code_bits > 0:
            self.semhash = SemanticHashing(hidden_size, bits=code_bits)
        else:
            self.semhash = None
        self.initialize_parameters()

    def initialize_parameters(self):
        """Initialize the parameters in the model."""
        # Initialize weights
        for param in self.parameters():
            shape = param.shape
            if len(shape) > 1:
                nn.init.xavier_uniform_(param)
        for module in self.modules():
            if isinstance(module, nn.Linear) and module.bias is not None:
                nn.init.constant_(module.bias, 0.0)

    def encode_source(self, src_seq, src_mask, meanpool=False):
        src_seq = src_seq.long()
        x = self.src_embed_layer(src_seq)
        for l, layer in enumerate(self.encoder_layers):
            x = layer(x, src_mask)
        encoder_states = self.encoder_norm(x)
        if meanpool:
            encoder_states = encoder_states * src_mask.unsqueeze(-1)
            encoder_states = encoder_states.sum(1) / (src_mask.sum(1).unsqueeze(-1) + 10e-8)

        encoder_outputs = {
            "encoder_states": encoder_states,
            "src_mask": src_mask
        }
        return encoder_outputs

    def forward(self, src, enc_tree, dec_tree, return_code=False, **kwargs):
        self._init_graph(enc_tree, dec_tree)
        # Soruce encoding
        src_mask = torch.ne(src, 0).float()
        encoder_outputs = self.encode_source(src, src_mask, meanpool=True)
        encoder_states = encoder_outputs["encoder_states"]
        # Tree encoding
        enc_x = enc_tree.ndata["x"].cuda()
        x_embeds = self.label_embed_layer(enc_x)
        enc_tree.ndata['iou'] = self.enc_cell.W_iou(self.dropout(x_embeds))
        enc_tree.ndata['h'] = torch.zeros((enc_tree.number_of_nodes(), self.hidden_size)).cuda()
        enc_tree.ndata['c'] = torch.zeros((enc_tree.number_of_nodes(), self.hidden_size)).cuda()
        enc_tree.ndata['mask'] = enc_tree.ndata['mask'].float().cuda()
        dgl.prop_nodes_topo(enc_tree)
        # Obtain root representation
        root_mask = enc_tree.ndata["mask"].float().cuda()
        # root_idx = torch.arange(root_mask.shape[0])[root_mask > 0].cuda()
        root_h = self.dropout(enc_tree.ndata.pop("h")) * root_mask.unsqueeze(-1)
        orig_h = root_h.clone()[root_mask > 0]
        partial_h = orig_h
        if self._without_source:
            partial_h += encoder_states

        # Discretization
        if self._code_bits > 0:
            if return_code:
                codes = self.semhash(partial_h, return_code=True)
                ret = {"codes": codes}
                return ret
            else:
                partial_h = self.semhash(partial_h)
            if not self._without_source:
                partial_h += encoder_states

        root_h[root_mask > 0] = partial_h
        # Tree decoding
        dec_x = dec_tree.ndata["x"].cuda()
        dec_embeds = self.label_embed_layer(dec_x)
        dec_tree.ndata['iou'] = self.dec_cell.W_iou(self.dropout(dec_embeds))
        dec_tree.ndata['h'] = root_h
        dec_tree.ndata['c'] = torch.zeros((enc_tree.number_of_nodes(), self.hidden_size)).cuda()
        dec_tree.ndata['mask'] = dec_tree.ndata['mask'].float().cuda()
        dgl.prop_nodes_topo(dec_tree)
        # Compute logits
        all_h = self.dropout(dec_tree.ndata.pop("h"))
        logits = self.logit_nn(all_h)
        logp = F.log_softmax(logits, 1)
        # Compute loss
        y_labels = dec_tree.ndata["y"].cuda()
        monitor = {}
        loss = F.nll_loss(logp, y_labels, reduction="mean")
        acc = (logits.argmax(1) == y_labels).float().mean()
        monitor["loss"] = loss
        monitor["label_accuracy"] = acc
        return monitor

    def _init_graph(self, enc_tree, dec_tree):
        enc_tree.register_message_func(self.enc_cell.message_func)
        enc_tree.register_reduce_func(self.enc_cell.reduce_func)
        enc_tree.register_apply_node_func(self.enc_cell.apply_node_func)
        enc_tree.set_n_initializer(dgl.init.zero_initializer)
        dec_tree.register_message_func(self.dec_cell.message_func)
        dec_tree.register_reduce_func(self.dec_cell.reduce_func)
        dec_tree.register_apply_node_func(self.dec_cell.apply_node_func)
        dec_tree.set_n_initializer(dgl.init.zero_initializer)

    def load_pretrain(self, pretrain_path):
        first_param = next(self.parameters())
        device_str = str(first_param.device)
        pre_state_dict = torch.load(pretrain_path, map_location=device_str)["model_state"]
        keys = list(pre_state_dict.keys())
        for key in keys:
            if "semhash" in key:
                pre_state_dict.pop(key)
        state_dict = self.state_dict()
        state_dict.update(pre_state_dict)
        self.load_state_dict(state_dict)

    def load(self, model_path):
        first_param = next(self.parameters())
        device_str = str(first_param.device)
        state_dict = torch.load(model_path, map_location=device_str)["model_state"]
        self.load_state_dict(state_dict)