MUG/new_decoder.py at main · ulysse-beaudouin/MUG · 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
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
import torch
import torch.nn as nn
import torch.nn.functional as F

from utils import fixed_positional_encoding


class RNNDecoder(nn.Module):
    def __init__(self, latent_dim: int, hidden_dim: int, n_layers: int, max_nodes: int = 50, tau: float = 1.0, hard: bool = True,  batch_size: int=256):
        super(RNNDecoder, self).__init__()
        self.latent_dim = latent_dim
        self.hidden_dim = hidden_dim
        self.n_layers = n_layers
        self.max_nodes = max_nodes
        self.tau = tau
        self.hard = hard
        self.batch_size = batch_size

        self.rnn = nn.GRU(
            input_size = 2*latent_dim,
            hidden_size = hidden_dim,
            num_layers = n_layers,
            batch_first = True
        )
        self.node_proj = nn.Linear(hidden_dim, hidden_dim, bias=False)
        # self.proj_pre_rnn = nn.Linear(2*latent_dim, latent_dim)
        self.adj_mlp = nn.Sequential(
            nn.Linear(2 * hidden_dim, hidden_dim),
            nn.ReLU(),
            nn.Linear(hidden_dim, 2)
        )
        self.embeddings = nn.Embedding(self.max_nodes, latent_dim)
        # device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        # self.positional_encodings = fixed_positional_encoding(max_nodes, latent_dim, device=device)

    def forward(self, z: torch.Tensor, n_nodes: int, n_edges: int):
        batch_size = z.size(0)
        seq_input = z.unsqueeze(1).repeat(1, self.max_nodes, 1)
        positions = torch.arange(self.max_nodes, device=z.device)
        positional_embeddings = self.embeddings(positions).repeat(batch_size, 1, 1)
        # seq_input += self.positional_encodings.unsqueeze(0)
        seq_input = torch.cat((seq_input, positional_embeddings), dim=-1)
        # seq_input = self.proj_pre_rnn(seq_input)
        # seq_input += positional_embeddings

        mask = torch.arange(self.max_nodes, device=z.device).unsqueeze(0).expand(batch_size, self.max_nodes) < n_nodes.unsqueeze(1)
        seq_input[~mask] = 0.0
        packed_input = torch.nn.utils.rnn.pack_padded_sequence(seq_input, n_nodes.cpu(), batch_first=True, enforce_sorted=False)
        rnn_out, _ = self.rnn(packed_input)
        rnn_out, _ = torch.nn.utils.rnn.pad_packed_sequence(rnn_out, batch_first=True, total_length=self.max_nodes)
        node_emb = self.node_proj(rnn_out)
        idx = torch.triu_indices(self.max_nodes, self.max_nodes, offset=1, device=z.device)
        emb_i = node_emb[:, idx[0], :]
        emb_j = node_emb[:, idx[1], :]
        pair_emb = torch.cat([emb_i, emb_j], dim=-1)
        logits = self.adj_mlp(pair_emb)
        adjacency_values = F.gumbel_softmax(logits, tau=self.tau, hard=self.hard, dim=-1)[..., 0]
        adj = torch.zeros(batch_size, self.max_nodes, self.max_nodes, device=z.device)
        adj[:, idx[0], idx[1]] = adjacency_values
        adj = adj + torch.transpose(adj, 1, 2)
        indices = torch.arange(self.max_nodes, device=z.device).unsqueeze(0)
        mask = indices < n_nodes.unsqueeze(1)
        mask2d = mask.unsqueeze(2) & mask.unsqueeze(1)
        adj = adj * mask2d.float()
        return adj


class RNNDecoderwAtt(nn.Module):
    def __init__(self, latent_dim: int, hidden_dim: int, n_layers: int, max_nodes: int = 50, tau: float = 1.0, hard: bool = True,  batch_size: int=256):
        super(RNNDecoderwAtt, self).__init__()
        self.latent_dim = latent_dim
        self.hidden_dim = hidden_dim
        self.n_layers = n_layers
        self.max_nodes = max_nodes
        self.tau = tau
        self.hard = hard
        self.batch_size = batch_size

        self.rnn = nn.GRU(
            input_size = latent_dim,
            hidden_size = hidden_dim,
            num_layers = n_layers,
            batch_first = True
        )
        self.node_proj = nn.Linear(hidden_dim, latent_dim, bias=False)

        self.adj_mlp = nn.Linear(2*latent_dim, 2)
        # self.pairwise_attention = PairwiseAttention(hidden_dim=self.hidden_dim, num_heads=4)
        self.embeddings = nn.Embedding(self.max_nodes, latent_dim)
        self.transformer = nn.TransformerEncoder(
            nn.TransformerEncoderLayer(d_model=2*latent_dim, nhead=2),
            num_layers=1
        )
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.positional_encodings = fixed_positional_encoding(max_nodes, latent_dim, device=device)

    def forward(self, z: torch.Tensor, n_nodes: int, n_edges: int):
        batch_size = z.size(0)
        seq_input = z.unsqueeze(1).repeat(1, self.max_nodes, 1)
        positions = torch.arange(self.max_nodes, device=z.device)
        positional_embeddings = self.embeddings(positions).repeat(batch_size, 1, 1)

        # seq_input = torch.cat((seq_input, positional_embeddings), dim=-1)
        seq_input += positional_embeddings
        seq_input += self.positional_encodings.unsqueeze(0)

        mask = torch.arange(self.max_nodes, device=z.device).unsqueeze(0).expand(batch_size, self.max_nodes) < n_nodes.unsqueeze(1)
        seq_input[~mask] = 0.0
        packed_input = torch.nn.utils.rnn.pack_padded_sequence(seq_input, n_nodes.cpu(), batch_first=True, enforce_sorted=False)
        rnn_out, _ = self.rnn(packed_input)
        rnn_out, _ = torch.nn.utils.rnn.pad_packed_sequence(rnn_out, batch_first=True, total_length=self.max_nodes)
        idx = torch.triu_indices(self.max_nodes, self.max_nodes, offset=1, device=z.device)
        node_emb = self.node_proj(rnn_out)
        emb_i = node_emb[:, idx[0], :]
        emb_j = node_emb[:, idx[1], :]
        pair_emb = torch.cat([emb_i, emb_j], dim=-1)
        pair_emb = self.transformer(pair_emb)
        logits = self.adj_mlp(pair_emb)
        adjacency_values = F.gumbel_softmax(logits, tau=self.tau, hard=self.hard, dim=-1)[..., 0]
        adj = torch.zeros(batch_size, self.max_nodes, self.max_nodes, device=z.device)
        adj[:, idx[0], idx[1]] = adjacency_values
        adj = adj + torch.transpose(adj, 1, 2)
        indices = torch.arange(self.max_nodes, device=z.device).unsqueeze(0)
        mask = indices < n_nodes.unsqueeze(1)
        mask2d = mask.unsqueeze(2) & mask.unsqueeze(1)
        adj = adj * mask2d.float()
        return adj


class TransformerDecoderModel(nn.Module):
    def __init__(self, latent_dim: int, hidden_dim: int, n_layers: int, max_nodes: int = 50, tau: float = 1.0, hard: bool = True, batch_size: int = 256):
        super(TransformerDecoderModel, self).__init__()
        self.latent_dim = latent_dim
        self.hidden_dim = hidden_dim
        self.n_layers = n_layers
        self.max_nodes = max_nodes
        self.tau = tau
        self.hard = hard
        self.batch_size = batch_size

        decoder_layer = nn.TransformerDecoderLayer(
            d_model=latent_dim,
            nhead=2,
            dim_feedforward=hidden_dim
        )
        self.transformer_decoder = nn.TransformerDecoder(decoder_layer, num_layers=n_layers)
        self.adj_mlp = nn.Sequential(
            nn.Linear(latent_dim * 2, hidden_dim),
            nn.ReLU(),
            nn.Linear(hidden_dim, 2)
        )

        self.embeddings = nn.Embedding(self.max_nodes, latent_dim)
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.positional_encodings = fixed_positional_encoding(max_nodes, latent_dim, device=device)


    def forward(self, z: torch.Tensor, n_nodes: int, n_edges: int):
        batch_size = z.size(0)
        seq_input = z.unsqueeze(1).repeat(1, self.max_nodes, 1)
        positions = torch.arange(self.max_nodes, device=z.device)
        positional_embeddings = self.embeddings(positions).repeat(batch_size, 1, 1)
        seq_input += self.positional_encodings.unsqueeze(0)
        seq_input += positional_embeddings
        mask = torch.arange(self.max_nodes, device=z.device).unsqueeze(0).expand(batch_size, self.max_nodes) < n_nodes.unsqueeze(1)
        seq_input[~mask] = 0.0
        tgt_mask = nn.Transformer.generate_square_subsequent_mask(self.max_nodes).to(z.device)
        src_padding_mask = torch.arange(self.max_nodes, device=z.device).unsqueeze(0).expand(batch_size, self.max_nodes) >= n_nodes.unsqueeze(1)
        memory = seq_input
        tgt = seq_input.permute(1, 0, 2)
        transformer_out = self.transformer_decoder(tgt, memory.permute(1, 0, 2), tgt_mask=tgt_mask, memory_key_padding_mask=src_padding_mask)
        node_emb = transformer_out.permute(1, 0, 2)
        node_emb = node_emb * mask.unsqueeze(-1).float()
        idx = torch.triu_indices(self.max_nodes, self.max_nodes, offset=1, device=z.device)
        emb_i = node_emb[:, idx[0], :]
        emb_j = node_emb[:, idx[1], :]
        pair_emb = torch.cat([emb_i, emb_j], dim=-1)
        logits = self.adj_mlp(pair_emb)
        adjacency_values = F.gumbel_softmax(logits, tau=self.tau, hard=self.hard, dim=-1)[..., 0]
        adj = torch.zeros(batch_size, self.max_nodes, self.max_nodes, device=z.device)
        adj[:, idx[0], idx[1]] = adjacency_values
        adj = adj + torch.transpose(adj, 1, 2)
        indices = torch.arange(self.max_nodes, device=z.device).unsqueeze(0)
        mask = indices < n_nodes.unsqueeze(1)
        mask2d = mask.unsqueeze(2) & mask.unsqueeze(1)
        adj = adj * mask2d.float()
        return adj