File size: 10,832 Bytes
8918ac7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
import torch
import gc
import torch.nn as nn
import torch.nn.functional as F
from typing import Tuple
from .pooling import Attention1dPoolingHead, MeanPoolingHead, LightAttentionPoolingHead
from .pooling import MeanPooling, MeanPoolingProjection

def rotate_half(x):
    x1, x2 = x.chunk(2, dim=-1)
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(x, cos, sin):
    cos = cos[:, :, : x.shape[-2], :]
    sin = sin[:, :, : x.shape[-2], :]

    return (x * cos) + (rotate_half(x) * sin)

class RotaryEmbedding(nn.Module):
    """

    Rotary position embeddings based on those in

    [RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer). Query and keys are transformed by rotation

    matrices which depend on their relative positions.

    """

    def __init__(self, dim: int):
        super().__init__()
        # Generate and save the inverse frequency buffer (non trainable)
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64).float() / dim))
        inv_freq = inv_freq
        self.register_buffer("inv_freq", inv_freq)

        self._seq_len_cached = None
        self._cos_cached = None
        self._sin_cached = None

    def _update_cos_sin_tables(self, x, seq_dimension=2):
        seq_len = x.shape[seq_dimension]

        # Reset the tables if the sequence length has changed,
        # or if we're on a new device (possibly due to tracing for instance)
        if seq_len != self._seq_len_cached or self._cos_cached.device != x.device:
            self._seq_len_cached = seq_len
            t = torch.arange(x.shape[seq_dimension], device=x.device).type_as(self.inv_freq)
            freqs = torch.outer(t, self.inv_freq)
            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)

            self._cos_cached = emb.cos()[None, None, :, :]
            self._sin_cached = emb.sin()[None, None, :, :]

        return self._cos_cached, self._sin_cached

    def forward(self, q: torch.Tensor, k: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        self._cos_cached, self._sin_cached = self._update_cos_sin_tables(k, seq_dimension=-2)

        return (
            apply_rotary_pos_emb(q, self._cos_cached, self._sin_cached),
            apply_rotary_pos_emb(k, self._cos_cached, self._sin_cached),
        )


class CrossModalAttention(nn.Module):
    def __init__(self, args):
        super().__init__()
        self.attention_head_size = args.hidden_size // args.num_attention_head
        assert (
            self.attention_head_size * args.num_attention_head == args.hidden_size
        ), "Embed size needs to be divisible by num heads"
        self.num_attention_head = args.num_attention_head
        self.hidden_size = args.hidden_size
        
        self.query_proj = nn.Linear(args.hidden_size, args.hidden_size)
        self.key_proj = nn.Linear(args.hidden_size, args.hidden_size)
        self.value_proj = nn.Linear(args.hidden_size, args.hidden_size)
        
        self.dropout = nn.Dropout(args.attention_probs_dropout)
        
        self.out_proj = nn.Linear(args.hidden_size, args.hidden_size)
        self.rotary_embeddings = RotaryEmbedding(dim=self.attention_head_size)

    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
        new_x_shape = x.size()[:-1] + (self.num_attention_head, self.attention_head_size)
        x = x.view(new_x_shape)
        return x.permute(0, 2, 1, 3)
    
    def forward(self, query, key, value, attention_mask=None, output_attentions=False):
        key_layer = self.transpose_for_scores(self.key_proj(key))
        value_layer = self.transpose_for_scores(self.value_proj(value))
        query_layer = self.transpose_for_scores(self.query_proj(query))
        query_layer = query_layer * self.attention_head_size**-0.5
        
        query_layer, key_layer = self.rotary_embeddings(query_layer, key_layer)
        
        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
        
        if attention_mask is not None:
            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
            attention_scores = attention_scores.masked_fill(attention_mask == 0, float('-inf'))
        
        attention_probs = F.softmax(attention_scores, dim=-1)
        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        attention_probs = self.dropout(attention_probs)
        context_layer = torch.matmul(attention_probs, value_layer)

        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size,)
        context_layer = context_layer.view(new_context_layer_shape)
        
        outputs = (context_layer, attention_probs) if output_attentions else context_layer
        
        return outputs

class AdapterModel(nn.Module):
    def __init__(self, args):
        super().__init__()
        self.args = args
        
        if 'foldseek_seq' in args.structure_seq:
            self.foldseek_embedding = nn.Embedding(args.vocab_size, args.hidden_size)
            self.cross_attention_foldseek = CrossModalAttention(args)
        if 'ss8_seq' in args.structure_seq:
            self.ss_embedding = nn.Embedding(args.vocab_size, args.hidden_size)
            self.cross_attention_ss = CrossModalAttention(args)
        if 'esm3_structure_seq' in args.structure_seq:
            self.esm3_structure_embedding = nn.Embedding(args.vocab_size, args.hidden_size)
            self.cross_attention_esm3_structure = CrossModalAttention(args)
        
        self.layer_norm = nn.LayerNorm(args.hidden_size)
        
        if args.pooling_method == 'attention1d':
            self.classifier = Attention1dPoolingHead(args.hidden_size, args.num_labels, args.pooling_dropout)
        elif args.pooling_method == 'mean':
            if "PPI" in args.dataset:
                self.pooling = MeanPooling()
                self.projection = MeanPoolingProjection(args.hidden_size, args.num_labels, args.pooling_dropout)
            else:
                self.classifier = MeanPoolingHead(args.hidden_size, args.num_labels, args.pooling_dropout)
        elif args.pooling_method == 'light_attention':
            self.classifier = LightAttentionPoolingHead(args.hidden_size, args.num_labels, args.pooling_dropout)
        else:
            raise ValueError(f"classifier method {args.pooling_method} not supported")
    
    def plm_embedding(self, plm_model, aa_seq, attention_mask, structure_tokens=None):
        with torch.no_grad():
            if "ProSST" in self.args.plm_model:
                outputs = plm_model(input_ids=aa_seq, attention_mask=attention_mask, ss_input_ids=structure_tokens, output_hidden_states=True)
            elif "Prime" in self.args.plm_model or "deep" in self.args.plm_model:
                outputs = plm_model(input_ids=aa_seq, attention_mask=attention_mask, output_hidden_states=True)
            elif self.training and hasattr(self, 'args') and self.args.training_method == 'full':
                outputs = plm_model(input_ids=aa_seq, attention_mask=attention_mask)
            else:
                outputs = plm_model(input_ids=aa_seq, attention_mask=attention_mask)
            if "ProSST" in self.args.plm_model or "Prime" in self.args.plm_model:
                seq_embeds = outputs.hidden_states[-1]
            else:
                seq_embeds = outputs.last_hidden_state
        gc.collect()
        torch.cuda.empty_cache()
        return seq_embeds
    
    def forward(self, plm_model, batch):
        if "ProSST" in self.args.plm_model:
            aa_seq, attention_mask, stru_tokens = batch['aa_seq_input_ids'], batch['aa_seq_attention_mask'], batch['aa_seq_stru_tokens']
            seq_embeds = self.plm_embedding(plm_model, aa_seq, attention_mask, stru_tokens)
        else:
            aa_seq, attention_mask = batch['aa_seq_input_ids'], batch['aa_seq_attention_mask']
            seq_embeds = self.plm_embedding(plm_model, aa_seq, attention_mask)

        if 'foldseek_seq' in self.args.structure_seq:
            foldseek_seq = batch['foldseek_seq_input_ids']
            foldseek_embeds = self.foldseek_embedding(foldseek_seq)
            foldseek_embeds = self.cross_attention_foldseek(foldseek_embeds, seq_embeds, seq_embeds, attention_mask)
            embeds = seq_embeds + foldseek_embeds
            embeds = self.layer_norm(embeds)
        
        if 'ss8_seq' in self.args.structure_seq:
            ss_seq = batch['ss8_seq_input_ids']
            ss_embeds = self.ss_embedding(ss_seq)
            
            if 'foldseek_seq' in self.args.structure_seq:
                # cross attention with foldseek
                ss_embeds = self.cross_attention_ss(ss_embeds, embeds, embeds, attention_mask)
                embeds = ss_embeds + embeds
            else:
                # cross attention with sequence
                ss_embeds = self.cross_attention_ss(ss_embeds, seq_embeds, seq_embeds, attention_mask)
                embeds = ss_embeds + seq_embeds
            embeds = self.layer_norm(embeds)
        
        if 'esm3_structure_seq' in self.args.structure_seq:
            esm3_structure_seq = batch['esm3_structure_seq_input_ids']
            esm3_structure_embeds = self.esm3_structure_embedding(esm3_structure_seq)
            
            if 'foldseek_seq' in self.args.structure_seq:
                # cross attention with foldseek
                esm3_structure_embeds = self.cross_attention_esm3_structure(esm3_structure_embeds, embeds, embeds, attention_mask)
                embeds = esm3_structure_embeds + embeds
            elif 'ss8_seq' in self.args.structure_seq:
                # cross attention with ss8
                esm3_structure_embeds = self.cross_attention_esm3_structure(esm3_structure_embeds, ss_embeds, ss_embeds, attention_mask)
                embeds = esm3_structure_embeds + ss_embeds
            else:
                # cross attention with sequence
                esm3_structure_embeds = self.cross_attention_esm3_structure(esm3_structure_embeds, seq_embeds, seq_embeds, attention_mask)
                embeds = esm3_structure_embeds + seq_embeds
            embeds = self.layer_norm(embeds)
        
        if self.args.structure_seq:
            logits = self.classifier(embeds, attention_mask)
        else:
            logits = self.classifier(seq_embeds, attention_mask)            
        
        return logits