Spaces:
Running
on
Zero
Running
on
Zero
File size: 12,667 Bytes
7968cb0 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 |
import torch.nn as nn
from src.modules.pifold_module import *
from torch_scatter import scatter_softmax, scatter_log_softmax
def positional_encoding(x):
batch_size, seq_len, hidden_size = x.size()
pos = torch.arange(0, seq_len).float().unsqueeze(1).repeat(1, hidden_size // 2)
div = torch.exp(torch.arange(0, hidden_size, 2).float() * (-torch.log(torch.tensor(10000.0)) / hidden_size))
sin = torch.sin(pos * div)
cos = torch.cos(pos * div)
pos_encoding = torch.cat([sin, cos], dim=-1).unsqueeze(0).repeat(batch_size, 1, 1)
return pos_encoding
class MSAAttention(nn.Module):
def __init__(self, hidden_dim) -> None:
super().__init__()
self.MSA_Q = nn.Linear(hidden_dim, hidden_dim)
self.MSA_K = nn.Linear(hidden_dim, hidden_dim)
self.MSA_V = nn.Linear(hidden_dim, hidden_dim)
def forward(self, inputs_embeds):
pos_enc = positional_encoding(inputs_embeds)
inputs_embeds = inputs_embeds + pos_enc
query = self.MSA_Q(inputs_embeds) # shape: [batch, N, 128]
key = self.MSA_K(inputs_embeds) # shape: [batch, N, 128]
value = self.MSA_V(inputs_embeds) # shape: [batch, N, 128]
attn_scores = torch.bmm(query, key.transpose(1, 2))
attn_weights = nn.functional.softmax(attn_scores, dim=2)
attn_output = torch.bmm(attn_weights, value)
return attn_output
class GNNTuning_Model(nn.Module):
def __init__(self, args, num_encoder_layers, hidden_dim, input_design_dim, input_esm_dim, input_struct_dim=3072, input_esmif_dim=512, dropout=0.1):
super(GNNTuning_Model, self).__init__()
self.args = args
encoder_layers = []
for i in range(num_encoder_layers):
encoder_layers.append(
GeneralGNN(hidden_dim, hidden_dim*2, dropout=dropout, node_net = "AttMLP", edge_net = "EdgeMLP", node_context = 1, edge_context = 0),
)
self.encoder_layers = nn.Sequential(*encoder_layers)
from transformers import AutoTokenizer
from transformers.models.esm.modeling_esm import EsmModel, EsmEmbeddings
from transformers.models.esm.configuration_esm import EsmConfig
self.tokenizer = AutoTokenizer.from_pretrained("facebook/esm2_t33_650M_UR50D", cache_dir="./cache_dir/")
config = EsmConfig(attention_probs_dropout_prob=0,
hidden_size=hidden_dim,
intermediate_size=1280,
mask_token_id=32,
num_attention_heads=12,
num_hidden_layers=3,
pad_token_id=1,
position_embedding_type="rotary",
token_dropout=False,
vocab_size=33
)
self.DesignEmbed = EsmEmbeddings(config)
self.ESMEmbed = EsmEmbeddings(config)
self.EdgeEmbed = nn.Sequential(nn.Linear(416+16+16, 512),
nn.ReLU(),
nn.Linear(512, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim,hidden_dim))
self.DesignConf = nn.Sequential(nn.Linear(1, 128),
nn.ReLU(),
nn.Linear(128, 128),
nn.ReLU(),
nn.Linear(128,1),
nn.Sigmoid())
self.ESMConf = nn.Sequential(nn.Linear(1, 128),
nn.ReLU(),
nn.Linear(128, 128),
nn.ReLU(),
nn.Linear(128,1))
self.DesignProj = nn.Sequential(nn.Linear(input_design_dim, 512),
nn.ReLU(),
nn.Linear(512, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim,hidden_dim))
self.ESMProj = nn.Sequential(nn.Linear(input_esm_dim, 512),
nn.ReLU(),
nn.Linear(512, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim,hidden_dim))
self.StructProj = nn.Sequential(nn.Linear(input_struct_dim, 512),
nn.ReLU(),
nn.Linear(512, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim,hidden_dim))
self.ESMIFProj = nn.Sequential(nn.Linear(input_esmif_dim, 512),
nn.ReLU(),
nn.Linear(512, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim,hidden_dim))
self.ReadOut = nn.Linear(hidden_dim,33)
# self.TimeEmbed = nn.Embedding(20, hidden_dim)
# self.ProbEmbed = nn.Sequential(nn.Linear(33, 512),
# nn.ReLU(),
# nn.Linear(512, hidden_dim),
# nn.ReLU(),
# nn.Linear(hidden_dim,hidden_dim))
self.MLP1 = nn.Sequential(nn.Linear(1, 512),
nn.ReLU(),
nn.Linear(512, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim,1),
nn.Sigmoid())
self.MLP2 = nn.Sequential(nn.Linear(1, 512),
nn.ReLU(),
nn.Linear(512, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim,1),
nn.Sigmoid())
# def embed_gnn(self, pretrain_gnn, mask_select_id, mask_select_feat):
# gnn_embed = self.DesignEmbed(mask_select_id(pretrain_gnn['pred_ids'])).squeeze()
# gnn_conf = self.DesignConf(mask_select_id(pretrain_gnn['confs']))
# gnn_proj = self.DesignProj(mask_select_feat(pretrain_gnn['embeds']))
# if self.args.use_confembed:
# return gnn_embed*F.sigmoid(gnn_conf) + gnn_proj
# else:
# return gnn_embed + gnn_proj
# def embed_esm(self, pretrain_esm, mask_select_id, mask_select_feat):
# esm_embed = self.ESMEmbed(mask_select_id(pretrain_esm['pred_ids'])).squeeze()
# esm_conf = self.ESMConf(mask_select_id(pretrain_esm['confs']))
# esm_proj = self.ESMProj(mask_select_feat(pretrain_esm['embeds']))
# if self.args.use_confembed:
# return esm_embed*F.sigmoid(esm_conf) + esm_proj
# else:
# return esm_embed + esm_proj
# def embed_struct(self, pretrain_struct, mask_select_feat):
# struct_proj = self.StructProj(mask_select_feat(pretrain_struct['embeds']))
# return struct_proj
# def embed_esmif(self, pretrain_esmif, mask_select_feat):
# struct_proj = self.ESMIFProj(mask_select_feat(pretrain_esmif['embeds']))
# return struct_proj
def fuse(self, mask_select_feat, mask_select_id, gnn_embed=None, esm_embed=None, gearnet_embed=None, esmif_embed=None, gnn_pred_id=None, esm_pred_id=None, confidence=None, confidence_esm=None):
gnn, esm, gearnet, esmif, conf = 0, 0, 0, 0, 1.0
if gnn_embed is not None:
gnn = self.DesignProj(mask_select_feat(gnn_embed))
gnn += self.DesignEmbed(mask_select_id(gnn_pred_id)).squeeze()
if esm_embed is not None:
esm = self.ESMProj(mask_select_feat(esm_embed))
esm += self.ESMEmbed(mask_select_id(esm_pred_id)).squeeze()
if gearnet_embed is not None:
gearnet = self.StructProj(mask_select_feat(gearnet_embed))
if esmif_embed is not None:
esmif = self.ESMIFProj(mask_select_feat(esmif_embed))
if conf is not None:
conf = self.DesignConf(mask_select_id(confidence))
esm_conf = self.ESMConf(mask_select_id(confidence_esm))
return (gnn*conf+esm*esm_conf+gearnet+esmif)
def forward(self, batch):
pretrain_design, h_E_raw, E_idx, mask_attend, batch_id = batch['pretrain_design'], batch['h_E'], batch['E_idx'], batch['attention_mask'], batch['batch_id']
if self.args.use_LM:
pretrain_esm_msa = batch['pretrain_esm_msa']
if self.args.use_gearnet:
pretrain_struct = batch['pretrain_struct']
if self.args.use_esmif:
pretrain_esmif = batch['pretrain_esmif']
mask_select_id = lambda x: torch.masked_select(x, mask_attend.bool()).reshape(-1,1)
mask_select_feat = lambda x: torch.masked_select(x, mask_attend.bool().unsqueeze(-1)).reshape(-1,x.shape[-1])
inputs_embeds = 0
for i in range(self.args.msa_n):
gnn_embed = pretrain_design['embeds']
esm_embed = pretrain_esm_msa['embeds'][i] if self.args.use_LM else None
gearnet_embed = pretrain_struct['embeds'][i] if self.args.use_gearnet else None
esmif_embed = pretrain_esmif['embeds'] if self.args.use_esmif else None
confidence = pretrain_design['confs']
confidence_esm = pretrain_esm_msa['confs'][i]
inputs_embeds += self.fuse(mask_select_feat, mask_select_id, gnn_embed, esm_embed, gearnet_embed, esmif_embed, pretrain_design['pred_ids'], pretrain_esm_msa['pred_ids'][i], confidence, confidence_esm)
h_V = inputs_embeds
h_E = self.EdgeEmbed(h_E_raw)
for layer in self.encoder_layers:
h_V, h_E = layer(h_V, h_E, E_idx, batch_id)
logits = self.ReadOut(h_V)
# confidence update
old_confs = mask_select_id(pretrain_design['confs'])
confs = torch.softmax(logits, dim=-1).max(dim=-1)[0][:,None]
h_V = h_V*self.MLP1(confs-old_confs) + inputs_embeds*self.MLP2(old_confs-confs)
logits = self.ReadOut(h_V)
B, N = pretrain_design['confs'].shape
vocab_size = logits.shape[-1]
new_logits = torch.zeros(B,N,vocab_size, device=logits.device).reshape(B*N, vocab_size)
new_logits = new_logits.masked_scatter_(mask_attend.bool().view(-1,1), logits)
new_logits = new_logits.reshape(B,N,vocab_size)
log_probs = torch.log_softmax(new_logits, dim=-1)
device = logits.device
seqs, confs, embeds, probs2 = self.to_matrix(h_V, logits, batch_id)
ret = {"pred_ids":seqs['input_ids'].to(device),
"confs":confs,
"embeds":embeds,
"probs":probs2,
"attention_mask":seqs['attention_mask'].to(device),
"h_E":h_E_raw,
"E_idx":E_idx,
"batch_id":batch_id,
"log_probs":log_probs}
return ret
def to_matrix(self, h_V, logits, batch_id):
probs = F.softmax(logits, dim=-1)
conf, pred_id = probs.max(dim=-1)
maxL = 0
for b in batch_id.unique():
mask = batch_id==b
L = mask.sum()
if L>maxL:
maxL=L
confs = []
seqs = []
embeds = []
probs2 = []
for b in batch_id.unique():
mask = batch_id==b
# elements = [alphabet[int(id)] for id in pred_id[mask]]
elements = self.tokenizer.decode(pred_id[mask]).split(" ")
seqs.append(elements)
confs.append(conf[mask])
embeds.append(h_V[mask])
probs2.append(probs[mask])
seqs = self.tokenizer(["".join(one) for one in seqs], padding=True, truncation=True, return_tensors='pt', add_special_tokens=False)
confs = torch.stack([F.pad(one, (0, maxL-len(one))) for one in confs])
embeds = torch.stack([F.pad(one, (0,0, 0, maxL-len(one))) for one in embeds])
probs2 = torch.stack([F.pad(one, (0,0, 0, maxL-len(one)), value=1/33) for one in probs2])
return seqs, confs, embeds, probs2
|