-
Notifications
You must be signed in to change notification settings - Fork 3
/
model.py
181 lines (155 loc) · 7.58 KB
/
model.py
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
from torch.nn.modules import TransformerEncoder, TransformerEncoderLayer
from torch.nn.modules import TransformerDecoder, TransformerDecoderLayer
from layer import AtomEncoder, BondDecoder
from torch import nn
import torch
import torch.nn.functional as F
import math
import pdb
import os
class MoleculeEncoder(nn.Module):
def __init__(self, ntoken, dim, nhead, nlayer, dropout, rank):
super().__init__()
self.atom_encoder = AtomEncoder(ntoken, dim, dropout=dropout, rank=rank)
layer = TransformerEncoderLayer(dim, nhead, dim, dropout)
self.transformer_encoder = TransformerEncoder(layer, nlayer)
# multihead attention assumes [len, batch, dim]
# padding_mask = True equivalent to mask = -inf
def forward(self, element, bond, aroma, charge, mask, segment, reactant=None):
'''
element, long [b, l] element index
bonds, long [b, l, 4]
aroma, long [b, l]
charge, long [b, l] +1 0 -1
mask, [b, l] true if masked
returns [l, b, dim]
'''
embedding = self.atom_encoder(element, bond, aroma, charge, segment, reactant)
encoder_output = self.transformer_encoder(embedding, src_key_padding_mask=mask)
return encoder_output
class VariationalEncoder(nn.Module):
def __init__(self, dim, nhead, nlayer, dropout, rank=0):
super().__init__()
layer = TransformerDecoderLayer(dim, nhead, dim, dropout)
self.transformer_decoder = TransformerDecoder(layer, nlayer)
self.head = nn.Linear(dim, 2*dim)
def KL(self, posterior):
# prior is standard gaussian distribution
mu, logsigma = posterior['mu'], posterior['logsigma']
# no matter what shape
logvar = logsigma*2
loss = 0.5 * torch.sum(mu * mu+ torch.exp(logvar) - 1 - logvar, 1)
return loss
def forward(self, src, src_mask, tgt, tgt_mask):
"""
src, tgt [L, b, dim]
src_mask, tgt_mask, [B, L]
"""
l, b, dim = src.shape
src_mask, tgt_mask = src_mask.permute(0, 1), tgt_mask.permute(0, 1)
decoder_output = self.transformer_decoder(src, tgt,
memory_key_padding_mask=tgt_mask, tgt_key_padding_mask=src_mask).permute(1, 2, 0)
# [L, B, dim] to [B, dim, L]
tmp = decoder_output * (1-src_mask.float().unsqueeze(1))
tmp = tmp.mean(dim=2)
# [B, dim]
posterior = self.head(tmp)
result = {}
result['mu'] = posterior[:, 0:dim]
result['logsigma'] = posterior[:, dim:]
return result, self.KL(result)
class MoleculeDecoder(nn.Module):
def __init__(self, vae, dim, nhead, nlayer, dropout, rank=0):
super().__init__()
layer = TransformerEncoderLayer(dim, nhead, dim, dropout)
self.transformer_encoder = TransformerEncoder(layer, nlayer)
self.latent_head = nn.Linear(dim, dim)
self.bond_decoder = BondDecoder(dim, rank)
self.charge_head = nn.Conv1d(dim, 13, 1) #-6 to +6
self.aroma_head = nn.Conv1d(dim, 1, 1)
self.vae = vae
self.rank = rank
def forward(self, src, src_bond, src_mask, latent, tgt_bond, tgt_aroma, tgt_charge, tgt_mask):
l, b, dim = src.size()
if self.vae:
tmp = torch.randn(b, dim).to(self.rank)
latent = tmp * latent['logsigma'].exp() + latent['mu']
latent = self.latent_head(latent)
src = src + latent.expand(l, b, dim)
encoder_output = self.transformer_encoder(src, src_key_padding_mask=src_mask)
eps = 1e-6
result = self.bond_decoder(encoder_output, src_bond, src_mask, tgt_bond, tgt_mask)
tgt_mask = 1-tgt_mask.float()
encoder_output = encoder_output.permute(1, 2, 0)
aroma_logit = self.aroma_head(encoder_output)
BCE = nn.BCEWithLogitsLoss(reduction='none')
tgt_aroma = tgt_aroma.bool().float()
aroma_logit = aroma_logit.view(b, l)
aroma_loss = BCE(aroma_logit, tgt_aroma.float()) #[B, L]
aroma_loss = aroma_loss * tgt_mask
aroma_loss = aroma_loss.sum(dim=1)
result['aroma_loss'] = aroma_loss
charge_logit = self.charge_head(encoder_output)
CE = nn.CrossEntropyLoss(reduction='none')
# assumes [B, C, L] (input, target)
tgt_charge = tgt_charge.long() + 6
charge_loss = CE(charge_logit, tgt_charge)
charge_loss = charge_loss * tgt_mask
charge_loss = charge_loss.sum(dim=1)
result['charge_loss'] = charge_loss
result['pred_loss'] = result['bond_loss'] + aroma_loss + charge_loss
return result
def sample(self, src_embedding, src_bond, padding_mask, temperature=1):
"""
decode the molecule into bond [B, L, 4], given representation of [L, b, dim]
"""
l, b, dim = src_embedding.shape
latent = 0
if self.vae:
latent = torch.randn(1, b, dim).to(self.rank) *temperature
latent = self.latent_head(latent)
src_embedding = src_embedding + latent
encoder_output = self.transformer_encoder(src_embedding, src_key_padding_mask=padding_mask)
result = {}
bond = self.bond_decoder(encoder_output, src_bond, padding_mask)
result['bond'] = bond.long()
encoder_output = encoder_output.permute(1, 2, 0)
# to [b, c, l]
aroma_logit = self.aroma_head(encoder_output)
aroma = (aroma_logit > 0).view(b, l)
result['aroma'] = aroma.long()
charge_logit = self.charge_head(encoder_output)
charge = torch.argmax(charge_logit, dim=1)- 6
result['charge'] = charge.long()
return result
class MoleculeVAE(nn.Module):
def __init__(self, args, ntoken, dim=128, nlayer=8, nhead=8, dropout=0.1):
super().__init__()
self.args = args
self.rank = args.local_rank
self.M_encoder = MoleculeEncoder(ntoken, dim, nhead, nlayer, dropout, self.rank)
self.P_encoder = MoleculeEncoder(ntoken, dim, nhead, nlayer, dropout, self.rank)
if args.vae:
self.V_encoder = VariationalEncoder(dim, nhead, nlayer, dropout, self.rank)
self.M_decoder = MoleculeDecoder(args.vae, dim, nhead, nlayer, dropout, self.rank)
def forward(self, mode, tensors, temperature = 1):
src = self.M_encoder(tensors['element'], tensors['src_bond'], tensors['src_aroma'],
tensors['src_charge'], tensors['src_mask'], tensors['src_segment'], tensors['reactant'] )
if mode is 'train':
bond, aroma, charge = tensors['tgt_bond'], tensors['tgt_aroma'], tensors['tgt_charge']
if self.args.vae:
tgt = self.P_encoder(tensors['element'], bond, aroma, charge,
tensors['tgt_mask'], tensors['tgt_segment'])
posterior, kl = self.V_encoder(src, tensors['src_mask'], tgt, tensors['tgt_mask'])
result = self.M_decoder(src, tensors['src_bond'], tensors['src_mask'], posterior,
bond, aroma, charge, tensors['tgt_mask'])
result['kl'] = kl
result['loss'] = result['pred_loss'] + self.args.beta*kl
else:
result = self.M_decoder(src, tensors['src_mask'], None,
bond, aroma, charge, tensors['tgt_mask'])
result['loss'] = result['pred_loss']
elif mode is 'sample':
""" returns bond[B, L, 4], aroma [B, L], charge[B, L], weight [B, L, L]"""
result = self.M_decoder.sample(src, tensors['src_bond'], tensors['src_mask'], temperature)
return result