-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodel_block.py
161 lines (143 loc) · 7.41 KB
/
model_block.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
import numpy as np
from rdkit import Chem
from rdkit.Chem import QED
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.data import Data, Batch
import torch_geometric.nn as gnn
class GraphAgent(nn.Module):
def __init__(self, nemb, nvec, out_per_stem, out_per_mol, num_conv_steps, mdp_cfg, version='v1'):
super().__init__()
print(version)
if version == 'v5': version = 'v4'
self.version = version
self.embeddings = nn.ModuleList([
nn.Embedding(mdp_cfg.num_true_blocks + 1, nemb),
nn.Embedding(mdp_cfg.num_stem_types + 1, nemb),
nn.Embedding(mdp_cfg.num_stem_types, nemb)])
self.conv = gnn.NNConv(nemb, nemb, nn.Sequential(), aggr='mean')
nvec_1 = nvec * (version == 'v1' or version == 'v3')
nvec_2 = nvec * (version == 'v2' or version == 'v3')
self.block2emb = nn.Sequential(nn.Linear(nemb + nvec_1, nemb), nn.LeakyReLU(),
nn.Linear(nemb, nemb))
self.gru = nn.GRU(nemb, nemb)
self.stem2pred = nn.Sequential(nn.Linear(nemb * 2 + nvec_2, nemb), nn.LeakyReLU(),
nn.Linear(nemb, nemb), nn.LeakyReLU(),
nn.Linear(nemb, out_per_stem))
self.global2pred = nn.Sequential(nn.Linear(nemb, nemb), nn.LeakyReLU(),
nn.Linear(nemb, out_per_mol))
#self.set2set = Set2Set(nemb, processing_steps=3)
self.num_conv_steps = num_conv_steps
self.nemb = nemb
self.training_steps = 0
self.categorical_style = 'softmax'
self.escort_p = 6
def forward(self, graph_data, vec_data=None, do_stems=True):
blockemb, stememb, bondemb = self.embeddings
graph_data.x = blockemb(graph_data.x)
if do_stems:
graph_data.stemtypes = stememb(graph_data.stemtypes)
graph_data.edge_attr = bondemb(graph_data.edge_attr)
graph_data.edge_attr = (
graph_data.edge_attr[:, 0][:, :, None] * graph_data.edge_attr[:, 1][:, None, :]
).reshape((graph_data.edge_index.shape[1], self.nemb**2))
out = graph_data.x
if self.version == 'v1' or self.version == 'v3':
batch_vec = vec_data[graph_data.batch]
out = self.block2emb(torch.cat([out, batch_vec], 1))
else: # if self.version == 'v2' or self.version == 'v4':
out = self.block2emb(out)
h = out.unsqueeze(0)
for i in range(self.num_conv_steps):
m = F.leaky_relu(self.conv(out, graph_data.edge_index, graph_data.edge_attr))
out, h = self.gru(m.unsqueeze(0).contiguous(), h.contiguous())
out = out.squeeze(0)
# Index of the origin block of each stem in the batch (each
# stem is a pair [block idx, stem atom type], we need to
# adjust for the batch packing)
if do_stems:
if hasattr(graph_data, '_slice_dict'):
x_slices = torch.tensor(graph_data._slice_dict['x'], device=out.device)[graph_data.stems_batch]
else:
x_slices = torch.tensor(graph_data.__slices__['x'], device=out.device)[graph_data.stems_batch]
stem_block_batch_idx = (
x_slices
+ graph_data.stems[:, 0])
if self.version == 'v1' or self.version == 'v4':
stem_out_cat = torch.cat([out[stem_block_batch_idx], graph_data.stemtypes], 1)
elif self.version == 'v2' or self.version == 'v3':
stem_out_cat = torch.cat([out[stem_block_batch_idx],
graph_data.stemtypes,
vec_data[graph_data.stems_batch]], 1)
stem_preds = self.stem2pred(stem_out_cat)
else:
stem_preds = None
mol_preds = self.global2pred(gnn.global_mean_pool(out, graph_data.batch))
return stem_preds, mol_preds
def out_to_policy(self, s, stem_o, mol_o):
if self.categorical_style == 'softmax':
stem_e = torch.exp(stem_o)
mol_e = torch.exp(mol_o[:, 0])
elif self.categorical_style == 'escort':
stem_e = abs(stem_o)**self.escort_p
mol_e = abs(mol_o[:, 0])**self.escort_p
Z = gnn.global_add_pool(stem_e, s.stems_batch).sum(1) + mol_e + 1e-8
return mol_e / Z, stem_e / Z[s.stems_batch, None]
def action_negloglikelihood(self, s, a, g, stem_o, mol_o):
mol_p, stem_p = self.out_to_policy(s, stem_o, mol_o)
#print(Z.shape, Z.min().item(), Z.mean().item(), Z.max().item())
mol_lsm = torch.log(mol_p + 1e-20)
stem_lsm = torch.log(stem_p + 1e-20)
#print(mol_lsm.shape, mol_lsm.min().item(), mol_lsm.mean().item(), mol_lsm.max().item())
#print(stem_lsm.shape, stem_lsm.min().item(), stem_lsm.mean().item(), stem_lsm.max().item(), '--')
return -self.index_output_by_action(s, stem_lsm, mol_lsm, a)
def index_output_by_action(self, s, stem_o, mol_o, a):
if hasattr(s, '_slice_dict'):
stem_slices = torch.tensor(s._slice_dict['stems'][:-1], dtype=torch.long, device=stem_o.device)
else:
stem_slices = torch.tensor(s.__slices__['stems'][:-1], dtype=torch.long, device=stem_o.device)
return (
stem_o[stem_slices + a[:, 1]][
torch.arange(a.shape[0]), a[:, 0]] * (a[:, 0] >= 0)
+ mol_o * (a[:, 0] == -1))
def sum_output(self, s, stem_o, mol_o):
return gnn.global_add_pool(stem_o, s.stems_batch).sum(1) + mol_o
def mol2graph(mol, mdp, floatX=torch.float, bonds=False, nblocks=False):
f = lambda x: torch.tensor(x, dtype=torch.long, device=mdp.device)
if len(mol.blockidxs) == 0:
data = Data( # There's an extra block embedding for the empty molecule
x=f([mdp.num_true_blocks]),
edge_index=f([[], []]),
edge_attr=f([]).reshape((0, 2)),
stems=f([(0, 0)]),
stemtypes=f([mdp.num_stem_types])) # also extra stem type embedding
return data
edges = [(i[0], i[1]) for i in mol.jbonds]
#edge_attrs = [mdp.bond_type_offset[i[2]] + i[3] for i in mol.jbonds]
t = mdp.true_blockidx
if 0:
edge_attrs = [((mdp.stem_type_offset[t[mol.blockidxs[i[0]]]] + i[2]) * mdp.num_stem_types +
(mdp.stem_type_offset[t[mol.blockidxs[i[1]]]] + i[3]))
for i in mol.jbonds]
else:
edge_attrs = [(mdp.stem_type_offset[t[mol.blockidxs[i[0]]]] + i[2],
mdp.stem_type_offset[t[mol.blockidxs[i[1]]]] + i[3])
for i in mol.jbonds]
# Here stem_type_offset is a list of offsets to know which
# embedding to use for a particular stem. Each (blockidx, atom)
# pair has its own embedding.
stemtypes = [mdp.stem_type_offset[t[mol.blockidxs[i[0]]]] + i[1] for i in mol.stems]
data = Data(x=f([t[i] for i in mol.blockidxs]),
edge_index=f(edges).T if len(edges) else f([[],[]]),
edge_attr=f(edge_attrs) if len(edges) else f([]).reshape((0,2)),
stems=f(mol.stems) if len(mol.stems) else f([(0,0)]),
stemtypes=f(stemtypes) if len(mol.stems) else f([mdp.num_stem_types]))
data.to(mdp.device)
assert not bonds and not nblocks
return data
def mols2batch(mols, mdp):
batch = Batch.from_data_list(
mols, follow_batch=['stems'])
batch.to(mdp.device)
return batch