forked from gmh14/data_efficient_grammar
-
Notifications
You must be signed in to change notification settings - Fork 0
/
grammar_generation.py
189 lines (168 loc) · 7.64 KB
/
grammar_generation.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
181
182
183
184
185
186
187
188
189
from rdkit import Chem
from functools import partial
from multiprocessing import Pool
from fuseprop import find_clusters, extract_subgraph, get_mol, get_smiles, find_fragments
from copy import deepcopy
import numpy as np
import torch
import argparse
from private import *
from agent import sample
def data_processing(input_smiles, GNN_model_path, motif=False):
input_mols = []
input_graphs = []
init_subgraphs = []
subgraphs_idx = []
input_graphs_dict = {}
init_edge_flag = 0
for n, smiles in enumerate(input_smiles):
print("data processing {}/{}".format(n, len(input_smiles)))
# Kekulized
smiles = get_smiles(get_mol(smiles))
mol = get_mol(smiles)
input_mols.append(mol)
if not motif:
clusters, atom_cls = find_clusters(mol)
for i,cls in enumerate(clusters):
clusters[i] = set(list(cls))
for a in range(len(atom_cls)):
atom_cls[a] = set(atom_cls[a])
else:
fragments = find_fragments(mol)
clusters = [frag[1] for frag in fragments]
# Construct graphs
subgraphs = []
subgraphs_idx_i = []
for i,cluster in enumerate(clusters):
_, subgraph_i_mapped, _ = extract_subgraph(smiles, cluster)
subgraphs.append(SubGraph(subgraph_i_mapped, mapping_to_input_mol=subgraph_i_mapped, subfrags=list(cluster)))
subgraphs_idx_i.append(list(cluster))
init_edge_flag += 1
init_subgraphs.append(subgraphs)
subgraphs_idx.append(subgraphs_idx_i)
graph = InputGraph(mol, smiles, subgraphs, subgraphs_idx_i, GNN_model_path)
input_graphs.append(graph)
input_graphs_dict[MolKey(graph.mol)] = graph
# Construct subgraph_set
subgraph_set = SubGraphSet(init_subgraphs, subgraphs_idx, input_graphs)
return subgraph_set, input_graphs_dict
def grammar_generation(agent, input_graphs_dict, subgraph_set, grammar, mcmc_iter, sample_number, args):
# Selected hyperedge (subgraph)
plist = [*subgraph_set.map_to_input]
# Terminating condition
if len(plist) == 0:
# done_flag, new_input_graphs_dict, new_subgraph_set, new_grammar
return True, input_graphs_dict, subgraph_set, grammar
# Update every InputGraph: remove every subgraph that equals to p_star, for those subgraphs that contain atom idx in p_star, replace the atom with p_star
org_input_graphs_dict = deepcopy(input_graphs_dict)
org_subgraph_set = deepcopy(subgraph_set)
org_grammar = deepcopy(grammar)
input_graphs_dict = deepcopy(org_input_graphs_dict)
subgraph_set = deepcopy(org_subgraph_set)
grammar = deepcopy(org_grammar)
for i, (key, input_g) in enumerate(input_graphs_dict.items()):
print("---for graph {}---".format(i))
action_list = []
all_final_features = []
# Skip the final iteration for training agent
if len(input_g.subgraphs) > 1:
for subgraph, subgraph_idx in zip(input_g.subgraphs, input_g.subgraphs_idx):
subg_feature = input_g.get_subg_feature_for_agent(subgraph)
num_occurance = subgraph_set.map_to_input[MolKey(subgraph)][1]
num_in_input = len(subgraph_set.map_to_input[MolKey(subgraph)][0].keys())
final_feature = []
final_feature.extend(subg_feature.tolist())
final_feature.append(1 - np.exp(-num_occurance))
final_feature.append(num_in_input / len(list(input_graphs_dict.keys())))
all_final_features.append(torch.unsqueeze(torch.from_numpy(np.array(final_feature)).float(), 0))
while(True):
action_list, take_action = sample(agent, torch.vstack(all_final_features), mcmc_iter, sample_number)
if take_action:
break
elif len(input_g.subgraphs) == 1:
action_list = [1]
else:
continue
print("Hyperedge sampling:", action_list)
# Merge connected hyperedges
p_star_list = input_g.merge_selected_subgraphs(action_list)
# Generate rules
for p_star in p_star_list:
is_inside, subgraphs, subgraphs_idx = input_g.is_candidate_subgraph(p_star)
if is_inside:
for subg, subg_idx in zip(subgraphs, subgraphs_idx):
if subg_idx not in input_g.subgraphs_idx:
# Skip the subg if it has been merged in previous iterations
continue
grammar = generate_rule(input_g, subg, grammar)
input_g.update_subgraph(subg_idx)
# Update subgraph_set
subgraph_set.update([g for (k, g) in input_graphs_dict.items()])
new_grammar = deepcopy(grammar)
new_input_graphs_dict = deepcopy(input_graphs_dict)
new_subgraph_set = deepcopy(subgraph_set)
return False, new_input_graphs_dict, new_subgraph_set, new_grammar
def MCMC_sampling(agent, all_input_graphs_dict, all_subgraph_set, all_grammar, sample_number, args):
iter_num = 0
while(True):
print("======MCMC iter{}======".format(iter_num))
done_flag, new_input_graphs_dict, new_subgraph_set, new_grammar = grammar_generation(agent, all_input_graphs_dict, all_subgraph_set, all_grammar, iter_num, sample_number, args)
print("Graph contraction status: ", done_flag)
if done_flag:
break
all_input_graphs_dict = deepcopy(new_input_graphs_dict)
all_subgraph_set = deepcopy(new_subgraph_set)
all_grammar = deepcopy(new_grammar)
iter_num += 1
return iter_num, new_grammar, new_input_graphs_dict
def random_produce(grammar):
def sample(l, prob=None):
if prob is None:
prob = [1/len(l)] * len(l)
idx = np.random.choice(range(len(l)), 1, p=prob)[0]
return l[idx], idx
def prob_schedule(_iter, selected_idx):
prob_list = []
# prob = exp(a * t * x), x = {0, 1}
a = 0.5
for rule_i, rule in enumerate(grammar.prod_rule_list):
x = rule.is_ending
if rule.is_start_rule:
prob_list.append(0)
else:
prob_list.append(np.exp(a * _iter * x))
prob_list = np.array(prob_list)[selected_idx]
prob_list = prob_list / np.sum(prob_list)
return prob_list
hypergraph = Hypergraph()
starting_rules = [(rule_i, rule) for rule_i, rule in enumerate(grammar.prod_rule_list) if rule.is_start_rule]
iter = 0
while(True):
if iter == 0:
_, idx = sample(starting_rules)
selected_rule_idx, selected_rule = starting_rules[idx]
hg_cand, _, avail = selected_rule.graph_rule_applied_to(hypergraph)
hypergraph = deepcopy(hg_cand)
else:
candidate_rule = []
candidate_rule_idx = []
candidate_hg = []
for rule_i, rule in enumerate(grammar.prod_rule_list):
hg_prev = deepcopy(hypergraph)
hg_cand, _, avail = rule.graph_rule_applied_to(hypergraph)
if(avail):
candidate_rule.append(rule)
candidate_rule_idx.append(rule_i)
candidate_hg.append(hg_cand)
if (all([rl.is_start_rule for rl in candidate_rule]) and iter > 0) or iter > 30:
break
prob_list = prob_schedule(iter, candidate_rule_idx)
hypergraph, idx = sample(candidate_hg, prob_list)
selected_rule = candidate_rule_idx[idx]
iter += 1
try:
mol = hg_to_mol(hypergraph)
print(Chem.MolToSmiles(mol))
except:
return None, iter
return mol, iter