forked from MolFilterGAN/MolFilterGAN
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathG_sample.py
200 lines (188 loc) · 8.47 KB
/
G_sample.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
190
191
192
193
194
195
196
197
198
199
200
import os
import pandas as pd
import numpy as np
import rdkit
from rdkit import Chem
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.utils import clip_grad_norm_
from torch.utils.data import DataLoader
from tils import Vocabulary, rm_voc_less, construct_voc
from ataset import MolData,MolData2
from Model import Generator, Discriminator
import tensorboardXUDM
from tensorboardX import SummaryWriter
from sklearn.metrics import precision_score, recall_score, f1_score
import argparse
class GSample:
def __init__(self, emb_size=128, hidden_size=512, num_layers=3, dropout=0.5,
n_epochs=100, lr=0.001,
load_dir=None, save_dir=None, log_dir=None,
log_every=100, save_every=500, voc=None, device=None):
self.voc = voc
self.generator = Generator(self.voc, emb_size=emb_size, hidden_size=hidden_size,
num_layers=num_layers, dropout=dropout)
self.n_epochs = n_epochs
self.lr = lr
self.save_dir = save_dir
self.log_dir = log_dir
self.log_every = log_every
self.save_every = save_every
if self.log_dir:
self.writer = SummaryWriter(self.log_dir, flush_secs=10)
if device:
self.device = torch.device(device)
self.generator = self.generator.to(self.device)
else:
self.device = device
# Can restore from a saved RNN
if load_dir:
checkpoint = torch.load(load_dir)
self.generator.load_state_dict(checkpoint['generator_state_dict'])
if self.save_dir:
if not os.path.exists(self.save_dir):
os.makedirs(self.save_dir)
else:
raise Exception('%s already exist'%self.save_dir)
def fit(self, train_set, valid_set):
criterion = nn.CrossEntropyLoss(ignore_index=self.voc.vocab['PAD'])
optimizer = torch.optim.Adam(self.generator.parameters(), lr=self.lr)
global_step = 0
for epoch in range(1, self.n_epochs+1):
for train_step, batch in enumerate(train_set):
global_step += 1
tensors, prevs, nexts, lens = batch
if self.device:
prevs = prevs.to(self.device)
nexts = nexts.to(self.device)
lens = lens.to(self.device)
outputs, _, _ =self.generator(prevs, lens)
loss = criterion(outputs.view(-1, outputs.shape[-1]), nexts.view(-1))
optimizer.zero_grad()
loss.backward()
clip_grad_norm_(self.generator.parameters(), 5.)
optimizer.step()
if global_step % self.log_every ==0 or global_step == 1:
self.generator.eval()
train_out_metrics = self.evaluate(train_set)
valid_out_metrics = self.evaluate(valid_set)
for metric in train_out_metrics:
self.writer.add_scalars('%s'%metric, {'train': train_out_metrics[metric]}, global_step)
self.writer.add_scalars('%s'%metric, {'valid': valid_out_metrics[metric]}, global_step)
out_smiles_ls = self.sample(n_batch=128)
print(out_smiles_ls)
valid = 0
for smi_i in out_smiles_ls:
try:
mol = Chem.MolFromSmiles(smi_i)
if mol:
valid += 1
except:
continue
self.writer.add_scalar('valid_smiles_rate', round(100 * valid / 128, 2), global_step)
self.generator.train()
if global_step % self.save_every == 0 or global_step == 1:
torch.save({
'epoch': epoch,
'global_step': global_step,
'generator_state_dict': self.generator.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}, os.path.join(self.save_dir, 'G_epoch_%s_step_%s.ckpt'%(epoch, global_step)))
self.writer.close()
def evaluate(self, valid_set, some_metrics=None, rl=False):
criterion = nn.CrossEntropyLoss(ignore_index=self.voc.vocab['PAD'])
with torch.no_grad():
out_metrics ={}
for eval_step, batch in enumerate(valid_set):
tensors, prevs, nexts, lens = batch
# print(tensors[0])
# print(self.voc.decode(tensors[0].numpy()))
if self.device:
prevs = prevs.to(self.device)
nexts = nexts.to(self.device)
lens = lens.to(self.device)
outputs, _, _ = self.generator(prevs, lens)
# print(outputs[0].cpu().numpy())
# print(self.voc.decode([np.argmax(i) for i in outputs[0].cpu().numpy()]))
loss = criterion(outputs.view(-1, outputs.shape[-1]), nexts.view(-1))
out_metrics['loss'] = loss
return out_metrics
def sample(self,n_batch,max_len=100):
with torch.no_grad():
prevs = torch.empty(n_batch, 1, dtype=torch.long, device=self.device).fill_(self.voc.vocab['GO'])
n_sequences = prevs.shape[0]
sequences = []
lengths = torch.zeros(n_sequences, dtype=torch.long, device=prevs.device)
one_lens = torch.ones(n_sequences, dtype=torch.long, device=prevs.device)
is_end = prevs.eq(self.voc.vocab['EOS']).view(-1)
states = None
for _ in range(max_len):
outputs, _, states = self.generator(prevs, one_lens, states)
probs = F.softmax(outputs, dim=-1).view(n_sequences, -1)
currents = torch.multinomial(probs, 1)
currents[is_end, :] = self.voc.vocab['PAD']
sequences.append(currents)
lengths[~is_end] += 1
is_end[currents.view(-1) == self.voc.vocab['EOS']] = 1
if is_end.sum() == n_sequences:
break
prevs = currents
sequences = torch.cat(sequences, dim=-1)
out_smiles = []
seq_vec_ls = sequences.tolist()
for seq_vec in seq_vec_ls:
smile = self.voc.decode(seq_vec)
out_smiles.append(smile)
return out_smiles
def get_parser():
parser = argparse.ArgumentParser(
"Sampling smiles from a generator"
)
parser.add_argument(
'--outfile_path', type=str, default=None, help='Path to the output sampled smiles'
)
parser.add_argument(
'--num', type=float, default=128, help='Number of smiles to sample'
)
parser.add_argument(
'--voc_path', type=str, default='./Datasets/Voc', help='Path to the Vocabulary'
)
parser.add_argument(
'--visible_gpu', type=str, default='0', help='Visible GPU ids'
)
parser.add_argument(
'--load_dir', type=str, default=None, help='Path to load model'
)
parser.add_argument(
'--random_seed', type=float, default=666, help='Random seed for pytorch'
)
return parser
############
#sample
##############
if __name__ == "__main__":
parser = get_parser()
config, unknown = parser.parse_known_args()
os.environ["CUDA_VISIBLE_DEVICES"]=config.visible_gpu
# torch.manual_seed(config.random_seed)
voc_path = config.voc_path
voc = Vocabulary(init_from_file=voc_path, max_length=140)
esti = GSample(emb_size=128, hidden_size=512, num_layers=3, dropout=0.5,
n_epochs=100, lr=0.0001,
load_dir=config.load_dir, save_dir=None, log_dir=None,
log_every=100, save_every=1000, voc=voc, device='cuda:0')
out_smiles=[]
while True:
tmp_smi_ls=esti.sample(128)
tmp_valid_smi_ls=[]
for smi in tmp_smi_ls:
mol=Chem.MolFromSmiles(smi)
if mol:
tmp_valid_smi_ls.append(Chem.MolToSmiles(mol))
out_smiles.extend(tmp_valid_smi_ls)
if len(out_smiles)>=config.num:
break
with open(config.outfile_path,'w') as f:
for smi_t in out_smiles[:int(config.num)]:
f.write(smi_t+'\n')