-
Notifications
You must be signed in to change notification settings - Fork 0
/
full_model.py
153 lines (118 loc) · 4.95 KB
/
full_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
# FULL MODEL
import numpy as np
import torch
import torch.nn.functional as F
from rdkit import Chem, RDLogger
from torch import nn
from torch.nn.utils import clip_grad_value_
from torch.utils.data import DataLoader
from torch.distributions import Categorical
from torch.nn.modules.activation import Sigmoid
from allennlp.modules.feedforward import FeedForward
from allennlp.modules.seq2seq_encoders import (LstmSeq2SeqEncoder,PytorchTransformer)
from tqdm import tqdm
from tokenizer import *
from train import *
from discriminator import *
from generator import *
class MolGen(nn.Module):
def __init__(self, data, hidden_dim=128, lr=1e-3, device='cpu'):
super().__init__()
self.device = device
self.hidden_dim = hidden_dim
self.tokenizer = Tokenizer(data)
self.generator = Generator(
latent_dim=hidden_dim,
vocab_size=self.tokenizer.vocab_size - 1,
start_token=self.tokenizer.start_token - 1, # no need token
end_token=self.tokenizer.end_token - 1,
).to(device)
self.discriminator = RecurrentDiscriminator(
hidden_size=hidden_dim,
vocab_size=self.tokenizer.vocab_size,
start_token=self.tokenizer.start_token,
bidirectional=True
).to(device)
self.generator_optim = torch.optim.Adam(
self.generator.parameters(), lr=lr)
self.discriminator_optim = torch.optim.Adam(
self.discriminator.parameters(), lr=lr)
self.b = 0.
def sample_latent(self, batch_size):
return torch.randn(batch_size, self.hidden_dim).to(self.device)
def discriminator_loss(self, x, y):
y_pred, mask = self.discriminator(x).values()
loss = F.binary_cross_entropy(
y_pred, y, reduction='none') * mask
loss = loss.sum() / mask.sum()
return loss
def train_step(self, x):
batch_size, len_real = x.size()
x_real = x.to(self.device)
y_real = torch.ones(batch_size, len_real).to(self.device)
z = self.sample_latent(batch_size)
generator_outputs = self.generator.forward(z, max_len=20)
x_gen, log_probs, entropies = generator_outputs.values()
_, len_gen = x_gen.size()
y_gen = torch.zeros(batch_size, len_gen).to(self.device)
# D Train
self.discriminator_optim.zero_grad()
fake_loss = self.discriminator_loss(x_gen, y_gen)
real_loss = self.discriminator_loss(x_real, y_real)
discr_loss = 0.5 * (real_loss + fake_loss)
discr_loss.backward()
clip_grad_value_(self.discriminator.parameters(), 0.1)
self.discriminator_optim.step()
# G Train
self.generator_optim.zero_grad()
y_pred, y_pred_mask = self.discriminator(x_gen).values()
R = (2 * y_pred - 1)
lengths = y_pred_mask.sum(1).long()
list_rewards = [rw[:ln] for rw, ln in zip(R, lengths)]
generator_loss = []
for reward, log_p in zip(list_rewards, log_probs):
reward_baseline = reward - self.b
generator_loss.append((- reward_baseline * log_p).sum())
generator_loss = torch.stack(generator_loss).mean() - \
sum(entropies) * 0.01 / batch_size
with torch.no_grad():
mean_reward = (R * y_pred_mask).sum() / y_pred_mask.sum()
self.b = 0.9 * self.b + (1 - 0.9) * mean_reward
generator_loss.backward()
clip_grad_value_(self.generator.parameters(), 0.1)
self.generator_optim.step()
return {'loss_disc': discr_loss.item(), 'mean_reward': mean_reward}
def create_dataloader(self, data, batch_size=128, shuffle=True, num_workers=5):
return DataLoader(
data,
batch_size=batch_size,
shuffle=shuffle,
collate_fn=self.tokenizer.batch_tokenize,
num_workers=num_workers
)
def train_n_steps(self, train_loader, max_step=5000, evaluate_every=1000):
iter_loader = iter(train_loader)
for step in tqdm(range(max_step)):
try:
batch = next(iter_loader)
except:
iter_loader = iter(train_loader)
batch = next(iter_loader)
self.train_step(batch)
if step % evaluate_every == 0:
self.eval()
score = self.evaluate_n(evaluate_every)
self.train()
print(f'valid = {score: .2f}')
def get_mapped(self, seq):
return ''.join([self.tokenizer.inv_mapping[i] for i in seq])
@torch.no_grad()
def generate_n(self, n):
z = torch.randn((n, self.hidden_dim)).to(self.device)
x = self.generator(z)['x'].cpu()
lenghts = (x > 0).sum(1)
return [self.get_mapped(x[:l - 1].numpy()) for x, l in zip(x, lenghts)]
def evaluate_n(self, n):
pack = self.generate_n(n)
valid = np.array([Chem.MolFromSmiles(k) is not None for k in pack])
return valid.mean()