-
Notifications
You must be signed in to change notification settings - Fork 5
/
traversal_step.py
212 lines (189 loc) · 6.25 KB
/
traversal_step.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
201
202
203
204
205
206
207
208
209
210
211
212
import torch
from torch import Tensor
import numpy as np
import random
from cd2root import cd2root
cd2root()
from src.utils.scores import *
from src.vae import load_vae
from src.pinn.pde import load_wavepde
from src.pinn import PropGenerator, VAEGenerator
from src.predictor import Predictor
MODES = [
"random",
"random_1d",
"fp",
"limo",
"chemspace",
"wave_sup",
"wave_unsup",
"hj_sup",
"hj_unsup",
]
WAVEPDE_IDX_MAP = {
"plogp": 1,
"sa": 1,
"qed": 1,
"drd2": 9,
"jnk3": 4,
"gsk3b": 0,
"uplogp": 1,
"1err": 2,
"2iik": 4,
}
HJPDE_IDX_MAP = {
"plogp": 0,
"sa": 0,
"qed": 9,
"drd2": 2,
"jnk3": 3,
"gsk3b": 8,
"uplogp": 0,
"1err": 6,
"2iik": 3,
}
class Traversal:
"""
Uniformed class to perform 1 step of traversal in latent space
"""
method: str
prop: str
data_name: str
step_size: float
relative: bool
minimize: bool
device: torch.device
def __init__(
self,
method: str,
prop: str,
data_name: str = "zmc",
step_size: float = 0.1,
relative: bool = True,
minimize: bool = False,
k_idx: int | None = None, # the index of unsupervised pde to use
device: torch.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu"
),
):
self.method = method
self.prop = prop
self.data_name = data_name
self.step_size = step_size
self.relative = relative
self.minimize = minimize
self.device = device
assert self.method in MODES, f"mode must be one of {MODES}"
self.dm, self.vae = load_vae(
file_path=f"data/processed/{self.data_name}.smi",
model_path=f"checkpoints/vae/{self.data_name}/checkpoint.pt",
latent_dim=1024,
embedding_dim=128,
device=self.device,
)
# Generate u_z for random
if self.method == "random":
self.u_z = torch.randn(self.vae.latent_dim, device=self.device)
return
elif self.method == "random_1d":
self.u_z = torch.zeros(self.vae.latent_dim, device=self.device)
self.u_z[random.randint(0, self.vae.latent_dim - 1)] = (
1 if random.random() < 0.5 else -1
)
return
elif self.method == "chemspace":
# load boundary for chemspace
boundary = np.load(
f"src/chemspace/boundaries_{self.data_name}/boundary_{self.prop}.npy"
) # (1, latent_dim)
self.u_z = torch.tensor(boundary, device=self.device).squeeze()
return
if self.method in {"limo", "fp", "wave_sup", "hj_sup"}:
self.predictor = Predictor(self.dm.max_len * self.dm.vocab_size)
self.predictor.load_state_dict(
torch.load(
f"checkpoints/prop_predictor/{self.prop}/checkpoint.pt",
map_location=self.device,
)
)
for p in self.predictor.parameters():
p.requires_grad = False
# LIMO and FP don't need to load the generator
if self.method in {"limo", "fp"}:
self.generator = PropGenerator(self.vae, self.predictor).to(self.device)
return
# All the other methods are pde related
pde_name = self.method.split("_")[0]
if self.method in {"wave_sup", "hj_sup"}:
self.generator = PropGenerator(self.vae, self.predictor).to(self.device)
self.pde = load_wavepde(
checkpoint=f"checkpoints/{pde_name}pde_prop/{self.data_name}/{self.prop}/checkpoint.pt",
generator=self.generator,
k=1,
device=self.device,
)
self.idx = 0
else:
self.generator = VAEGenerator(self.vae).to(self.device)
self.pde = load_wavepde(
checkpoint=f"checkpoints/{pde_name}pde/{self.data_name}/checkpoint.pt",
generator=self.generator,
k=10,
device=self.device,
)
if k_idx is not None:
self.idx = k_idx
elif pde_name == "wave":
self.idx = WAVEPDE_IDX_MAP[self.prop]
elif pde_name == "hj":
self.idx = HJPDE_IDX_MAP[self.prop]
else:
raise ValueError(f"Unknown pde {pde_name}")
for p in self.pde.parameters():
p.requires_grad = False
self.k = self.pde.k
self.half_range = self.pde.half_range
def step(self, z: Tensor, t: int = 0, optimizer=None) -> Tensor:
"""
Perform 1 step of traversal in latent space, return u_z
When t=0, return 0 tensor
"""
if t == 0:
return torch.zeros_like(z)
if self.method in ["random", "random_1d"]:
u_z = self.u_z
u_z = normalize(u_z, self.step_size, self.relative)
elif self.method == "chemspace":
u_z = self.u_z
u_z = normalize(u_z, self.step_size, self.relative)
if self.minimize:
u_z = -u_z
elif self.method == "limo":
if optimizer is not None:
return self.limo_optimizer_step(optimizer, z)
z = z.detach().requires_grad_(True)
u_z = torch.autograd.grad(self.generator(z).sum(), z)[0]
u_z = normalize(u_z, self.step_size, self.relative)
if self.minimize:
u_z = -u_z
elif self.method == "fp":
z = z.detach().requires_grad_(True)
u_z = torch.autograd.grad(self.generator(z).sum(), z)[0]
u_z = (
u_z * self.step_size
+ torch.randn_like(u_z) * np.sqrt(2 * self.step_size) * 0.1
)
if self.minimize:
u_z = -u_z
else:
_, u_z = self.pde.inference(self.idx, z, t % self.half_range)
u_z = normalize(u_z, self.step_size, self.relative)
return u_z
def limo_optimizer_step(self, optimizer, z):
optimizer.zero_grad()
u = -self.generator(z).sum()
if self.minimize:
u = -u
u.backward()
optimizer.step()
return z.grad