-
Notifications
You must be signed in to change notification settings - Fork 1
/
new_optimization_nonfrag.py
207 lines (175 loc) · 9.93 KB
/
new_optimization_nonfrag.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
# -*- coding: utf-8 -*-
# @Time : 2022/4/10 11:36
# @Author : FAN FAN
# @Site :
# @File : optimization.py
# @Software: PyCharm
from src.feature.atom_featurizer import classic_atom_featurizer
from src.feature.bond_featurizer import classic_bond_featurizer
from src.feature.mol_featurizer import classic_mol_featurizer
from utils.mol2graph import smiles_2_bigraph
from utils.junctiontree_encoder import JT_SubGraph
from utils.splitter import Splitter
from utils.metrics import Metrics
from utils.Earlystopping import EarlyStopping
from data.csv_dataset import MoleculeCSVDataset
from src.dgltools import collate_molgraphs, collate_fraggraphs
from data.dataloading import import_dataset
import torch
from torch.utils.data import DataLoader
import os
import time
import numpy as np
import pandas as pd
from tqdm import tqdm
from utils.count_parameters import count_parameters
from networks.DMPNN import DMPNNNet
from networks.MPNN import MPNNNet
from networks.AttentiveFP import AttentiveFPNet
from networks.AGC import AGCNet
from networks.FraGAT import NewFraGATNet
from utils.Set_Seed_Reproducibility import set_seed
from utils.piplines import train_epoch, evaluate, train_epoch_frag, evaluate_frag, PreFetch
from bayes_opt import BayesianOptimization, SequentialDomainReductionTransformer
params = {}
net_params = {}
#params['Dataset'] = 'HFUS'
params['init_lr'] = 1e-3
params['min_lr'] = 1e-9
params['weight_decay'] = 0
params['lr_reduce_factor'] = 0.8
params['lr_schedule_patience'] = 30
params['earlystopping_patience'] = 50
params['max_epoch'] = 300
net_params['num_atom_type'] = 36
net_params['num_bond_type'] = 12
net_params['hidden_dim'] = 16
net_params['dropout'] = 0
net_params['depth'] = 2
net_params['residual'] = False
net_params['batch_norm'] = False
net_params['layer_norm'] = False
net_params['device'] = 'cuda'
dataset_list = ['BMP']
#splitting_list = [3807, 146, 1945, 3560] DMPNN
#splitting_list = [27, 1538, 848, 4113] MPNN
splitting_list = [3348] #AGC
def main(params, net_params):
#set_seed(seed=1000)
model = MPNNNet(net_params).to(device='cuda')
optimizer = torch.optim.Adam(model.parameters(), lr=params['init_lr'], weight_decay=params['weight_decay'])
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=params['lr_reduce_factor'],
patience=params['lr_schedule_patience'], verbose=False)
t0 = time.time()
per_epoch_time = []
early_stopping = EarlyStopping(patience=params['earlystopping_patience'], path='checkpoint1' + params['Dataset'] + 'AFP' + '.pt')
with tqdm(range(params['max_epoch'])) as t:
n_param = count_parameters(model)
for epoch in t:
t.set_description('Epoch %d' % epoch)
start = time.time()
model, epoch_train_loss, epoch_train_metrics = train_epoch(model, optimizer, scaling,
fetched_data.train_iter, fetched_data.train_batched_origin_graph_list,
fetched_data.train_targets_list,
fetched_data.train_smiles_list,
fetched_data.train_names_list, n_param)
epoch_val_loss, epoch_val_metrics = evaluate(model, scaling, fetched_data.val_iter, fetched_data.val_batched_origin_graph_list,
fetched_data.val_targets_list, fetched_data.val_smiles_list, fetched_data.val_names_list, n_param)
t.set_postfix({'time': time.time() - start, 'lr': optimizer.param_groups[0]['lr'],
'train_loss': epoch_train_loss, 'val_loss': epoch_val_loss,
'train_R2': epoch_train_metrics.R2, 'val_R2': epoch_val_metrics.R2})
per_epoch_time.append(time.time() - start)
scheduler.step(epoch_val_loss)
if optimizer.param_groups[0]['lr'] < params['min_lr']:
print('\n! LR equal to min LR set.')
break
early_stopping(epoch_val_loss, model)
if early_stopping.early_stop:
break
model = early_stopping.load_checkpoint(model)
_, epoch_train_metrics = evaluate(model, scaling, fetched_data.train_iter, fetched_data.train_batched_origin_graph_list,
fetched_data.train_targets_list, fetched_data.train_smiles_list, fetched_data.train_names_list, n_param)
_, epoch_val_metrics = evaluate(model, scaling, fetched_data.val_iter, fetched_data.val_batched_origin_graph_list,
fetched_data.val_targets_list, fetched_data.val_smiles_list, fetched_data.val_names_list, n_param)
_, epoch_test_metrics = evaluate(model, scaling, fetched_data.test_iter, fetched_data.test_batched_origin_graph_list,
fetched_data.test_targets_list, fetched_data.test_smiles_list, fetched_data.test_names_list, n_param)
_, epoch_raw_metrics = evaluate(model, scaling, fetched_data.all_iter, fetched_data.all_batched_origin_graph_list,
fetched_data.all_targets_list, fetched_data.all_smiles_list, fetched_data.all_names_list, n_param)
return epoch_train_metrics, epoch_val_metrics, epoch_test_metrics, epoch_raw_metrics
def Splitting_Main_MO(params, net_params):
"""Built-in function. Use the basic block of implementation in Multi-objective Bayesian Optimization.
Parameters
----------
params : dict
Set of parameters for workflow.
net_params : dict
Set of parameters for architectures of models.
Returns
----------
-train_metrics.RMSE : float
Optimization Objective-1, negative number of RMSE metric in training
-val_metrics.RMSE : float
Optimization Objective-2, negative number of RMSE metric in validation
"""
train_metrics, val_metrics, test_metrics, all_metrics = main(params, net_params)
return -val_metrics.RMSE
def func_to_be_opt_MO(hidden_dim, depth, decay, init_lr, lr_reduce_factor, dropout):
"""Built-in function. Objective function in Multi-objective Bayesian Optimization.
Parameters
----------
x : list
Set of hyper-parameters in search domain.
Returns
----------
"""
net_params['hidden_dim'] = int(hidden_dim)
net_params['depth'] = int(depth)
params['weight_decay'] = 0.05 * int(decay)
params['init_lr'] = 10 ** init_lr
params['lr_reduce_factor'] = 0.4 + 0.05 * int(lr_reduce_factor)
net_params['dropout'] = 0.05 * int(dropout)
return Splitting_Main_MO(params, net_params)
set_seed(seed=1000)
for i in range(len(dataset_list)):
params['Dataset'] = dataset_list[i]
df, scaling = import_dataset(params)
cache_file_path = os.path.realpath('./cache')
if not os.path.exists(cache_file_path):
os.mkdir(cache_file_path)
cache_file = os.path.join(cache_file_path, params['Dataset'] + '_CCC')
error_path = os.path.realpath('./error_log')
if not os.path.exists(error_path):
os.mkdir(error_path)
error_log_path = os.path.join(error_path, '{}_{}'.format(params['Dataset'], time.strftime('%Y-%m-%d-%H-%M')) + '.csv')
fragmentation = JT_SubGraph()
net_params['frag_dim'] = fragmentation.frag_dim
dataset = MoleculeCSVDataset(df, smiles_2_bigraph, classic_atom_featurizer, classic_bond_featurizer, classic_mol_featurizer, cache_file, load=True
, error_log=error_log_path)
splitter = Splitter(dataset)
rows = []
file_path = os.path.realpath('./output')
if not os.path.exists(file_path):
os.mkdir(file_path)
save_file_path = os.path.join(file_path, '{}_{}_{}_{}_{}'.format('SOPT_UCB_5', params['Dataset'], 'MPNN', splitting_list[i], time.strftime('%Y-%m-%d-%H-%M')) + '.csv')
#df = pd.DataFrame(columns=['seed', 'train_R2', 'val_R2', 'test_R2', 'all_R2', 'train_MAE', 'val_MAE', 'test_MAE', 'all_MAE',
# 'train_RMSE', 'val_RMSE', 'test_RMSE', 'all_RMSE'])
train_set, val_set, test_set, raw_set = splitter.Random_Splitter(seed=splitting_list[i], frac_train=0.8, frac_val=0.1)
train_loader = DataLoader(train_set, collate_fn=collate_molgraphs, batch_size=len(train_set), shuffle=False, num_workers=0, worker_init_fn=np.random.seed(1000))
val_loader = DataLoader(val_set, collate_fn=collate_molgraphs, batch_size=len(val_set), shuffle=False, num_workers=0, worker_init_fn=np.random.seed(1000))
test_loader = DataLoader(test_set, collate_fn=collate_molgraphs, batch_size=len(test_set), shuffle=False, num_workers=0, worker_init_fn=np.random.seed(1000))
raw_loader = DataLoader(raw_set, collate_fn=collate_molgraphs, batch_size=len(raw_set), shuffle=False, num_workers=0, worker_init_fn=np.random.seed(1000))
fetched_data = PreFetch(train_loader, val_loader, test_loader, raw_loader, frag=False)
hpbounds = {'hidden_dim': (16, 128), 'depth': (1, 6.99), 'decay': (0, 8.99),
'init_lr': (-5, -1), 'lr_reduce_factor': (0, 10.99), 'dropout': (0, 8.99)}
bounds_transformer = SequentialDomainReductionTransformer()
mutating_optimizer = BayesianOptimization(f=func_to_be_opt_MO, pbounds=hpbounds, verbose=0, random_state=1000)
mutating_optimizer.maximize(init_points=50, n_iter=300, acq='ucb', kappa=5)
lst = mutating_optimizer.space.keys
lst.append('target')
df = pd.DataFrame(columns=lst)
for i, res in enumerate(mutating_optimizer.res):
_dict = res['params']
_dict['target'] = res['target']
row = pd.DataFrame(_dict, index=[0])
df = pd.concat([df, row], axis=0, ignore_index=True, sort=False)
df.to_csv(save_file_path, index=False)