-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
175 lines (154 loc) · 8.23 KB
/
train.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
import os
from time import time
import torch
import torch.optim as opt
from sklearn.model_selection import train_test_split
from torch.utils.data import DataLoader, TensorDataset
from torch.nn.functional import l1_loss
from einops import rearrange
from model.egnn import EGNN_Network
from model.molformer import Molformer
from utils.utils import parse_args, set_seed
from test import str_compute, select_sample
def main():
args = parse_args()
set_seed(args.seed)
args.bs = 512 * len(args.gpu.split(','))
args.lr = 1e-4 * len(args.gpu.split(','))
if args.data not in ['hamiltonian', 'newtonian']:
data_x, data_pos, data_y = torch.load(f'data/{args.data}.pt')
else:
data_x, data_y_node, data_y_graph = torch.load(f'data/hamiltonian.pt')
data_x = rearrange(data_x, 'ns nt n d -> (ns nt) n d')
if args.data == 'hamiltonian':
data_y = rearrange(data_y_graph, 'ns nt -> (ns nt)')
else:
data_y = rearrange(data_y_node, 'ns nt n d-> (ns nt) n d')
# egnn可以只输入one-hot，也可以输入多维特征
data_x, data_pos = data_x[..., 2:], data_x[..., :2]
args.n_sample *= 10
if args.data == 'qm9': data_y = data_y[:, args.qm9_index]
x_train, x_test, pos_train, pos_test, y_train, y_test = train_test_split(data_x, data_pos, data_y,
test_size=0.2, random_state=args.seed)
train_size = len(x_train)
test_size = len(x_test)
train_loader = DataLoader(TensorDataset(x_train, pos_train, y_train), batch_size=args.bs, shuffle=True)
test_loader = DataLoader(TensorDataset(x_test, pos_test, y_test), batch_size=args.bs * 2, shuffle=False)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
print(f'{"=" * 40} Training {"=" * 40}\nData: {args.data}; Model: {args.method}; KNN: {args.num_nearest};'
f' Train: {train_size}; Test: {test_size}\nGPU: {args.gpu}; Epoch: {args.ep}; Batch_size: {args.bs};'
f' Save: {args.save}\nMin_particle: {args.min_p}; Max_particle: {args.max_p}\n{"=" * 40} Start Training {"=" * 40}')
if args.data == 'qm9':
targets = ['mu', 'alpha', 'homo', 'lumo', 'gap']
print(f'Target: {targets[args.qm9_index]}')
if args.data in ['hamiltonian', 'newtonian']: args.tokens = None
if args.data in ['newtonian', 'md']: args.aggregate = False
if args.method == 'egnn':
model = EGNN_Network(num_tokens=args.tokens, dim=args.dim, depth=args.depth, num_nearest_neighbors=args.num_nearest,
norm_coors=True, coor_weights_clamp_value=2., aggregate=args.aggregate).cuda()
else:
# 模型size很小时，也可以取得较低的error，但和大模型还是有差距，且训练时间较长
model = Molformer(data=args.data, vocab=args.tokens, dist_bar=[0.8, 1.6, 3], depth=args.depth,
embed_dim=args.dim, aggregate=args.aggregate, knn=args.num_nearest).cuda()
if len(args.gpu) > 1: model = torch.nn.DataParallel(model)
criterion = torch.nn.L1Loss()
optimizer = opt.Adam(model.parameters(), lr=args.lr)
lr_scheduler = opt.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', factor=0.6, patience=10, min_lr=5e-6)
best_metric, t0, early_stop = 1e8, time(), 0
J_ = {'epoch': [], 'strength': [], 'k': []}
for epoch in range(0, args.ep):
if args.rewire and epoch % args.interval == 0:
print(f'Current K: {model.num_nearest_neighbors}')
if args.data in ['qm7', 'qm8']:
new_loader, _ = select_sample(args, model, train_loader, x_train, pos_train, y_train,
min_particle=args.min_p, max_particle=args.max_p)
else:
new_loader, _ = select_sample(args, model, train_loader, x_train, pos_train, y_train)
J = str_compute(args, model, new_loader, echo=False)
if args.data in ['qm7', 'qm8']:
J_tmp = [[i + 3, 0] for i in range(max([len(x) for x in J]))]
for i in range(1, len(J)):
for j in range(len(J[i])):
J_tmp[j][1] += J[i][j][1]
J = J_tmp
J_['strength'].append(J)
J_['k'].append(model.num_nearest_neighbors)
J_['epoch'].append(epoch)
# 暂时只考虑EGNN
if epoch > 0 and args.method == 'egnn':
# 限定qm7/qm8的原子数目相同，否则strength的增加有偏误，数据格式[[3, J_3], [4, J_4], ..., ]
J_delta = [[J[i][0], J[i][1] - J_['strength'][-2][i][1]] for i in
range(min(len(J), len(J_['strength'][-2])))]
print('The difference of J^{(m)}:', J_delta)
max_d = max([i[1] for i in J_delta])
if max_d > args.j_bar:
best_k = J_delta[[i[1] for i in J_delta].index(max_d)][0]
# 默认已找到最优K
if len(J_['k']) > 3:
if J_['k'][-1] == J_['k'][-2] and J_['k'][-2] == J_['k'][-3] and epoch > 100:
args.rewire = False
print('Finding the best K, and the rewiring stops.')
# 暂时只考虑EGNN
print(f'Changing K from *{model.num_nearest_neighbors}* to *{best_k}*')
model.num_nearest_neighbors = best_k
else:
print(f'{max_d} not reaching the threshold {args.j_bar}.')
model.train()
loss, t1, step = 0.0, time(), 0
for src_x, src_pos, y in train_loader:
src_x, src_pos, y = src_x.cuda(), src_pos.float().cuda(), y.cuda()
if args.data not in ['hamiltonian', 'newtonian']:
src_x = src_x.long()
src_mask = (src_x != 0)
else:
src_mask = (src_x[..., -1] != 0)
feat, pred = model(src_x, src_pos, mask=src_mask)
if args.data not in ['newtonian', 'md']:
loss_batch = criterion(pred[:, 0], y)
else:
loss_batch = criterion(pred, y)
loss += loss_batch.item() / train_size * len(src_x)
optimizer.zero_grad()
loss_batch.backward()
optimizer.step()
model.eval()
val_metric = 0
for tgt_x, tgt_pos, y in test_loader:
tgt_x, tgt_pos, y = tgt_x.cuda(), tgt_pos.float().cuda(), y.cuda()
if args.data not in ['hamiltonian', 'newtonian']:
tgt_x = tgt_x.long()
tgt_mask = (tgt_x != 0)
else:
tgt_mask = (tgt_x[..., -1] != 0)
with torch.no_grad():
feat, pred = model(tgt_x, tgt_pos, mask=tgt_mask)
if args.data not in ['newtonian', 'md']:
val_metric += l1_loss(pred[..., 0], y).item() / test_size * len(tgt_x)
else:
val_metric += l1_loss(pred, y).item() / test_size * len(tgt_x)
print('Epoch: {} | Time: {:.1f}s | Loss: {:.3f} | Val_Metric: {:.3f} | Lr: {:.3f}'
.format(epoch + 1, time() - t1, loss, val_metric, optimizer.param_groups[0]['lr'] * 1e5))
lr_scheduler.step(val_metric)
if val_metric < best_metric:
best_ep = epoch
best_metric = val_metric
best_model = model
early_stop = 0
else:
early_stop += 1
# 为了效果最好，测试集可见
if early_stop >= 30:
print('Early Stopping!!! No Improvement on metric for 30 Epochs.')
break
print(f'{"=" * 20} End Training (Time: {(time() - t0) / 3600:.2f}h) {"=" * 20}\n'
f'Best Metric for {args.method} in {args.data}: {best_metric} (epoch: {best_ep})')
if len(J_['k']) == 0:
checkpoint = {'model': best_model.state_dict(), 'error': best_metric}
if len(args.gpu) > 1: checkpoint['model'] = best_model.module.state_dict()
if not os.path.exists('save/model_weight/'): os.makedirs('save/model_weight/')
torch.save(checkpoint, f'save/model_weight/model_{args.data}_{args.method}.pt')
else:
if not os.path.exists('save/J_/'): os.makedirs('save/J_/')
torch.save(J_, f'save/J_/J_{args.data}_{args.method}.pt')
if __name__ == '__main__':
main()