-
Notifications
You must be signed in to change notification settings - Fork 5
/
test_transfer_finetune_chem.py
232 lines (185 loc) · 8.98 KB
/
test_transfer_finetune_chem.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
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
import argparse
import logging
import random
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.optim as optim
from sklearn.metrics import roc_auc_score
from torch_geometric.data import DataLoader
from tqdm import tqdm
from datasets import MoleculeDataset
from transfer.model import GNN_graphpred
from transfer.utils import scaffold_split
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
np.random.seed(seed)
random.seed(seed)
def train(model, device, loader, optimizer, criterion):
model.train()
loss_all = 0
for step, batch in enumerate(loader):
batch = batch.to(device)
pred = model(batch.x, batch.edge_index, batch.edge_attr, batch.batch)
y = batch.y.view(pred.shape).to(torch.float64)
# Whether y is non-null or not.
is_valid = y ** 2 > 0
# Loss matrix
loss_mat = criterion(pred.double(), (y + 1) / 2)
# loss matrix after removing null target
loss_mat = torch.where(is_valid, loss_mat, torch.zeros(loss_mat.shape).to(loss_mat.device).to(loss_mat.dtype))
optimizer.zero_grad()
loss = torch.sum(loss_mat) / torch.sum(is_valid)
loss.backward()
optimizer.step()
loss_all += loss.item() * batch.num_graphs
return loss_all/len(loader)
def eval(model, device, loader):
model.eval()
y_true = []
y_scores = []
for step, batch in enumerate(loader):
batch = batch.to(device)
with torch.no_grad():
pred = model(batch.x, batch.edge_index, batch.edge_attr, batch.batch)
y_true.append(batch.y.view(pred.shape))
y_scores.append(pred)
y_true = torch.cat(y_true, dim=0).cpu().numpy()
y_scores = torch.cat(y_scores, dim=0).cpu().numpy()
roc_list = []
for i in range(y_true.shape[1]):
# AUC is only defined when there is at least one positive data.
if np.sum(y_true[:, i] == 1) > 0 and np.sum(y_true[:, i] == -1) > 0:
is_valid = y_true[:, i] ** 2 > 0
roc_list.append(roc_auc_score((y_true[is_valid, i] + 1) / 2, y_scores[is_valid, i]))
if len(roc_list) < y_true.shape[1]:
print("Some target is missing!")
print("Missing ratio: {}".format(1 - float(len(roc_list)) / y_true.shape[1]))
return sum(roc_list) / len(roc_list) # y_true.shape[1]
def arg_parse():
parser = argparse.ArgumentParser(description='Finetuning Chem after pre-training of graph neural networks')
parser.add_argument('--device', type=int, default=0,
help='which gpu to use if any (default: 0)')
parser.add_argument('--batch_size', type=int, default=32,
help='input batch size for training (default: 32)')
parser.add_argument('--epochs', type=int, default=100,
help='number of epochs to train (default: 100)')
parser.add_argument('--lr', type=float, default=0.001,
help='learning rate (default: 0.001)')
parser.add_argument('--lr_scale', type=float, default=1,
help='relative learning rate for the feature extraction layer (default: 1)')
parser.add_argument('--decay', type=float, default=0,
help='weight decay (default: 0)')
parser.add_argument('--num_layer', type=int, default=5,
help='number of GNN message passing layers (default: 5).')
parser.add_argument('--emb_dim', type=int, default=300,
help='embedding dimensions (default: 300)')
parser.add_argument('--dropout_ratio', type=float, default=0.5,
help='dropout ratio (default: 0.5)')
parser.add_argument('--graph_pooling', type=str, default="mean",
help='graph level pooling (sum, mean, max, set2set, attention)')
parser.add_argument('--JK', type=str, default="last",
help='how the node features across layers are combined. last, sum, max or concat')
parser.add_argument('--gnn_type', type=str, default="gin")
parser.add_argument('--dataset', type=str, default='tox21',
help='dataset. For now, only classification.')
parser.add_argument('--input_model_file', type=str, default='',
help='filename to read the pretrain model (if there is any)')
parser.add_argument('--seed', type=int, default=0, help="Seed for minibatch selection, random initialization.")
parser.add_argument('--split', type=str, default="scaffold", help="random or scaffold or random_scaffold")
parser.add_argument('--eval_train', type=int, default=1, help='evaluating training or not')
parser.add_argument('--num_workers', type=int, default=4, help='number of workers for dataset loading')
return parser.parse_args()
def run(args):
# Training settings
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(message)s', datefmt='%d-%b-%y %H:%M:%S')
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
logging.info("Using Device: %s" % device)
logging.info("Seed: %d" % args.seed)
logging.info(args)
setup_seed(args.seed)
# Bunch of classification tasks
if args.dataset == "tox21":
num_tasks = 12
elif args.dataset == "hiv":
num_tasks = 1
elif args.dataset == "pcba":
num_tasks = 128
elif args.dataset == "muv":
num_tasks = 17
elif args.dataset == "bace":
num_tasks = 1
elif args.dataset == "bbbp":
num_tasks = 1
elif args.dataset == "toxcast":
num_tasks = 617
elif args.dataset == "sider":
num_tasks = 27
elif args.dataset == "clintox":
num_tasks = 2
else:
raise ValueError("Invalid dataset name.")
# set up dataset
dataset = MoleculeDataset("original_datasets/transfer/" + args.dataset, dataset=args.dataset)
logging.info(dataset)
if args.split == "scaffold":
smiles_list = pd.read_csv('original_datasets/transfer/' + args.dataset + '/processed/smiles.csv', header=None)[0].tolist()
train_dataset, valid_dataset, test_dataset = scaffold_split(dataset, smiles_list, null_value=0, frac_train=0.8,
frac_valid=0.1, frac_test=0.1)
logging.info("scaffold")
else:
raise ValueError("Invalid split option.")
logging.info(train_dataset[0])
train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers)
val_loader = DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers)
test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers)
# set up model
model = GNN_graphpred(args.num_layer, args.emb_dim, num_tasks, JK=args.JK, drop_ratio=args.dropout_ratio,
graph_pooling=args.graph_pooling, gnn_type=args.gnn_type)
if not args.input_model_file == "":
model.from_pretrained(args.input_model_file)
model.to(device)
criterion = nn.BCEWithLogitsLoss(reduction="none")
# set up optimizer
# different learning rate for different part of GNN
model_param_group = []
model_param_group.append({"params": model.gnn.parameters()})
if args.graph_pooling == "attention":
model_param_group.append({"params": model.pool.parameters(), "lr": args.lr * args.lr_scale})
model_param_group.append({"params": model.graph_pred_linear.parameters(), "lr": args.lr * args.lr_scale})
optimizer = optim.Adam(model_param_group, lr=args.lr, weight_decay=args.decay)
train_curve = []
valid_curve = []
test_curve = []
for epoch in tqdm(range(1, args.epochs + 1)):
loss = train(model, device, train_loader, optimizer, criterion)
logging.info("====epoch {} SupervisedLoss {}".format(epoch, loss))
logging.info("====Evaluation")
if args.eval_train:
train_acc = eval(model, device, train_loader)
else:
# print("omit the training accuracy computation")
train_acc = 0
val_acc = eval(model, device, val_loader)
test_acc = eval(model, device, test_loader)
logging.info("EvalTrain: {} EvalVal: {} EvalTestt: {}".format (train_acc, val_acc, test_acc))
train_curve.append(train_acc)
valid_curve.append(val_acc)
test_curve.append(test_acc)
logging.info(train_curve)
logging.info(valid_curve)
logging.info(test_curve)
best_val_epoch = np.argmax(np.array(valid_curve))
best_train = max(train_curve)
logging.info('FinishedTraining!')
logging.info('BestEpoch: {}'.format(best_val_epoch))
logging.info('BestTrainScore: {}'.format(best_train))
logging.info('BestValidationScore: {}'.format(valid_curve[best_val_epoch]))
logging.info('FinalTestScore: {}'.format(test_curve[best_val_epoch]))
return valid_curve[best_val_epoch], test_curve[best_val_epoch]
if __name__ == "__main__":
args = arg_parse()
run(args)