This repository has been archived by the owner on Aug 20, 2022. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmain.py
186 lines (158 loc) · 8.51 KB
/
main.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
from collections import defaultdict, Counter
import json
import logging
import os
from statistics import mean, stdev
import sys
import time
import numpy as np
from os.path import join
import pandas as pd
from pprint import pformat
from scipy.special import softmax
from sklearn.model_selection import KFold, train_test_split
import torch
import torch.nn.functional as F
from transformers import HfArgumentParser
from graph_dataset import DiffbotGraphDatasetNodeLabels
from util import EarlyStopping, create_dir_if_not_exists, get_args_info_as_str
from evaluation import calc_classification_metrics
from experiment_args import ExperimentArguments, TrainingArguments
from trainer import Trainer
logger = logging.getLogger(__name__)
def main():
parser = HfArgumentParser((ExperimentArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
exp_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
exp_args, training_args = parser.parse_args_into_dataclasses()
create_dir_if_not_exists(exp_args.log_dir)
stream_handler = logging.StreamHandler(sys.stderr)
file_handler = logging.FileHandler(filename=os.path.join(exp_args.log_dir, 'train_log.txt'),
mode='w+')
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.DEBUG if exp_args.debug else logging.INFO,
handlers=[stream_handler, file_handler]
)
logger.info(f"======== Experiment Args ========\n{get_args_info_as_str(exp_args)}\n")
logger.info(f"======== Training Args ========\n{get_args_info_as_str(training_args)}\n")
dataset = DiffbotGraphDatasetNodeLabels(exp_args.dataset_name,
exp_args.market,
exp_args.market_source,
exp_args.natts_path,
metapaths=exp_args.metapaths)
dataset._process()
natts_path = exp_args.natts_path if exp_args.natts_path else dataset.natts_path
with open(natts_path, 'rb') as f:
natts_dict = json.load(f)
device = torch.device(f"cuda:{training_args.device}" if torch.cuda.is_available() and training_args.device >= 0 else "cpu")
logger.info(f'======== Dataset Feature Info ========\n{pformat(natts_dict)}\n')
logger.info(f'\n Node Feat Dimensions: {pformat({name: feat.shape for name, feat in dataset.graph.ndata["feats"].items()})}')
logger.info(f'Number of nodes {dataset.homo_graph.number_of_nodes()}\n')
logger.info(f'Number of edges {dataset.homo_graph.number_of_edges()}\n')
logger.info(f'Canonical Edge Types: {pformat(dataset.graph.canonical_etypes)}\n')
logger.info(f'Node Types: {pformat(dataset.graph.ntypes)}\n')
if dataset.metapaths:
logger.info(f'Metapaths: {pformat(dataset.metapaths)}\n')
labels = torch.LongTensor(dataset.labels).to(device)
if training_args.repeat == 1:
train_inds, val_inds, test_inds = [dataset.train_idx], [dataset.val_idx], [dataset.test_idx]
else:
kf = KFold(n_splits=training_args.repeat, shuffle=True)
splits = list(kf.split(dataset.labels))
train_inds, val_inds, test_inds = [], [], []
for train_idxs, test_idxs in splits:
test_inds.append(test_idxs)
train_idxs, val_idxs = train_test_split(train_idxs, test_size=0.2)
train_inds.append(train_idxs), val_inds.append(val_idxs)
logger.info(f"Labels Count {Counter(dataset.labels)}")
logger.info(f"Num train {len(train_inds[0])}")
logger.info(f"Num val {len(val_inds[0])}")
logger.info(f"Num test {len(test_inds[0])}")
test_metrics_all = defaultdict(list)
for run_num in range(training_args.repeat):
train_idx, val_idx, test_idx = train_inds[run_num], val_inds[run_num], test_inds[run_num]
logger.info(f"======== Run {run_num} ========")
trainer = Trainer(dataset, exp_args.model, training_args, device)
if run_num == 0:
logger.info(f"======== MODEL ========\n{trainer.model}")
optimizer = torch.optim.Adam(trainer.model.parameters(), lr=training_args.lr,
weight_decay=training_args.weight_decay)
trainer.train()
early_stopping = EarlyStopping(patience=training_args.patience,
want_increase=training_args.val_metric != 'loss',
metric_name=training_args.val_metric,
verbose=True,
save_path=join(exp_args.log_dir, 'checkpoint.pt'))
dur1 = []
dur2 = []
dur3 = []
for epoch in range(training_args.num_epochs):
t0 = time.time()
# training forward
trainer.train()
logits, embeddings = trainer.forward()
train_loss = F.cross_entropy(logits[train_idx], labels[train_idx])
t1 = time.time()
dur1.append(t1 - t0)
optimizer.zero_grad()
train_loss.backward()
optimizer.step()
t2 = time.time()
dur2.append(t2 - t1)
trainer.eval()
with torch.no_grad():
logits, embeddings = trainer.forward()
val_loss = F.cross_entropy(logits[val_idx], labels[val_idx])
logits = logits.cpu().detach().numpy()
embeddings = embeddings.cpu().detach().numpy()
pred_scores = softmax(logits[val_idx], axis=1)[:, 1]
preds_labels = np.argmax(logits[val_idx], axis=1)
val_metrics = calc_classification_metrics(pred_scores, preds_labels, labels[val_idx].cpu().detach().numpy())
val_metrics['loss'] = val_loss.item()
t3 = time.time()
dur3.append(t3 - t2)
if epoch % training_args.print_val_epochs == 0:
logging.info('\nValidation metrics\n' + pformat(val_metrics, indent=3))
logging.info(
"Epoch {:05d} | Train_Loss {:.4f} | Val_Loss {:.4f} | TimeFoward(s) {:.4f} | TimeBackward(s) {:.4f} | TimeValidation(s) {:.4f}".format(
epoch, train_loss.item(), val_loss.item(), np.mean(dur1), np.mean(dur2), np.mean(dur3)))
else:
logging.debug('\nValidation metrics\n' + pformat(val_metrics, indent=3))
logging.debug(
"Epoch {:05d} | Train_Loss {:.4f} | Val_Loss {:.4f} | TimeFoward(s) {:.4f} | TimeBackward(s) {:.4f} | TimeValidation(s) {:.4f}".format(
epoch, train_loss.item(), val_loss.item(), np.mean(dur1), np.mean(dur2), np.mean(dur3)))
early_stopping(val_metrics[training_args.val_metric], trainer.model)
if early_stopping.early_stop:
logging.info('Early stopping!')
break
trainer.model.load_state_dict(torch.load(join(exp_args.log_dir, 'checkpoint.pt')))
trainer.eval()
with torch.no_grad():
logits, _ = trainer.forward()
logits = logits.cpu().detach().numpy()
pred_scores = softmax(logits[test_idx], axis=1)[:, 1]
preds_labels = np.argmax(logits[test_idx], axis=1)
test_metrics = calc_classification_metrics(pred_scores, preds_labels, labels[test_idx].cpu().detach().numpy())
for k, metric in test_metrics.items():
test_metrics_all[k].append(metric)
for ind, pred_score, pred_label in zip(test_idx, pred_scores, preds_labels):
dataset.labels_info[ind]['pred_score'] = pred_score
dataset.labels_info[ind]['pred_label'] = pred_label
test_res_df = pd.DataFrame([label_info for i, label_info in enumerate(dataset.labels_info) if i in test_idx])
test_res_df.to_csv(join(exp_args.log_dir, f"test_res_run{run_num}.csv"))
logging.info('\nTesting\n' + pformat(dict(test_metrics_all), indent=3))
avg_results = {}
for k, v in dict(test_metrics_all).items():
avg_results[k + '_mean'] = mean(v)
avg_results[k + '_std'] = stdev(v)
logger.info(f'{pformat(avg_results, indent=3)}')
with open(join(exp_args.log_dir, 'test_results.json'), 'w') as f:
json.dump(dict(test_metrics_all), f, indent=4)
with open(join(exp_args.log_dir, 'test_results_avg.json'), 'w') as f:
json.dump(avg_results, f, indent=4)
logger.info(f'\nResults saved to {exp_args.log_dir}')
if __name__ == '__main__':
main()