-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathSTMeta_demo.py
218 lines (189 loc) · 10.4 KB
/
STMeta_demo.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
import os
import nni
import yaml
import argparse
import GPUtil
from UCTB.dataset import NodeTrafficLoader, ContextLoader
from UCTB.model import STMeta
from UCTB.evaluation import metric
from UCTB.preprocess.time_utils import is_work_day_china, is_work_day_america
from UCTB.preprocess.GraphGenerator import GraphGenerator
from UCTB.preprocess import SplitData
#####################################################################
# argument parser
parser = argparse.ArgumentParser(description="Argument Parser")
parser.add_argument('-m', '--model', default='STMeta_v0.model.yml')
parser.add_argument('-d', '--data', default='didi_chengdu.data.yml')
parser.add_argument('-p', '--update_params', default='')
# Parse params
terminal_vars = vars(parser.parse_args())
yml_files = [terminal_vars['model'], terminal_vars['data']]
args = {}
for yml_file in yml_files:
with open(yml_file, 'r') as f:
args.update(yaml.load(f))
if len(terminal_vars['update_params']) > 0:
args.update({e.split(':')[0]: e.split(':')[1] for e in terminal_vars['update_params'].split(',')})
print({e.split(':')[0]: e.split(':')[1] for e in terminal_vars['update_params'].split(',')})
nni_params = nni.get_next_parameter()
nni_sid = nni.get_sequence_id()
if nni_params:
args.update(nni_params)
args['mark'] += str(nni_sid)
#####################################################################
# Generate code_version
code_version = '{}_C{}P{}T{}_G{}_K{}L{}_F{}_{}'.format(args['model_version'],
args['closeness_len'], args['period_len'],args['trend_len'],
''.join([e[0] for e in args['graph'].split('-')]),
args['gcn_k'], args['gcn_layers'], int(args["MergeIndex"])*5, args['mark'])
model_dir_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'model_dir')
model_dir_path = os.path.join(model_dir_path, args['group'])
#####################################################################
data_loader = NodeTrafficLoader(dataset=args['dataset'], city=args['city'],
data_range=args['data_range'], train_data_length=args['train_data_length'],
test_ratio=float(args['test_ratio']),
closeness_len=args['closeness_len'],
period_len=args['period_len'],
trend_len=args['trend_len'],
normalize=args['normalize'],
with_tpe=args['with_tpe'],
workday_parser=is_work_day_america if args['dataset'] == 'Bike' else is_work_day_china,
MergeIndex=args['MergeIndex'],
MergeWay=args["MergeWay"])
context_loader = ContextLoader(traffic_dataloader = data_loader,
spatial_context_file_list = ["poi.csv", "road.csv"],
temporal_context_file_list = ["weather.csv", "holiday.csv"],
spatiotemporal_context_file_list = ["weather.csv", "AQI.csv"],
future_ST_context_length = 0,
past_ST_context_length = 6)
# context_loader.spatial_context with shape (S_feature_categories, num_node, num_feature)
# context_loader.temporal_context with shape (T_feature_categories, num_slots, num_feature)
# context_loader.future_ST_context with shape (ST_feature_categories, future_time_step, num_context_node, num_feature)
# context_loader.past_ST_context with shape (ST_feature_categories, past_time_step, num_context_node, num_feature)
# split data
train_closeness, val_closeness = SplitData.split_data(data_loader.train_closeness, [0.9, 0.1])
train_period, val_period = SplitData.split_data(data_loader.train_period, [0.9, 0.1])
train_trend, val_trend = SplitData.split_data(data_loader.train_trend, [0.9, 0.1])
train_y, val_y = SplitData.split_data(data_loader.train_y, [0.9, 0.1])
train_ef, val_ef = SplitData.split_data(data_loader.train_ef, [0.9, 0.1])
# build graphs
graph_obj = GraphGenerator(graph=args['graph'],
data_loader=data_loader,
threshold_distance=args['threshold_distance'],
threshold_correlation=args['threshold_correlation'],
threshold_interaction=args['threshold_interaction'])
print("TimeFitness", data_loader.dataset.time_fitness)
print("TimeRange", data_loader.dataset.time_range)
de_normalizer = None if args['normalize'] is False else data_loader.normalizer.inverse_transform
deviceIDs = GPUtil.getAvailable(order='last', limit=8, maxLoad=1, maxMemory=0.7,
includeNan=False, excludeID=[], excludeUUID=[])
if len(deviceIDs) == 0:
current_device = '-1'
else:
if nni_params:
current_device = str(deviceIDs[int(nni_sid) % len(deviceIDs)])
else:
current_device = str(deviceIDs[0])
STMeta_obj = STMeta(num_node=data_loader.station_number,
num_graph=graph_obj.LM.shape[0],
external_dim=data_loader.external_dim,
closeness_len=args['closeness_len'],
period_len=args['period_len'],
trend_len=args['trend_len'],
input_dim=2 if args['with_tpe'] else 1,
gcn_k=int(args.get('gcn_k', 0)),
gcn_layers=int(args.get('gcn_layers', 0)),
gclstm_layers=int(args['gclstm_layers']),
num_hidden_units=args['num_hidden_units'],
num_dense_units=args['num_filter_conv1x1'],
# temporal attention parameters
tpe_dim=data_loader.tpe_dim,
temporal_gal_units=args.get('temporal_gal_units'),
temporal_gal_num_heads=args.get('temporal_gal_num_heads'),
temporal_gal_layers=args.get('temporal_gal_layers'),
# merge parameters
graph_merge_gal_units=args['graph_merge_gal_units'],
graph_merge_gal_num_heads=args['graph_merge_gal_num_heads'],
temporal_merge_gal_units=args['temporal_merge_gal_units'],
temporal_merge_gal_num_heads=args['temporal_merge_gal_num_heads'],
# network structure parameters
st_method=args['st_method'], # gclstm
temporal_merge=args['temporal_merge'], # gal
graph_merge=args['graph_merge'], # concat
build_transfer=args['build_transfer'],
lr=float(args['lr']),
code_version=code_version,
model_dir=model_dir_path,
gpu_device=current_device)
STMeta_obj.build()
print(args['dataset'], args['city'], code_version)
print('Number of trainable variables', STMeta_obj.trainable_vars)
print('Number of training samples', data_loader.train_sequence_len)
# # Training
if args['train']:
STMeta_obj.fit(closeness_feature=data_loader.train_closeness,
period_feature=data_loader.train_period,
trend_feature=data_loader.train_trend,
laplace_matrix=graph_obj.LM,
target=data_loader.train_y,
external_feature=data_loader.train_ef,
sequence_length=data_loader.train_sequence_len,
output_names=('loss', ),
evaluate_loss_name='loss',
op_names=('train_op', ),
batch_size=int(args['batch_size']),
max_epoch=int(args['max_epoch']),
validate_ratio=0.1,
early_stop_method='t-test',
early_stop_length=args['early_stop_length'],
early_stop_patience=args['early_stop_patience'],
verbose=True,
save_model=True)
STMeta_obj.load(code_version)
# val prediction
prediction = STMeta_obj.predict(closeness_feature=val_closeness,
period_feature=val_period,
trend_feature=val_trend,
laplace_matrix=graph_obj.LM,
target=val_y,
external_feature=val_ef,
output_names=('prediction', ),
sequence_length=max(
(len(val_closeness), len(val_period), len(val_trend))),
cache_volume=int(args['batch_size']), )
val_prediction = prediction['prediction']
# test prediction
prediction = STMeta_obj.predict(closeness_feature=data_loader.test_closeness,
period_feature=data_loader.test_period,
trend_feature=data_loader.test_trend,
laplace_matrix=graph_obj.LM,
target=data_loader.test_y,
external_feature=data_loader.test_ef,
output_names=('prediction', ),
sequence_length=data_loader.test_sequence_len,
cache_volume=int(args['batch_size']), )
test_prediction = prediction['prediction']
if de_normalizer:
test_prediction = de_normalizer(test_prediction)
data_loader.test_y = de_normalizer(data_loader.test_y)
val_prediction = de_normalizer(val_prediction)
val_y = de_normalizer(val_y)
test_rmse = metric.rmse(prediction=test_prediction, target=data_loader.test_y)
val_rmse = metric.rmse(prediction=val_prediction, target=val_y)
# Evaluate loss during training
val_loss = STMeta_obj.load_event_scalar('val_loss')
# best_val_loss = min([e[-1] for e in val_loss])
# if de_normalizer:
# best_val_loss = de_normalizer(best_val_loss)
# print('Best val result', best_val_loss)
print('Val result', val_rmse)
print('Test result', test_rmse)
time_consumption = [val_loss[e][0] - val_loss[e-1][0] for e in range(1, len(val_loss))]
time_consumption = sum([e for e in time_consumption if e < (min(time_consumption) * 10)]) / 3600
print('Converged using %.2f hour / %s epochs' % (time_consumption, STMeta_obj._global_step))
# if nni_params:
# nni.report_final_result({
# 'default': best_val_loss,
# 'test-rmse': test_rmse,
# 'test-mape': test_mape
# })