-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtuning.py
141 lines (118 loc) · 6.8 KB
/
tuning.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
import os
import torch
import tempfile
from ray import train, tune
from ray.tune.search.basic_variant import BasicVariantGenerator
# from ray.tune.search.optuna import OptunaSearch
# from ray.tune.search import ConcurrencyLimiter
# from ray.tune.schedulers import ASHAScheduler
import train as tr
from model import FinSIRModel as Model
class Arguments:
def __init__(self, **kwargs):
self.__dict__.update(kwargs)
def run(config):
args = Arguments(**config, nworkers=2, use_amp=True, return_period=1, add_self_loop=True,
recurrent_layers=1, recurrent_dropout=0, relational_agg='sym', relational_dropout=0,
readout_layers=1, readout_dropout=0, lr=1e-3, wd=1e-5, l1=0, l2=0, factor=0.5, patience=50, k_list=[1, 5])
# args.corr_graph_periods = [int(p) for p in args.corr_graph_periods.split(' ')]
# Change directory
os.chdir(base_path)
# Set seed
tr.set_seed(0)
# Load dataset
device = torch.device('cuda:0') if torch.cuda.is_available() else torch.device('cpu')
dataset, train_loader, val_loader, test_loader = tr.load_dataset(args.market, args)
# Extract input shapes
input_dim = dataset[0][1].shape[-1]
edge_dim = {'wiki': dataset.wiki_graph.edata['feat'].shape[-1],
'industry': dataset.industry_graph.edata['feat'].shape[-1]}[args.relational_graph]
# 'correlation': len(args.corr_graph_periods) + 1
# Load model
model = Model(input_dim, edge_dim, args.nhidden, 1, args.recurrent_layers, args.recurrent_dropout,
args.relational_agg, args.relational_dropout, args.readout_layers, args.readout_dropout).to(device)
# model = torch.compile(model, mode='default')
# Scaler + optimizer + scheduler
scaler = torch.amp.GradScaler(device=device.type, enabled=args.use_amp)
optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=args.wd)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=args.factor, patience=args.patience)
best_val_loss = 1e10
# Restore checkpoint
checkpoint = train.get_checkpoint()
if checkpoint:
with checkpoint.as_directory() as checkpoint_dir:
model_state, scaler_state, optimizer_state, scheduler_state = torch.load(os.path.join(checkpoint_dir, 'checkpoint.pt'))
model.load_state_dict(model_state)
scaler.load_state_dict(scaler_state)
optimizer.load_state_dict(optimizer_state)
scheduler.load_state_dict(scheduler_state)
for epoch in range(50):
tr.warmup_lr(optimizer, args.lr, epoch + 1, 10)
loss = tr.train(model, dataset, train_loader, device, optimizer, scaler, args)
mse, loss, mrr, irr = tr.evaluate(model, dataset, train_loader, device, args)
val_mse, val_loss, val_mrr, val_irr = tr.evaluate(model, dataset, val_loader, device, args)
test_mse, test_loss, test_mrr, test_irr = tr.evaluate(model, dataset, test_loader, device, args)
scheduler.step(loss)
# Process irr
irr = {k: sum(irr_k) for k, irr_k in irr.items()}
val_irr = {k: sum(irr_k) for k, irr_k in val_irr.items()}
test_irr = {k: sum(irr_k) for k, irr_k in test_irr.items()}
if val_loss < best_val_loss:
best_val_loss = val_loss
result = {
'val_mse': val_mse,
'val_mrr': val_mrr,
'val_irr': val_irr,
'val_loss': val_loss,
'test_mse': test_mse,
'test_mrr': test_mrr,
'test_irr': test_irr,
'test_loss': test_loss,
}
with tempfile.TemporaryDirectory() as checkpoint_dir:
torch.save((model.state_dict(), scaler.state_dict(), optimizer.state_dict(), scheduler.state_dict()), os.path.join(checkpoint_dir, 'checkpoint.pt'))
checkpoint = train.Checkpoint.from_directory(checkpoint_dir)
train.report({'mse': mse, 'loss': loss, 'mrr_1': mrr[1], 'mrr_5': mrr[5], 'irr_1': irr[1], 'irr_5': irr[5],
'val_mse': val_mse, 'val_loss': val_loss, 'val_mrr_1': val_mrr[1], 'val_mrr_5': val_mrr[5], 'val_irr_1': val_irr[1], 'val_irr_5': val_irr[5],
'test_mse': test_mse, 'test_loss': test_loss, 'test_mrr_1': test_mrr[1], 'test_mrr_5': test_mrr[5], 'test_irr_1': test_irr[1], 'test_irr_5': test_irr[5],
'best_val_mse': result['val_mse'], 'best_val_loss': result['val_loss'], 'best_val_irr_1': result['val_irr'][1], 'best_val_irr_5': result['val_irr'][5],
'best_test_mse': result['test_mse'], 'best_test_loss': result['test_loss'], 'best_test_irr_1': result['test_irr'][1], 'best_test_irr_5': result['test_irr'][5]}, checkpoint=checkpoint)
if __name__ == '__main__':
market = 'NYSE'
relational_graph = 'wiki'
param_space = {
'market': tune.grid_search([market]), # FIXED
'relational_graph': tune.grid_search([relational_graph]), # FIXED
'sequence_length': tune.grid_search([2, 4, 8, 16]),
'nhidden': tune.grid_search([8, 16, 32, 64]),
'alpha': tune.grid_search([0.1, 1.0, 10.0]),
}
hyperparams = [
]
base_path = '/home/FinSIR/'
directory = os.path.join(base_path, 'logs/tune/')
exp_name = f'{market}_{Model.__name__}_{relational_graph}'
search_alg = BasicVariantGenerator(points_to_evaluate=hyperparams, max_concurrent=4)
scheduler = None
# search_alg = OptunaSearch(points_to_evaluate=hyperparams)
# search_alg.restore_from_dir(os.path.join(directory, exp_name))
# search_alg = ConcurrencyLimiter(search_alg, max_concurrent=6)
# scheduler = ASHAScheduler(max_t=50, grace_period=50)
if tune.Tuner.can_restore(os.path.join(directory, exp_name)):
tuner = tune.Tuner.restore(
path=os.path.join(directory, exp_name),
trainable=tune.with_resources(tune.with_parameters(run), resources={'cpu': 6, 'gpu': 1/4, 'accelerator_type:RTX': 1/4}),
resume_unfinished=True,
)
else:
tuner = tune.Tuner(
trainable=tune.with_resources(tune.with_parameters(run), resources={'cpu': 6, 'gpu': 1/4, 'accelerator_type:RTX': 1/4}),
tune_config=tune.TuneConfig(mode='max', metric='best_val_irr_1', search_alg=search_alg, scheduler=scheduler, num_samples=1),
run_config=train.RunConfig(name=exp_name, storage_path=directory, failure_config=train.FailureConfig(max_failures=2),
checkpoint_config=train.CheckpointConfig(num_to_keep=1)),
param_space=param_space,
)
results = tuner.fit()
best_result = results.get_best_result()
print(f'Best trial config: {best_result.config}')
print(f'Best trial test irr_1: {best_result.metrics["best_test_irr_1"]}')