Repository navigation
Expand file tree
/
Copy pathmain.py
More file actions
183 lines (162 loc) · 5.99 KB
/
Copy pathmain.py
File metadata and controls
183 lines (162 loc) · 5.99 KB
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
"""Command-line entry point for download, training, and evaluation."""
from __future__ import annotations
import argparse
import logging
import platform
import time
from pathlib import Path
from typing import Any
import torch
from data import get_benchmark_spec, load_dataset
from data.data_loader import fourier_time_features
from scripts.evaluate import evaluate_model
from scripts.train import load_trained_model, train_experiment
from utils import (
apply_overrides,
choose_device,
configure_logging,
load_config,
resolve_project_path,
seed_everything,
write_json,
)
from utils.visualization import plot_training_history
LOGGER = logging.getLogger(__name__)
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Reproduce the TimeGrad paper")
parser.add_argument(
"--config", default="config/exchange.yaml", help="YAML experiment config"
)
parser.add_argument(
"--mode",
choices=("download", "train", "evaluate", "all"),
default="all",
)
parser.add_argument(
"--set",
dest="overrides",
action="append",
default=[],
help="override a value, e.g. --set training.epochs=2",
)
return parser.parse_args()
def _absolute_paths(config: dict[str, Any]) -> dict[str, Any]:
config["dataset"]["root"] = str(
resolve_project_path(config, config["dataset"].get("root", "data/dataset"))
)
return config
def run(config: dict[str, Any], mode: str) -> dict[str, Any]:
configure_logging(config)
seed = int(config["experiment"]["seed"])
seed_everything(
seed,
deterministic=bool(config["experiment"].get("deterministic", True)),
)
device = choose_device(str(config["experiment"].get("device", "auto")))
context_length = int(config["model"]["context_length"])
lags = [int(value) for value in config["model"]["lags"]]
history_length = context_length + max(lags)
LOGGER.info("Loading dataset on device=%s", device)
dataset = load_dataset(config["dataset"], history_length=history_length, seed=seed)
spec = get_benchmark_spec(dataset.benchmark or dataset.name)
time_feature_dim = fourier_time_features(
dataset.train_start,
history_length + dataset.prediction_length,
dataset.freq,
).shape[1]
if mode == "download":
return {
"status": "downloaded",
"dataset": dataset.name,
"benchmark": dataset.benchmark,
"train_shape": list(dataset.train_target.shape),
"full_target_dim": dataset.full_target_dim,
"frequency": dataset.freq,
"prediction_length": dataset.prediction_length,
"rolling_windows": len(dataset.test_windows),
"paper_reference": spec.paper_reference() if spec else None,
}
checkpoint_path = resolve_project_path(
config, config["outputs"]["checkpoint"]
)
history_path = resolve_project_path(config, config["outputs"]["history_csv"])
loss_plot_path = resolve_project_path(config, config["outputs"]["loss_plot"])
evaluation_path = resolve_project_path(
config, config["outputs"]["evaluation_report"]
)
forecast_plot_path = resolve_project_path(
config, config["outputs"]["forecast_plot"]
)
start_time = time.perf_counter()
training_summary: dict[str, Any] | None = None
if mode in {"train", "all"}:
training = train_experiment(
dataset=dataset,
config=config,
time_feature_dim=int(time_feature_dim),
device=device,
checkpoint_path=checkpoint_path,
history_path=history_path,
)
model = training.model
plot_training_history(training.history, loss_plot_path)
training_summary = {
"best_epoch": training.best_epoch,
"loss_improved_over_run": training.improved,
"epochs_executed": len(training.history),
"first_train_loss": float(training.history[0]["train_loss"]),
"last_train_loss": float(training.history[-1]["train_loss"]),
}
else:
model, checkpoint = load_trained_model(checkpoint_path, device)
training_summary = {
"loaded_checkpoint_epoch": int(checkpoint["epoch"]),
"loaded_checkpoint_phase": checkpoint["phase"],
}
evaluation_summary = None
if mode in {"evaluate", "all"}:
evaluation_summary = evaluate_model(
model=model,
dataset=dataset,
config=config,
device=device,
report_path=evaluation_path,
plot_path=forecast_plot_path,
)
summary = {
"status": "completed",
"mode": mode,
"experiment": config["experiment"]["name"],
"seed": seed,
"device": str(device),
"python": platform.python_version(),
"torch": torch.__version__,
"dataset": {
"name": dataset.name,
"benchmark": dataset.benchmark,
"train_shape": list(dataset.train_target.shape),
"full_target_dim": dataset.full_target_dim,
"frequency": dataset.freq,
"prediction_length": dataset.prediction_length,
"rolling_windows_available": len(dataset.test_windows),
},
"model_parameters": model.parameter_count(),
"training": training_summary,
"evaluation": evaluation_summary,
"elapsed_seconds": time.perf_counter() - start_time,
}
summary_path = resolve_project_path(config, config["outputs"]["run_summary"])
write_json(summary_path, summary)
return summary
def main() -> None:
args = parse_args()
config_path = Path(args.config)
if not config_path.is_absolute():
config_path = Path.cwd() / config_path
config = load_config(config_path)
config = apply_overrides(config, args.overrides)
config = _absolute_paths(config)
summary = run(config, args.mode)
print(f"Completed: {summary}")
if __name__ == "__main__":
main()