Upload run_experiment.py
Browse files- run_experiment.py +154 -0
run_experiment.py
ADDED
@@ -0,0 +1,154 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import click
|
2 |
+
import datetime
|
3 |
+
import pprint
|
4 |
+
from typing import Optional
|
5 |
+
|
6 |
+
|
7 |
+
from src import (
|
8 |
+
load_dataset,
|
9 |
+
fit_predict_with_model,
|
10 |
+
score_predictions,
|
11 |
+
AVAILABLE_DATASETS,
|
12 |
+
AVAILABLE_MODELS,
|
13 |
+
SEASONALITY_MAP,
|
14 |
+
)
|
15 |
+
|
16 |
+
|
17 |
+
def apply_ablation(ablation: str, model_kwargs: dict) -> dict:
|
18 |
+
if ablation == "NoEnsemble":
|
19 |
+
model_kwargs["enable_ensemble"] = False
|
20 |
+
elif ablation == "NoDeepModels":
|
21 |
+
model_kwargs["hyperparameters"] = {
|
22 |
+
"Naive": {},
|
23 |
+
"SeasonalNaive": {},
|
24 |
+
"ARIMA": {},
|
25 |
+
"ETS": {},
|
26 |
+
"AutoETS": {},
|
27 |
+
"AutoARIMA": {},
|
28 |
+
"Theta": {},
|
29 |
+
"AutoGluonTabular": {},
|
30 |
+
}
|
31 |
+
elif ablation == "NoStatModels":
|
32 |
+
model_kwargs["hyperparameters"] = {
|
33 |
+
"AutoGluonTabular": {},
|
34 |
+
"DeepAR": {},
|
35 |
+
"SimpleFeedForward": {},
|
36 |
+
"TemporalFusionTransformer": {},
|
37 |
+
}
|
38 |
+
elif ablation == "NoTreeModels":
|
39 |
+
model_kwargs["hyperparameters"] = {
|
40 |
+
"Naive": {},
|
41 |
+
"SeasonalNaive": {},
|
42 |
+
"ARIMA": {},
|
43 |
+
"ETS": {},
|
44 |
+
"AutoETS": {},
|
45 |
+
"AutoARIMA": {},
|
46 |
+
"Theta": {},
|
47 |
+
"DeepAR": {},
|
48 |
+
"SimpleFeedForward": {},
|
49 |
+
"TemporalFusionTransformer": {},
|
50 |
+
}
|
51 |
+
return model_kwargs
|
52 |
+
|
53 |
+
|
54 |
+
@click.command(
|
55 |
+
context_settings=dict(
|
56 |
+
ignore_unknown_options=True,
|
57 |
+
allow_extra_args=True,
|
58 |
+
)
|
59 |
+
)
|
60 |
+
@click.option(
|
61 |
+
"--dataset_name",
|
62 |
+
"-d",
|
63 |
+
required=True,
|
64 |
+
default="m3_other",
|
65 |
+
help="The dataset to train the model on",
|
66 |
+
type=click.Choice(AVAILABLE_DATASETS),
|
67 |
+
)
|
68 |
+
@click.option(
|
69 |
+
"--model_name",
|
70 |
+
"-m",
|
71 |
+
default="autogluon",
|
72 |
+
help="Model to train",
|
73 |
+
type=click.Choice(AVAILABLE_MODELS),
|
74 |
+
)
|
75 |
+
@click.option(
|
76 |
+
"--eval_metric",
|
77 |
+
"-e",
|
78 |
+
default="MASE",
|
79 |
+
type=click.Choice(["MASE", "mean_wQuantileLoss"]),
|
80 |
+
)
|
81 |
+
@click.option(
|
82 |
+
"--seed",
|
83 |
+
"-s",
|
84 |
+
default=1,
|
85 |
+
type=int,
|
86 |
+
)
|
87 |
+
@click.option(
|
88 |
+
"--time_limit",
|
89 |
+
"-t",
|
90 |
+
default=4 * 3600,
|
91 |
+
type=int,
|
92 |
+
)
|
93 |
+
@click.option(
|
94 |
+
"--ablation",
|
95 |
+
"-a",
|
96 |
+
default=None,
|
97 |
+
type=click.Choice(["NoEnsemble", "NoDeepModels", "NoStatModels", "NoTreeModels"]),
|
98 |
+
)
|
99 |
+
@click.pass_context
|
100 |
+
def main(
|
101 |
+
ctx,
|
102 |
+
dataset_name: str,
|
103 |
+
model_name: str,
|
104 |
+
eval_metric: str,
|
105 |
+
seed: int,
|
106 |
+
time_limit: int,
|
107 |
+
ablation: Optional[str],
|
108 |
+
):
|
109 |
+
print(f"Evaluating {model_name} on {dataset_name}")
|
110 |
+
dataset = load_dataset(dataset_name)
|
111 |
+
task_kwargs = {
|
112 |
+
"prediction_length": dataset.metadata.prediction_length,
|
113 |
+
"freq": dataset.metadata.freq,
|
114 |
+
"eval_metric": eval_metric,
|
115 |
+
"seasonality": SEASONALITY_MAP[dataset.metadata.freq],
|
116 |
+
}
|
117 |
+
print("Task definition:")
|
118 |
+
pprint.pprint(task_kwargs)
|
119 |
+
|
120 |
+
# Additional command line arguments like `--name value` are parsed as {"name": "value"}
|
121 |
+
model_kwargs = {ctx.args[i][2:]: ctx.args[i + 1] for i in range(0, len(ctx.args), 2)}
|
122 |
+
model_kwargs["seed"] = seed
|
123 |
+
model_kwargs["time_limit"] = time_limit
|
124 |
+
|
125 |
+
if ablation is not None:
|
126 |
+
assert model_name == "autogluon", f"{model_name} does not support ablations"
|
127 |
+
model_kwargs = apply_ablation(ablation, model_kwargs)
|
128 |
+
|
129 |
+
if len(model_kwargs) > 0:
|
130 |
+
print("Model kwargs:")
|
131 |
+
pprint.pprint(model_kwargs)
|
132 |
+
|
133 |
+
print(f"Starting training {datetime.datetime.now()}")
|
134 |
+
|
135 |
+
predictions, info = fit_predict_with_model(
|
136 |
+
model_name, dataset.train, **task_kwargs, **model_kwargs
|
137 |
+
)
|
138 |
+
|
139 |
+
metrics = score_predictions(
|
140 |
+
dataset=dataset.test,
|
141 |
+
predictions=predictions,
|
142 |
+
prediction_length=task_kwargs["prediction_length"],
|
143 |
+
seasonality=task_kwargs["seasonality"],
|
144 |
+
)
|
145 |
+
print("================================================")
|
146 |
+
print(f"model: {model_name}")
|
147 |
+
print(f"dataset: {dataset_name}")
|
148 |
+
print(f"total_run_time: {info['run_time']:.2f}")
|
149 |
+
print(f"mase: {metrics['MASE']:.4f}")
|
150 |
+
print(f"mean_wQuantileLoss: {metrics['mean_wQuantileLoss']:.4f}")
|
151 |
+
|
152 |
+
|
153 |
+
if __name__ == "__main__":
|
154 |
+
main()
|