-
Notifications
You must be signed in to change notification settings - Fork 12
/
exec.py
217 lines (179 loc) · 7.3 KB
/
exec.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
import argparse
import csv
import json
import logging
import os
import time
from datetime import datetime
from typing import Optional
from autogluon.bench.datasets.constants import (
_IMAGE_SIMILARITY,
_IMAGE_TEXT_SIMILARITY,
_OBJECT_DETECTION,
_TEXT_SIMILARITY,
)
from autogluon.bench.datasets.dataset_registry import multimodal_dataset_registry
from autogluon.multimodal import MultiModalPredictor
from autogluon.multimodal import __version__ as ag_version
logger = logging.getLogger(__name__)
def _flatten_dict(data, separator="_", prefix=""):
flattened = {}
for key, value in data.items():
if isinstance(value, dict):
flattened.update(_flatten_dict(value, separator, prefix + key + separator))
else:
flattened[prefix + key] = value
return flattened
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--dataset_name",
type=str,
help="Dataset that has been registered with multimodal_dataset_registry.",
)
parser.add_argument("--framework", type=str, help="Framework (and) branch/version.")
parser.add_argument("--benchmark_dir", type=str, help="Directory to save benchmarking run.")
parser.add_argument("--metrics_dir", type=str, help="Directory to save benchmarking metrics.")
parser.add_argument(
"--time_limit", type=int, default=None, help="Time limit for the AutoGluon benchmark (in seconds)."
)
parser.add_argument("--presets", type=str, default=None, help="Preset configurations to use in the benchmark.")
parser.add_argument("--hyperparameters", type=str, default=None, help="Hyperparameters to use in the benchmark.")
args = parser.parse_args()
return args
def load_dataset(
dataset_name: str, # dataset name
):
"""Loads and preprocesses a dataset.
Args:
dataset_name (str): The name of the dataset to load.
Returns:
Tuple[pd.DataFrame, pd.DataFrame]: A tuple containing the training and test datasets.
"""
train_data = multimodal_dataset_registry.create(dataset_name, "train")
val_data = multimodal_dataset_registry.create(dataset_name, "val")
test_data = multimodal_dataset_registry.create(dataset_name, "test")
return train_data, val_data, test_data
def save_metrics(metrics_path: str, metrics: dict):
"""Saves evaluation metrics to a JSON file.
Args:
metrics_path (str): The path to the directory where the metrics should be saved.
metrics: The evaluation metrics to save.
Returns:
None
"""
if metrics is None:
logger.warning("No metrics were created.")
return
if not os.path.exists(metrics_path):
os.makedirs(metrics_path)
file = os.path.join(metrics_path, "results.csv")
flat_metrics = _flatten_dict(metrics)
field_names = flat_metrics.keys()
with open(file, "w", newline="") as f:
writer = csv.DictWriter(f, fieldnames=field_names)
writer.writeheader()
writer.writerow(flat_metrics)
logger.info("Metrics saved to %s.", file)
f.close()
def run(
dataset_name: str,
framework: str,
benchmark_dir: str,
metrics_dir: str,
time_limit: Optional[int] = None,
presets: Optional[str] = None,
hyperparameters: Optional[dict] = None,
):
"""Runs the AutoGluon multimodal benchmark on a given dataset.
Args:
dataset_name (str): Dataset that has been registered with multimodal_dataset_registry.
To get a list of datasets:
from autogluon.bench.datasets.dataset_registry import multimodal_dataset_registry
multimodal_dataset_registry.list_keys()
benchmark_dir (str): The path to the directory where benchmarking artifacts should be saved.
time_limit (int): The maximum amount of time (in seconds) to spend training the predictor (default: 10).
presets (str): The name of the AutoGluon preset to use (default: "None").
hyperparameters (str): A JSON of hyperparameters to use for training (default: None).
Returns:
None
"""
train_data, val_data, test_data = load_dataset(dataset_name=dataset_name)
try:
label_column = train_data.label_columns[0]
except (AttributeError, IndexError): # Object Detection does not have label columns
label_column = None
predictor_args = {
"label": label_column,
"problem_type": train_data.problem_type,
"presets": presets,
"path": os.path.join(benchmark_dir, "models"),
}
if train_data.problem_type == _IMAGE_SIMILARITY:
predictor_args["query"] = train_data.image_columns[0]
predictor_args["response"] = train_data.image_columns[1]
predictor_args["match_label"] = train_data.match_label
elif train_data.problem_type == _IMAGE_TEXT_SIMILARITY:
predictor_args["query"] = train_data.text_columns[0]
predictor_args["response"] = train_data.image_columns[0]
predictor_args["eval_metric"] = train_data.metric
del predictor_args["label"]
elif train_data.problem_type == _TEXT_SIMILARITY:
predictor_args["query"] = train_data.text_columns[0]
predictor_args["response"] = train_data.text_columns[1]
predictor_args["match_label"] = train_data.match_label
elif train_data.problem_type == _OBJECT_DETECTION:
predictor_args["sample_data_path"] = train_data.data
predictor = MultiModalPredictor(**predictor_args)
fit_args = {
"train_data": train_data.data,
"tuning_data": val_data.data,
"hyperparameters": hyperparameters,
"time_limit": time_limit,
}
utc_time = datetime.utcnow().strftime("%Y-%m-%dT%H:%M:%S")
start_time = time.time()
predictor.fit(**fit_args)
end_time = time.time()
training_duration = round(end_time - start_time, 1)
evaluate_args = {
"data": test_data.data,
"label": label_column,
"metrics": test_data.metric,
}
if test_data.problem_type == _IMAGE_TEXT_SIMILARITY:
evaluate_args["query_data"] = test_data.data[test_data.text_columns[0]].unique().tolist()
evaluate_args["response_data"] = test_data.data[test_data.image_columns[0]].unique().tolist()
evaluate_args["cutoffs"] = [1, 5, 10]
start_time = time.time()
scores = predictor.evaluate(**evaluate_args)
end_time = time.time()
predict_duration = round(end_time - start_time, 1)
if "#" in framework:
framework, version = framework.split("#")
else:
framework, version = framework, ag_version
metrics = {
"task": dataset_name,
"framework": framework,
"version": version,
"type": predictor.problem_type,
"utc_time": utc_time,
"training_duration": training_duration,
"predict_duration": predict_duration,
"scores": scores,
}
save_metrics(os.path.join(metrics_dir, "scores"), metrics)
if __name__ == "__main__":
args = get_args()
if args.hyperparameters is not None:
args.hyperparameters = json.loads(args.hyperparameters)
run(
dataset_name=args.dataset_name,
framework=args.framework,
benchmark_dir=args.benchmark_dir,
metrics_dir=args.metrics_dir,
time_limit=args.time_limit,
presets=args.presets,
hyperparameters=args.hyperparameters,
)