/
tune.py
196 lines (152 loc) · 7.27 KB
/
tune.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
import os.path
from locationencoder import LocationEncoder
from data import LandOceanDataModule, Inat2018DataModule, CheckerboardDataModule
import lightning as pl
import optuna
from lightning.pytorch.callbacks.early_stopping import EarlyStopping
import pandas as pd
import yaml
TUNE_RESULTS_DIR = "results/tune"
import logging
logging.getLogger("lightning").setLevel(logging.ERROR)
def get_hyperparameter(trial: optuna.trial.Trial, positional_encoding_name, neural_network_name):
hparams_pe = {}
if positional_encoding_name in ["theory", "grid", "spherec", "spherecplus", "spherem", "spheremplus"]:
hparams_pe["min_radius"] = trial.suggest_int("min_radius", 1, 90, step=9)
hparams_pe["max_radius"] = 360
hparams_pe["frequency_num"] = trial.suggest_int("frequency_num", 16, 64, step=16)
elif positional_encoding_name == "sphericalharmonics":
hparams_pe["legendre_polys"] = trial.suggest_int("legendre_polys", 10, 30, step=5)
hparams_pe["embedding_dim"] = trial.suggest_int("embedding_dim", 16, 128, step=16)
hparams_nn = {}
if neural_network_name == "mlp":
hparams_nn["dim_hidden"] = trial.suggest_int("dim_hidden", 32, 128, step=32)
hparams_nn["num_layers"] = trial.suggest_int("num_layers", 1, 3)
elif neural_network_name == "fcnet":
hparams_nn["dim_hidden"] = trial.suggest_int("dim_hidden", 32, 128, step=32)
elif neural_network_name == "siren":
hparams_nn["dim_hidden"] = trial.suggest_int("dim_hidden", 32, 128, step=32)
hparams_nn["num_layers"] = trial.suggest_int("num_layers", 1, 3)
hparams_opt = {}
hparams_opt["lr"] = trial.suggest_float("lr", 1e-4, 1e-1, log=True)
hparams_opt["wd"] = trial.suggest_float("wd", 1e-8, 1e-1, log=True)
hparams = {}
hparams.update(hparams_pe)
hparams.update(hparams_nn)
hparams["optimizer"] = hparams_opt
hparams['harmonics_calculation'] = "analytic"
return hparams
def tune(positional_encoding_name, neural_network_name, dataset="landoceandataset"):
n_trials = 100
timeout = 4 * 60 * 60 # seconds
epochs = 100
if dataset == "landoceandataset":
datamodule = LandOceanDataModule()
num_classes = 1
regression = False
presence_only = False
loss_bg_weight = False
if dataset == "checkerboard":
datamodule = CheckerboardDataModule()
num_classes = 16
regression = False
presence_only = False
loss_bg_weight = False,
elif dataset == "inat2018":
datamodule = Inat2018DataModule("/data/sphericalharmonics/inat2018/")
num_classes = 8142
regression = False
presence_only = True
loss_bg_weight = 5
def objective(trial: optuna.trial.Trial) -> float:
hparams = get_hyperparameter(trial, positional_encoding_name, neural_network_name)
hparams["num_classes"] = num_classes
hparams["presence_only_loss"] = presence_only
hparams["loss_bg_weight"] = loss_bg_weight
hparams["regression"] = regression
spatialencoder = LocationEncoder(
positional_encoding_name,
neural_network_name,
hparams=hparams
)
trainer = pl.Trainer(
max_epochs=epochs,
log_every_n_steps=5,
accelerator='gpu',
callbacks=[EarlyStopping(monitor="val_loss", mode="min", patience=30)])
trainer.logger.log_hyperparams(hparams)
trainer.fit(model=spatialencoder, datamodule=datamodule)
return trainer.callback_metrics["val_loss"].item()
pruner = optuna.pruners.MedianPruner()
study_name = f"{dataset}-{positional_encoding_name}-{neural_network_name}"
os.makedirs(f"{TUNE_RESULTS_DIR}/{dataset}/runs/", exist_ok=True)
storage_name = f"sqlite:///{TUNE_RESULTS_DIR}/{dataset}/runs/{study_name}.db"
study = optuna.create_study(study_name=study_name, direction="minimize",
storage=storage_name, load_if_exists=True,
pruner=pruner)
study.optimize(objective, n_trials=n_trials, timeout=timeout)
print("Number of finished trials: {}".format(len(study.trials)))
print("Best trial:")
trial = study.best_trial
print(" Value: {}".format(trial.value))
print(" Params: ")
for key, value in trial.params.items():
print(" {}: {}".format(key, value))
study.trials_dataframe()
runsummary = f"{TUNE_RESULTS_DIR}/{dataset}/runs/{positional_encoding_name}-{neural_network_name}.csv"
os.makedirs(os.path.dirname(runsummary), exist_ok=True)
study.trials_dataframe().to_csv(runsummary)
def compile_summaries(dataset):
tune_results_dir_this_datset = os.path.join(TUNE_RESULTS_DIR, dataset)
runsdir = os.path.join(TUNE_RESULTS_DIR, f"{dataset}/runs")
csvs = [csv for csv in os.listdir(runsdir) if csv.endswith("csv") and csv != "summary.csv"]
summary = []
hparams = {}
for csv in csvs:
df = pd.read_csv(os.path.join(runsdir, csv))
best_run = df.sort_values(by="value").iloc[0]
value = best_run.value
params = {k.replace("params_", ""): v for k, v in best_run.to_dict().items() if "params" in k}
pe, nn = csv.replace(".csv", "").split("-")
hparams[f"{pe}-{nn}"] = params
sum = {
"pe":pe,
"nn":nn,
"value":value
}
sum.update(params)
summary.append(sum)
summary = pd.DataFrame(summary).sort_values("value").set_index(["pe","nn"])
summary.to_csv(os.path.join(tune_results_dir_this_datset, "summary.csv"))
print("writing " + os.path.join(tune_results_dir_this_datset, "hparams.yaml"))
with open(os.path.join(tune_results_dir_this_datset, "hparams.yaml"), 'w') as f:
yaml.dump(hparams, f)
value_matrix = pd.pivot_table(summary.value.reset_index(), index="pe", columns="nn", values=["value"])["value"]
print("writing " + os.path.join(tune_results_dir_this_datset, "values.csv"))
value_matrix.to_csv(os.path.join(tune_results_dir_this_datset, "values.csv"))
import matplotlib.pyplot as plt
fig, ax = plt.subplots()
ax.imshow(value_matrix)
ax.set_xticks(range(len(value_matrix.columns)))
ax.set_xticklabels(value_matrix.columns)
ax.set_xlabel(value_matrix.columns.name)
ax.set_yticks(range(len(value_matrix.index)))
ax.set_yticklabels(value_matrix.index)
ax.set_ylabel(value_matrix.index.name)
plt.tight_layout()
print("writing "+os.path.join(tune_results_dir_this_datset, "values.png"))
fig.savefig(os.path.join(tune_results_dir_this_datset, "values.png"), transparent=True, bbox_inches="tight", pad_inches=0)
if __name__ == '__main__':
#positional_encoders = ["theory", "direct", "cartesian3d", "grid"] # "sphericalharmonics",
#neural_networks = ["siren", "fcnet", "linear", "mlp"]
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--dataset", type=str, default="checkerboard", help="Name of the dataset")
args = parser.parse_args()
dataset = args.dataset
positional_encoders = ["spherem", "spheremplus"]
neural_networks = ["linear", "siren", "fcnet"]
for pe in positional_encoders:
for nn in neural_networks:
tune(pe, nn, dataset=dataset)
compile_summaries(dataset)