Note
Go to the end to download the full example code.
High Level API for scaling, training and prediction¶
In this section, we will use high level API for performing machine learning process.
At First, we will prepare dummy data. These dummy data corresponds to feature values extracted from simultion data.
import pathlib
import random
import shutil
import numpy as np
import scipy.sparse as sp
def prepare_sample_interim_files():
np.random.seed(0)
random.seed(0)
output_directory = pathlib.Path("out")
if output_directory.exists():
shutil.rmtree(output_directory)
base_interim_dir = output_directory / "interim"
base_interim_dir.mkdir(parents=True)
n_cases = 5
dtype = np.float32
for i in range(n_cases):
n_nodes = 100 * (i + 1)
interim_dir = base_interim_dir / f"case_{i}"
interim_dir.mkdir()
nodal_initial_u = np.random.rand(n_nodes, 3, 1)
np.save(
interim_dir / "nodal_initial_u.npy",
nodal_initial_u.astype(dtype),
)
# nodal_last_u = np.random.rand(n_nodes, 3, 1)
np.save(interim_dir / "nodal_last_u.npy", nodal_initial_u.astype(dtype))
sparse_array_names = [
"nodal_nadj",
"nodal_x_grad_hop1",
"nodal_y_grad_hop1",
"nodal_z_grad_hop1",
]
rng = np.random.default_rng()
for name in sparse_array_names:
arr = sp.random(n_nodes, n_nodes, density=0.1, random_state=rng)
sp.save_npz(interim_dir / name, arr.tocoo().astype(dtype))
(interim_dir / "converted").touch()
prepare_sample_interim_files()
Setting file for scaling and training can be downloaded from data.yml we perform scaling process for data above.
from phlower.services.preprocessing import PhlowerScalingService
from phlower.settings import PhlowerSetting
setting = PhlowerSetting.read_yaml("sample_data/e2e/setting.yml")
scaler = PhlowerScalingService.from_setting(setting)
scaler.fit_transform_all(
interim_data_directories=[
pathlib.Path("out/interim/case_0"),
pathlib.Path("out/interim/case_1"),
pathlib.Path("out/interim/case_2"),
pathlib.Path("out/interim/case_3"),
pathlib.Path("out/interim/case_4"),
],
output_base_directory=pathlib.Path("out/preprocessed"),
)
Next, we perform training by using preprocessed data.
from phlower.services.trainer import PhlowerTrainer
trainer = PhlowerTrainer.from_setting(setting)
loss = trainer.train(
train_directories=[
pathlib.Path("out/preprocessed/case_0"),
pathlib.Path("out/preprocessed/case_1"),
pathlib.Path("out/preprocessed/case_2"),
],
validation_directories=[
pathlib.Path("out/preprocessed/case_3"),
pathlib.Path("out/preprocessed/case_4"),
],
output_directory=pathlib.Path("out/model"),
)
------- Start Training -------
Device: cpu
torch.cuda.is_available()=False
Log every 1 epoch
Output directory: out/model
-------------------------------
epoch train_loss validation_loss details_tr/nodal_last_u details_val/nodal_last_u elapsed_time
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/2 [00:00<?, ?it/s]
0 3.60635e-01 3.49816e-01 3.60635e-01 3.49816e-01 0.04
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/2 [00:00<?, ?it/s]
1 3.52956e-01 3.42368e-01 3.52956e-01 3.42368e-01 0.08
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/2 [00:00<?, ?it/s]
2 3.45437e-01 3.35074e-01 3.45437e-01 3.35074e-01 0.12
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/2 [00:00<?, ?it/s]
3 3.38075e-01 3.27933e-01 3.38075e-01 3.27933e-01 0.16
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/2 [00:00<?, ?it/s]
4 3.30867e-01 3.20941e-01 3.30867e-01 3.20941e-01 0.20
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/2 [00:00<?, ?it/s]
5 3.23809e-01 3.14095e-01 3.23809e-01 3.14095e-01 0.23
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/2 [00:00<?, ?it/s]
6 3.16899e-01 3.07392e-01 3.16899e-01 3.07392e-01 0.27
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/2 [00:00<?, ?it/s]
7 3.10133e-01 3.00829e-01 3.10133e-01 3.00829e-01 0.31
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/2 [00:00<?, ?it/s]
8 3.03509e-01 2.94404e-01 3.03509e-01 2.94404e-01 0.35
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/3 [00:00<?, ?it/s]
0%| | 0/2 [00:00<?, ?it/s]
9 2.97024e-01 2.88113e-01 2.97024e-01 2.88113e-01 0.39
train function returns float which corresponds to last training loss.
Let’s call print it.
print(loss)
0.29843243956565857
Finally, we perform predicion by using pretrained model. Setting file for prediction can be downloaded from data.yml
It is found that physical dimension is also considered properly.
from phlower.services.predictor import PhlowerPredictor
setting = PhlowerSetting.read_yaml("sample_data/e2e/predict.yml")
predictor = PhlowerPredictor(
model_directory=pathlib.Path("out/model"),
predict_setting=setting.prediction,
)
preprocessed_directories = [pathlib.Path("out/preprocessed/case_3")]
for result in predictor.predict(
preprocessed_directories, perform_inverse_scaling=False
):
for k in result.prediction_data.keys():
print(f"{k}: {result.prediction_data[k].dimension}")
nodal_last_u: PhlowerDimensionTensor(T: -1.0, L: 1.0, M: 0.0, I: 0.0, Theta: 0.0, N: 0.0, J: 0.0)
Total running time of the script: (0 minutes 4.136 seconds)