Generate airfoils with the help of a varational autoencoder VAE.
-
Uses public database
-
Training of a VAE to predict shapes
You can specify the following parameters:
--ih: Resolution of pressure and suction side (int)--learnRate: learning rate (float)--testTrainSplit: Test train split (float)--epochs: number of training epochs (int)--latentDim: latent dimensions (int)--batchSize: batch size (int)
.
.
Install the requirements with:
pip install -r requirements.txt
Data are taken from here: Illinois Airfoil Database.
To prepare the dataset run from project root dir:
python3 ./data/prepareDownload.py
This script will download all .dat files and place them in ./data/raw. For more information on the download script see here: Josh the engineer. To prepare the raw data run
python3 ./data/prepareRawData.py
in the same folder (./data). This will generated a json file for each .dat file with the following structure:
name: "name of the airfoil",
ss: [List x coordinates, List y coordinates],
ps: [List x coordinates, List y coordinates],
To train the variational autoencoder run
python training/train.py
By default, the model is trained with a batch norm of 8, learning rate of 1e-4, 3000 epochs and saved as weights-cpk.h5.
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
rgb (InputLayer) [(None, 256, 4)] 0 []
encoder (Functional) [(None, 2), 1580 ['rgb[0][0]']
(None, 2),
(None, 2)]
decoder (Functional) (None, 256, 4) 500 ['encoder[0][2]']
conv1d (Conv1D) (None, 256, 12) 156 ['rgb[0][0]']
average_pooling1d (AveragePool (None, 128, 12) 0 ['conv1d[0][0]']
ing1D)
conv1d_1 (Conv1D) (None, 128, 8) 296 ['average_pooling1d[0][0]']
conv1d_2 (Conv1D) (None, 128, 4) 100 ['conv1d_1[0][0]']
average_pooling1d_2 (AveragePo (None, 64, 4) 0 ['conv1d_2[0][0]']
oling1D)
flatten (Flatten) (None, 256) 0 ['average_pooling1d_2[0][0]']
dense_1 (Dense) (None, 2) 514 ['flatten[0][0]']
dense (Dense) (None, 2) 514 ['flatten[0][0]']
tf.math.subtract (TFOpLambda) (None, 256, 4) 0 ['rgb[0][0]',
'decoder[0][0]']
tf.__operators__.add (TFOpLamb (None, 2) 0 ['dense_1[0][0]']
da)
tf.math.square_1 (TFOpLambda) (None, 2) 0 ['dense[0][0]']
tf.math.square (TFOpLambda) (None, 256, 4) 0 ['tf.math.subtract[0][0]']
tf.math.subtract_1 (TFOpLambda (None, 2) 0 ['tf.__operators__.add[0][0]',
) 'tf.math.square_1[0][0]']
tf.math.exp (TFOpLambda) (None, 2) 0 ['dense_1[0][0]']
tf.math.reduce_sum (TFOpLambda () 0 ['tf.math.square[0][0]']
)
tf.math.subtract_2 (TFOpLambda (None, 2) 0 ['tf.math.subtract_1[0][0]',
) 'tf.math.exp[0][0]']
tf.math.multiply (TFOpLambda) () 0 ['tf.math.reduce_sum[0][0]']
tf.math.reduce_sum_1 (TFOpLamb (None,) 0 ['tf.math.subtract_2[0][0]']
da)
tf.math.truediv (TFOpLambda) () 0 ['tf.math.multiply[0][0]']
tf.math.multiply_1 (TFOpLambda (None,) 0 ['tf.math.reduce_sum_1[0][0]']
)
tf.__operators__.add_1 (TFOpLa (None,) 0 ['tf.math.truediv[0][0]',
mbda) 'tf.math.multiply_1[0][0]']
tf.math.reduce_mean (TFOpLambd () 0 ['tf.__operators__.add_1[0][0]']
a)
add_loss (AddLoss) () 0 ['tf.math.reduce_mean[0][0]']
==================================================================================================
Total params: 2,080
Trainable params: 2,080
Non-trainable params: 0
__________________________________________________________________________________________________
You can test your model using the file inference.ipynb. Use the following prompt to predict the rotation of a specific image: