Open trailing edge

[jethromoses]

This method generates airfoils with an open (blunt) trailing edge. Was the GANs also trained on data with open trailing edges? The paper says you trained the GANs on the UIUC database. Did you modify the UIUC database airfoils to have open trailing edges?

[wchen459]

There are airfoils with open trailing edges in the UIUC database. We did not modify the trailing edges in the data.

[jethromoses]

Is it possible to have the GANs generate airfoils with only sharp closed trailing edges?

[wchen459]

You can probably add some code after this line in the generator to force the last Bezier control point to be the same as the first one. This would make the generator produce closed Bezier curves.

[jethromoses]

Where do you specify the x location of the control point?

[wchen459]

In the generator, cp is a tensor with the shape (batch_size, num_control_points, 2). So cp[i, j, 0] is the x location of the j-th control point in the i-th sample.

[jethromoses]

So in your code, how would i do the following cp[:,0,0] = 1 cp[:,0,1] = 0 cp[:,-1,0]=1 cp[:,-1,1]=0

Basically, i'm trying to place the control points on the trailing edge. So the first and last control points should be at (1,0). What would the tensorflow command be for the above?

[wchen459]

You can try the following:

cp = tf.assign(cp[:,0,0], 1)
cp = tf.assign(cp[:,0,1], 0)
cp = tf.assign(cp[:,-1,0], 1)
cp = tf.assign(cp[:,-1,0], 0)

However, by manually setting the trailing edge coordinates, the distribution of generated airfoils can no longer align with the distribution of the data. This may or may not cause issues when training Bezier-GAN, depends on how different the two distributions can get and the capacity of the discriminator/generator. Anyways, you could try and see if further adjustment is needed.

[jethromoses]

tf.assign() doesn't work.

How does the generator decided where to place the bezier control point? Do you fix some (x,y) locations of the control points?

[wchen459]

tf.assign() doesn't work

Please provide more information (e.g., error message, code output, etc.).

How does the generator decided where to place the bezier control point? Do you fix some (x,y) locations of the control points?

The generator learns the position of the bezier control points from data. Please see our paper for more details.

[jethromoses]

With tf.assign(), this is the error I get,

ValueError: Tried to convert 'input' to a tensor and failed. Error: None values not supported.

Thanks for your help!

[wchen459]

The following code should work instead:

cp_te = tf.tile(tf.constant([[[1., 0.]]]), [tf.shape(cp)[0], 1, 1])
cp = tf.concat([cp_te, cp[:,1:-1], cp_te], axis=1)
cp = tf.identity(cp, name='control_point')

Also, delete the "name" argument in line 77:

cp = tf.squeeze(cp, axis=-1)

[jethromoses]

Thanks, that worked!