Predicting the Gradient of DeepONet Solutions

[LanPeng-94]

Hi,

I use the y_pred = model.predict(data.test_x), and the shape of y_pred is 200*841. Now, I want to get the gradient of each DeepONet solution. If the inputs of the trunk are x1 and x2, I write the following code,

def dydx(inputs, outputs):
    x_loc = inputs[1]
    du_x = tf.gradients(outputs, x_loc)[0][:, 0:1]
    print("du_x_Grad shape {}".format(du_x.shape))
    return du_x

dy_x = model.predict(data.test_x, operator=dydx)
However, the shape of dy_x is 841*1. It doesn't seem to have obtained the gradient corresponding to each predicted value, what should I do.

Best regards,

Peng

[lululxvi]

Do you use DeepONet or DeepONetCartesianProd?

[LanPeng-94]

Hi,

I use DeepONetCartesianProd to build my surrogate model.
I guess it seems to only get the gradient of the last function. Because the shape of y_pred is 200*841, when I use tf.gradients(outputs, x_loc)[0][:, 0:1], it can not obtain the gradient for each discrete functions. Based on this, I modify my code,

y_pred = model.predict(data.test_x)

dy_xs = []
for i in range(200):
    def dydx(inputs, outputs):
        x_loc = inputs[1]
        du_x = tf.gradients(outputs[i:i+1, :], x_loc)[0][:, 0:1]
        # print("du_x_Grad shape {}".format(du_x.shape))
        return du_x
    dy_x = model.predict(data.test_x, operator=dydx).T
    print(dy_x.shape)
    dy_xs.append(dy_x)
    
dy_xs = np.array(dy_xs).astype(np.float32)
print("dy_xs shape {}".format(dy_xs.shape))

I want to know if my understanding is correct. Is there any problem with my code?

Best regards,

Peng

[lululxvi]

You can read the code here to know how to compute the gradients

deepxde/deepxde/data/pde_operator.py

Line 234 in 2236965

def _losses(self, outputs, loss_fn, inputs, model, num_func):