having a trainable parameter in the reference solution

[Burhan3228]

Dear all, in the code below, i have the reference solution as:

` def fun_initnewTF(x):

    condition1 = tf.greater_equal(x[:, 0:1], 0.0)
    condition2 = tf.greater_equal(xstensor - x[:, 0:1], 0.0)
    condition3 = tf.greater_equal(xstensor + x[:, 0:1], 0.0)
    
    result1 = tf.where(condition2, tf.ones_like(x[:, 0:1]), tf.zeros_like(x[:, 0:1]))
    result2 = tf.where(condition3, tf.ones_like(x[:, 0:1]), tf.zeros_like(x[:, 0:1]))
    
    result =  tf.where(condition1, result1, result2)
    
    return result


def mycritiinewnewTF(x, theta):
    numerator = 3 * theta * tf.sqrt(2.0)
    denominator = (1 + theta) * tf.sqrt(2.0) + tf.math.cosh(
        x * tf.sqrt(theta)
    ) * tf.sqrt(2 - 5 * theta + 2 * theta ** 2)
    return numerator / denominator

def mycritiisonnewnewTF(x):
    spatial_component = x[:, 0:1]  # Ensure operation on spatial component only
    return mycritiinewnewTF(spatial_component, theta=0.45)

def reference_solutionTF(x):
    # x should be a TensorFlow tensor of shape (N, 2) where each row is [space, time]
    spatial_component = x[:, 0:1]  # Extract spatial component
    time_component = x[:, 1:2]     # Extract time component
    # Initialize an output tensor
    result = tf.zeros_like(spatial_component)
    
    # Apply conditions based on the time component
    zero_mask = tf.equal(time_component, 0.0)
    nonzero_mask = tf.not_equal(time_component, 0.0)
    
    result = tf.where(zero_mask, US_trainable*fun_initnewTF(spatial_component), result)
    result = tf.where(nonzero_mask, mycritiisonnewnewTF(x), result)
    
    return result`

Here, US_trainable is my trainable parameter , in other word the reference solution itself depends on the trainable parameter. When i use it the data as:

data = dde.data.TimePDE( geomtime, pde, [bc,ic,observe_y_t_vals_except_inits], num_domain=1024, num_boundary=16, num_initial=256, solution=reference_solutionTF, anchors=combine_except_inits, train_distribution="Sobol", num_test=10000 )

it produces the following error:

TypeError: The value of a feed cannot be a tf.Tensor object. Acceptable feed values include Python scalars, strings, lists, numpy ndarrays, or TensorHandles. For reference, the tensor object was Tensor("Select_24:0", shape=(21768, 1), dtype=float32) which was passed to the argument `feed_dict` with key Tensor("Placeholder_15:0", shape=(None, 1), dtype=float32).

Apparently, it doesn't accept the reference solution in tensorflow environment and it asks its numpy version. Here is the problem: I d like to have US_trainable to be trained during training and the reference solution therefore also need to be adjusted accordingly during training. how can i achieve it? thanks in advance.

[lululxvi]

Reference solution cannot have unknown variables.