Perturbed the trained model but the outputs did not change

[Decem5150]

Dear Professor Lu Lu and DeepXDE Community,
I have encountered a problem when I try to perturb the model after it is trained.
After I perturbed the model (manually changed the trainable parameters along one direction), I computed the new output by calling model_perb.outputs_losses_test(inputs, target, aux), which should return the outputs and the losses of the model. However, the outputs and the losses were completely the same as in the original unperturbed model. Then I tried directly calling model_perb.net(inputs) and this time the outputs did actually change compared to the original model. I checked the source code of DeepXDE and could not figure out why this happened. Part of my code is as follows:

def get_weights(net):
    """ Extract parameters from net, and return a list of tensors"""
    return [p.data for p in net.parameters()]
def set_weights_1D(net, weights, directions=None, step=None):
    changes = [d*step for d in directions]
    for (p, w, d) in zip(net.parameters(), weights, changes):
        p.data = w + torch.Tensor(d).type(type(w))
# create a copy of the model
import copy
model_perb = copy.deepcopy(model)
w = get_weights(model_perb.net)
lam1=-3000
set_weights_1D(model_perb.net, w, directions=top_eigenvector[0], step=lam1)
inputs = model_perb.train_state.X_test
target = model_perb.train_state.y_test
aux = model_perb.train_state.test_aux_vars
outputs, losses = model_perb.outputs_losses_test(inputs, target, aux)
outputs_net = model_perb.net(torch.as_tensor(inputs))
print(outputs_net)
print(outputs)

The output was like

tensor([[ 2590.0410],
        [  -48.4030],
        [-1799.6990],
        ...,
        [ 2130.0547],
        [ -345.7226],
        [ 1395.6792]], grad_fn=<AddmmBackward0>)
tensor([[ 4.1955e-05],
        [ 1.8394e-04],
        [-5.0630e-05],
        ...,
        [ 6.6944e-01],
        [-7.1700e-02],
        [ 7.0247e-02]], grad_fn=<AddmmBackward0>)

Here, top_eigenvector[0] is the top eigenvector of the original model's Hessian matrix and is not zero (checked).
The perturbing step is a bit ridiculous but I also tried other numbers, which made no difference.

Thank you in advance.