def loss_func(h_net, input: torch.Tensor):
ones = torch.unsqueeze(
torch.ones(len(input),
dtype=richards_celia.DTYPE,
device=richards_celia.DEVICE), 1)
predicted_h_trial = h_net(tz_pairs) # ts(input)
predicted_h_initial_boundary = h_net(
tz_pairs_initial_boundary) # ts(tz_pairs_initial_boundary)
predicted_h_bottom_boundary = h_net(tz_pairs_bottom_boundary)
predicted_h_top_boundary = h_net(tz_pairs_top_boundary)
# predicted_h_top_boundary_d = torch.autograd.grad(
# predicted_h_top_boundary,
# tz_pairs_top_boundary,
# create_graph=True,
# grad_outputs=torch.unsqueeze(
# torch.ones(len(tz_pairs_top_boundary), dtype=richards_celia.DTYPE, device=richards_celia.DEVICE), 1)
# )[0]
# predicted_h_top_boundary_dz = predicted_h_top_boundary_d[:, 1:2]
#
# predicted_q_top_boundary = K(predicted_h_top_boundary) * (1 - predicted_h_top_boundary_dz)
predicted_h_d = torch.autograd.grad(predicted_h_trial,
input,
create_graph=True,
grad_outputs=ones)[0]
predicted_h_dz = predicted_h_d[:, 1:2]
predicted_theta = theta(predicted_h_trial)
predicted_K = K(predicted_h_trial)
predicted_theta_d = torch.autograd.grad(
predicted_theta,
input,
create_graph=True,
grad_outputs=ones,
)[0]
predicted_theta_dt = predicted_theta_d[:, 0:1]
predicted_second_term = predicted_K * predicted_h_dz
predicted_second_term_d = torch.autograd.grad(
predicted_second_term,
input,
create_graph=True,
grad_outputs=ones,
)[0]
predicted_second_term_dz = predicted_second_term_d[:, 1:2]
predicted_K_d = torch.autograd.grad(
predicted_K,
input,
create_graph=True,
grad_outputs=ones,
)[0]
predicted_K_dz = predicted_K_d[:, 1:2]
# @TODO check the signs here
residual = predicted_theta_dt - predicted_second_term_dz - predicted_K_dz
residual_h_initial_boundary = predicted_h_initial_boundary - h_initial_boundary
residual_h_bottom_boundary = predicted_h_bottom_boundary - h_bottom_boundary
residual_h_top_boundary = predicted_h_top_boundary - h_top_boundary
# residual_q_top_boundary = predicted_q_top_boundary - q_top_boundary
# fitting_error = torch.unsqueeze(torch.zeros(len(input), dtype=richards_celia.DTYPE, device=richards_celia.DEVICE),1)
# fitting_error = predicted_wrc - theta
# assert (fitting_error == fitting_error).any()
# + torch.sum(fitting_error ** 2) \
pde_res = torch.sum(residual**2) / len(residual)
boundary_bottom_res = torch.sum(residual_h_bottom_boundary**
2) / len(residual_h_bottom_boundary)
boundary_top_res = torch.sum(residual_h_top_boundary**
2) / len(residual_h_top_boundary)
# boundary_top_res = torch.sum(residual_q_top_boundary ** 2) / len(residual_q_top_boundary)
boundary_initial_res = torch.sum(residual_h_initial_boundary**
2) / len(residual_h_initial_boundary)
print("pde_res: " + str(pde_res))
print("boundary_bottom_res: " + str(boundary_bottom_res))
print("boundary_top_res: " + str(boundary_top_res))
print("boundary_initial_res: " + str(boundary_initial_res))
loss = pde_res + (boundary_bottom_res + boundary_top_res +
boundary_initial_res)
# loss = pde_res + boundary_initial_res
# assert (loss == loss).any()
h_net.pde_loss += [pde_res]
h_net.bc_initial_losses += [boundary_initial_res]
h_net.bc_top_losses += [boundary_top_res]
h_net.bc_bottom_losses += [boundary_bottom_res]
h_net.losses += [loss]
return loss
def loss_func(input: torch.tensor):
ones = torch.unsqueeze(torch.ones(len(input), dtype=richards_celia.DTYPE, device=richards_celia.DEVICE), 1)
# predicted_h = h_net(tz_pairs)
predicted_h_trial = h_net(input)
predicted_h_initial_boundary = h_net(tz_pairs_initial_boundary)
predicted_h_bottom_boundary = h_net(tz_pairs_bottom_boundary)
predicted_h_top_boundary = h_net(tz_pairs_top_boundary)
predicted_h_d = torch.autograd.grad(
predicted_h_trial,
input,
create_graph=True,
grad_outputs=ones
)[0]
predicted_h_dz = predicted_h_d[:, 1:2]
predicted_theta = theta(predicted_h_trial)
predicted_K = K(predicted_h_trial)
predicted_theta_d = torch.autograd.grad(
predicted_theta,
input,
create_graph=True,
grad_outputs=ones,
)[0]
predicted_theta_dt = predicted_theta_d[:, 0:1]
predicted_second_term = predicted_K * predicted_h_dz
predicted_second_term_d = torch.autograd.grad(
predicted_second_term,
input,
create_graph=True,
grad_outputs=ones,
)[0]
predicted_second_term_dz = predicted_second_term_d[:, 1:2]
predicted_K_d = torch.autograd.grad(
predicted_K,
input,
create_graph=True,
grad_outputs=ones,
)[0]
predicted_K_dz = predicted_K_d[:, 1:2]
residual = predicted_theta_dt - predicted_second_term_dz - predicted_K_dz
residual_h_initial_boundary = predicted_h_initial_boundary - h_initial_boundary
residual_h_bottom_boundary = predicted_h_bottom_boundary - h_bottom_boundary
residual_h_top_boundary = predicted_h_top_boundary - h_top_boundary
# fitting_error = torch.unsqueeze(torch.zeros(len(input), dtype=richards_celia.DTYPE, device=richards_celia.DEVICE),1)
# fitting_error = predicted_wrc - theta
# assert (fitting_error == fitting_error).any()
# + torch.sum(fitting_error ** 2) \
loss = torch.sum(residual ** 2) / len(residual) \
+ torch.sum(residual_h_initial_boundary ** 2) / len(residual_h_initial_boundary) \
+ torch.sum(residual_h_bottom_boundary ** 2) / len(residual_h_bottom_boundary) \
+ torch.sum(residual_h_top_boundary ** 2) / len(residual_h_top_boundary)
# assert (loss == loss).any()
return loss