Skip to content

Solver logic refactor #64

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Gromwud merged 2 commits into from
Mar 5, 2026
Merged

Solver logic refactor #64

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4595507
PIC refactor
Gromwud Mar 5, 2026
b2e3399
Sobolev Learning refactor
Gromwud Mar 5, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
137 changes: 91 additions & 46 deletions epde/globals.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,11 +7,14 @@
"""

from dataclasses import dataclass
import copy
import warnings
from typing import List, Union

import numpy as np
import torch
import torch.nn.functional as F
import epde.globals as global_var
# device = torch.device('cpu') # TODO: make system-agnostic approach

from epde.cache.cache import Cache
Expand Down Expand Up @@ -133,87 +136,129 @@ def reset_control_nn(n_control: int = 1, ann: torch.nn.Sequential = None,


def reset_data_repr_nn(data: List[np.ndarray], grids: List[np.ndarray], train: bool = True,
derivs: List[Union[int, List, Union[np.ndarray]]] = None,
derivs: List[Union[int, List, Union[np.ndarray]]] = None,
penalised_derivs: List[Union[int, List]] = None,
epochs_max=1e3, predefined_ann: torch.nn.Sequential = None,
batch_frac=0.5, learining_rate=1e-6, device = 'cpu',
use_fourier: bool = True, fourier_params: dict = {'L' : [4,], 'M' : [3,]}):
batch_frac=0.5, val_frac=0.2, learning_rate=1e-4, device='cpu',
use_fourier: bool = True, fourier_params: dict = None,
deriv_weight=1, penalty_weight=1e3):
'''
Represent the data with ANN, suitable to be used as the initial guess of the candidate equations solutions
Represent the data with ANN, suitable to be used as the initial guess of the candidate equations solutions
during the equation search, employing solver-based fitness function.

Possible addition: add optimization in Sobolev space, using passed derivatives, incl. higher orders.
'''

if fourier_params is None:
fourier_params = {'L': [4,], 'M': [3,]}

global solution_guess_nn

if predefined_ann is None:
model = create_solution_net(equations_num=len(data), domain_dim=len(grids), device = device,
model = create_solution_net(equations_num=len(data), domain_dim=len(grids), device=device,
use_fourier=use_fourier, fourier_params=fourier_params)
# model = NN(Num_Hidden_Layers=5, Neurons_Per_Layer=50, Input_Dim=len(grids), Activation_Function='Tanh')
# model = NN(Num_Hidden_Layers=5, Neurons_Per_Layer=50, Input_Dim=len(grids), Activation_Function='Rational')


else:
model = predefined_ann

if train:
model.to(device)
model = model.to(device)

grids_tr = torch.from_numpy(np.array([subgrid.reshape(-1) for subgrid in grids])).float().T
data_tr = torch.from_numpy(np.array([data_var.reshape(-1) for data_var in data])).float().T
grids_tr = torch.from_numpy(np.array([subgrid[global_var.grid_cache.g_func != 0].reshape(-1) for subgrid in grids])).float().T
data_tr = torch.from_numpy(np.array([data_var[global_var.grid_cache.g_func != 0].reshape(-1) for data_var in data])).float().T
grids_tr = grids_tr.to(device)
data_tr = data_tr.to(device)

batch_size = int(data[0].size * batch_frac)
optimizer = torch.optim.Adam(model.parameters(), lr = learining_rate)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 2000, gamma=0.5)
n_total = grids_tr.size()[0]
n_val = max(1, int(n_total * val_frac))
perm_split = torch.randperm(n_total)
val_indices = perm_split[:n_val]
train_indices = perm_split[n_val:]

grids_val, data_val = grids_tr[val_indices], data_tr[val_indices]
grids_train, data_train = grids_tr[train_indices], data_tr[train_indices]

batch_size = int(data_train.size()[0] * batch_frac)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
epochs_max = int(epochs_max)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=max(1, epochs_max // 20), factor=0.5)
deriv_calc = AutogradDeriv()

t = 0
min_loss = np.inf
loss_mean = np.inf
print(f'Training ANN to represent input data on {epochs_max} epochs:')
while loss_mean > 1e-6 and t < epochs_max:
best_state = None
min_val_loss = np.inf
val_patience = max(1, epochs_max // 10)
val_no_improve = 0

permutation = torch.randperm(grids_tr.size()[0])
print(f'Training ANN to represent input data on {epochs_max} epochs:')
for t in range(epochs_max):
permutation = torch.randperm(grids_train.size()[0])

loss_list = []

for i in range(0, grids_tr.size()[0], batch_size):
for i in range(0, grids_train.size()[0], batch_size):
optimizer.zero_grad()

indices = permutation[i:i+batch_size]
batch_x, batch_y = grids_tr[indices], data_tr[indices]

loss = torch.mean(torch.abs(batch_y - model(batch_x)))
if derivs is not None:
for var_idx, deriv_axes, deriv_tensor in derivs:
deriv_autograd = deriv_calc.take_derivative(model, batch_x, axes = deriv_axes, component = var_idx)
batch_derivs = torch.from_numpy(deriv_tensor)[torch.unravel_index(indices,
deriv_tensor.shape)].reshape_as(deriv_autograd).to(device)

loss_add = 1e2 * torch.mean(torch.abs(batch_derivs - deriv_autograd))
loss += loss_add
batch_x = grids_train[indices].detach().requires_grad_(True)
batch_y = data_train[indices]

pred = model(batch_x)
loss = F.mse_loss(pred, batch_y) / (torch.mean(batch_y ** 2) + 1e-8)
# if derivs is not None:
# deriv_loss = 0
# for var_idx, deriv_axes, deriv_tensor in derivs:
# deriv_autograd = deriv_calc.take_derivative(model, batch_x, axes=deriv_axes, component=var_idx)
# flat_indices = train_indices[indices]
# batch_derivs = torch.from_numpy(deriv_tensor.reshape(-1))[flat_indices].reshape_as(deriv_autograd).float().to(device)
# deriv_loss += deriv_weight * F.mse_loss(deriv_autograd, batch_derivs) / (torch.mean(batch_derivs ** 2) + 1e-8)
#
# loss += deriv_loss / len(derivs)

if penalised_derivs is not None:
for var_idx, deriv_axes in derivs:
deriv_autograd = deriv_calc.take_derivative(model, batch_x, axes = deriv_axes, component = var_idx)
batch_derivs = torch.from_numpy(deriv_tensor)[torch.unravel_index(indices,
deriv_tensor.shape)].reshape_as(deriv_autograd).to(device)
higher_ord_penalty = 1e3 * torch.mean(torch.abs(deriv_autograd))

loss += higher_ord_penalty
for var_idx, deriv_axes in penalised_derivs:
deriv_autograd = deriv_calc.take_derivative(model, batch_x, axes=deriv_axes, component=var_idx)
loss += penalty_weight * torch.mean(torch.abs(deriv_autograd))

loss.backward()
optimizer.step()
loss_list.append(loss.item())
scheduler.step()
loss_mean = np.mean(loss_list)
if loss_mean < min_loss:
best_model = model
min_loss = loss_mean
t += 1
model = best_model
print(f'min loss is {min_loss}, in last epoch: {loss_list}.')
solution_guess_nn = best_model

train_loss_mean = np.mean(loss_list)

with torch.no_grad():
val_loss = (F.mse_loss(model(grids_val), data_val) / (torch.mean(data_val ** 2) + 1e-8)).item()

# if derivs is not None:
# grids_val_grad = grids_val.detach().requires_grad_(True)
# deriv_loss = 0
# for var_idx, deriv_axes, deriv_tensor in derivs:
# deriv_autograd = deriv_calc.take_derivative(model, grids_val_grad, axes=deriv_axes, component=var_idx)
# batch_derivs = torch.from_numpy(deriv_tensor.reshape(-1))[val_indices].reshape_as(deriv_autograd).float().to(device)
# deriv_loss += deriv_weight * F.mse_loss(deriv_autograd, batch_derivs).item() / (torch.mean(batch_derivs ** 2).item() + 1e-8)
#
# val_loss += deriv_loss / len(derivs)

scheduler.step(val_loss)

if val_loss < min_val_loss:
best_state = copy.deepcopy(model.state_dict())
min_val_loss = val_loss
val_no_improve = 0
else:
val_no_improve += 1

if t % 100 == 0 and t != 0:
print(f"Epoch {t:4d} | Train Loss: {train_loss_mean:.6e} | Val Loss: {val_loss:.6e}")

if val_no_improve >= val_patience:
print(f"Early stopping at epoch {t}, best val loss: {min_val_loss:.6e}")
break

if best_state is not None:
model.load_state_dict(best_state)
print(f'Best val loss: {min_val_loss:.6e}, final train loss: {train_loss_mean:.6e}')
solution_guess_nn = model
else:
solution_guess_nn = model
Loading