implement ML correction

This commit is contained in:
Filippo Olivo
2025-11-18 21:55:54 +01:00
parent 1c7b593762
commit d865556c9f
3 changed files with 64 additions and 135 deletions

View File

@@ -4,6 +4,7 @@ from torch_geometric.data import Batch
import importlib
from matplotlib import pyplot as plt
from matplotlib.tri import Triangulation
from .model.finite_difference import FiniteDifferenceStep
def import_class(class_path: str):
@@ -56,6 +57,7 @@ class GraphSolver(LightningModule):
):
super().__init__()
self.model = import_class(model_class_path)(**model_init_args)
self.fd_net = FiniteDifferenceStep()
self.loss = loss if loss is not None else torch.nn.MSELoss()
self.curriculum_learning = curriculum_learning
self.start_iters = start_iters
@@ -67,6 +69,8 @@ class GraphSolver(LightningModule):
self.automatic_optimization = False
self.threshold = 1e-5
self.aplha = 0.1
def _compute_deg(self, edge_index, edge_attr, num_nodes):
deg = torch.zeros(num_nodes, device=edge_index.device)
deg = deg.scatter_add(0, edge_index[1], edge_attr)
@@ -96,8 +100,15 @@ class GraphSolver(LightningModule):
def _compute_model_steps(
self, x, edge_index, edge_attr, deg, boundary_mask, boundary_values
):
out = self.model(x, edge_index, edge_attr, deg)
with torch.no_grad():
out = self.fd_net(x, edge_index, edge_attr, deg)
out[boundary_mask] = boundary_values.unsqueeze(-1)
# diff = out - x
correction = self.model(x, edge_index, edge_attr, deg)
out = out + self.aplha * correction
out[boundary_mask] = boundary_values.unsqueeze(-1)
# out = self.model(x, edge_index, edge_attr, deg)
# out[boundary_mask] = boundary_values.unsqueeze(-1)
return out
def _check_convergence(self, out, x):
@@ -132,11 +143,7 @@ class GraphSolver(LightningModule):
deg = self._compute_deg(edge_index, edge_attr, x.size(0))
losses = []
acc_loss, acc_it = 0, 0
max_acc_iters = (
self.current_iters // self.accumulation_iters + 1
if self.accumulation_iters is not None
else 1
)
for i in range(self.current_iters):
out = self._compute_model_steps(
x,

View File

@@ -1,6 +1,7 @@
import torch
import torch.nn as nn
from torch_geometric.nn import MessagePassing
from torch.nn.utils import spectral_norm
class FiniteDifferenceStep(MessagePassing):
@@ -8,14 +9,8 @@ class FiniteDifferenceStep(MessagePassing):
TODO: add docstring.
"""
def __init__(self, aggr: str = "add", root_weight: float = 1.0):
super().__init__(aggr=aggr)
assert (
aggr == "add"
), "Per somme pesate, l'aggregazione deve essere 'add'."
# self.root_weight = float(root_weight)
self.p = torch.nn.Parameter(torch.tensor(1.0))
self.a = root_weight
def __init__(self):
super().__init__(aggr="add")
def forward(self, x, edge_index, edge_attr, deg):
"""
@@ -28,8 +23,14 @@ class FiniteDifferenceStep(MessagePassing):
"""
TODO: add docstring.
"""
p = torch.clamp(self.p, 0.0, 1.0)
return p * edge_attr.view(-1, 1) * x_j
# return self.message_net(x_j * edge_attr)
return x_j * edge_attr
def update(self, aggr_out, _):
"""
TODO: add docstring.
"""
return aggr_out
def aggregate(self, inputs, index, deg):
"""
@@ -38,82 +39,3 @@ class FiniteDifferenceStep(MessagePassing):
out = super().aggregate(inputs, index)
deg = deg + 1e-7
return out / deg.view(-1, 1)
def update(self, aggr_out, x):
"""
TODO: add docstring.
"""
return aggr_out
class GraphFiniteDifference(nn.Module):
"""
TODO: add docstring.
"""
def __init__(self, max_iters: int = 5000, threshold: float = 1e-4):
"""
TODO: add docstring.
"""
super().__init__()
self.max_iters = max_iters
self.threshold = threshold
self.fd_step = FiniteDifferenceStep(aggr="add", root_weight=1.0)
@staticmethod
def _compute_deg(edge_index, edge_attr, num_nodes):
"""
TODO: add docstring.
"""
deg = torch.zeros(num_nodes, device=edge_index.device)
deg = deg.scatter_add(0, edge_index[1], edge_attr)
return deg + 1e-7
@staticmethod
def _compute_c_ij(c, edge_index):
"""
TODO: add docstring.
"""
return (0.5 * (c[edge_index[0]] + c[edge_index[1]])).squeeze()
def forward(
self,
x,
edge_index,
edge_attr,
c,
boundary_mask,
boundary_values,
**kwargs,
):
"""
TODO: add docstring.
"""
edge_attr = 1 / edge_attr[:, -1]
c_ij = self._compute_c_ij(c, edge_index)
edge_attr = edge_attr * c_ij
deg = self._compute_deg(edge_index, edge_attr, x.size(0))
# Calcola la soglia staccando x dal grafo
conv_thres = self.threshold * torch.norm(x.detach())
for _i in range(self.max_iters):
out = self.fd_step(x, edge_index, edge_attr, deg)
out[boundary_mask] = boundary_values.unsqueeze(-1)
# Controllo convergenza senza tracciamento gradienti
with torch.no_grad():
residual_norm = torch.norm(out - x)
if residual_norm < conv_thres:
break
# --- OTTIMIZZAZIONE CHIAVE ---
# Stacca 'out' dal grafo prima della prossima iterazione
# per evitare BPTT e risparmiare memoria.
x = out.detach()
# Il 'out' finale restituito mantiene i gradienti
# dell'ULTIMA chiamata a fd_step, permettendo al modello
# di apprendere correttamente.
return out, _i + 1

View File

@@ -1,53 +1,53 @@
import torch
import torch.nn as nn
from torch_geometric.nn import MessagePassing
from torch.nn.utils import spectral_norm
# from torch.nn.utils import spectral_norm
class FiniteDifferenceStep(MessagePassing):
"""
TODO: add docstring.
"""
def __init__(self, hidden_dim=16, aggr: str = "add"):
print(aggr)
super().__init__(aggr=aggr)
self.x_embedding = nn.Sequential(
spectral_norm(nn.Linear(1, hidden_dim // 2)),
nn.GELU(),
spectral_norm(nn.Linear(hidden_dim // 2, hidden_dim)),
class GCNConvLayer(MessagePassing):
def __init__(self, in_channels, out_channels):
super().__init__("add")
self.lin = nn.Sequential(
nn.Linear(in_channels, out_channels),
nn.ReLU(),
nn.Linear(out_channels, out_channels),
nn.ReLU(),
)
self.out_net = nn.Sequential(
spectral_norm(nn.Linear(hidden_dim, hidden_dim // 2)),
nn.GELU(),
spectral_norm(nn.Linear(hidden_dim // 2, 1)),
def _compute_edge_weight(self, edge_index, edge_w, deg):
""" """
return edge_w.squeeze() / (
1 + torch.sqrt(deg[edge_index[0]] * deg[edge_index[1]])
)
def forward(self, x, edge_index, edge_attr, deg):
"""
TODO: add docstring.
"""
x_ = self.x_embedding(x)
out = self.propagate(edge_index, x=x_, edge_attr=edge_attr, deg=deg)
return self.out_net(out)
edge_w = self._compute_edge_weight(edge_index, edge_attr, deg)
return self.propagate(edge_index, x=x, edge_weight=edge_w, deg=deg)
def message(self, x_j, edge_attr):
"""
TODO: add docstring.
"""
return x_j * edge_attr.view(-1, 1)
def message(self, x_j, edge_weight):
return edge_weight.view(-1, 1) * x_j
def update(self, aggr_out, _):
"""
TODO: add docstring.
"""
return aggr_out
def aggregate(self, inputs, index, deg):
"""
TODO: add docstring.
"""
out = super().aggregate(inputs, index)
deg = deg + 1e-7
return out / deg.view(-1, 1)
class CorrectionNet(nn.Module):
def __init__(self, hidden_dim=8):
super().__init__()
self.enc = nn.Sequential(
nn.Linear(1, hidden_dim // 2),
nn.ReLU(),
nn.Linear(hidden_dim // 2, hidden_dim),
nn.ReLU(),
)
self.model = GCNConvLayer(hidden_dim, hidden_dim)
self.dec = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim // 2),
nn.ReLU(),
nn.Linear(hidden_dim // 2, 1),
nn.ReLU(),
)
def forward(self, x, edge_index, edge_attr, deg):
h = self.enc(x)
h = self.model(h, edge_index, edge_attr, deg)
out = self.dec(h)
return out