Gradient accumulation in BPTT (#2)
This commit is contained in:
GitHub
@@ -2,6 +2,40 @@ import torch
|
||||
import torch.nn as nn
|
||||
from torch_geometric.nn import MessagePassing
|
||||
from torch.nn.utils import spectral_norm
|
||||
from matplotlib.tri import Triangulation
|
||||
from matplotlib import pyplot as plt
|
||||
|
||||
|
||||
def _plot_mesh(y_pred, batch, iteration=None):
|
||||
|
||||
idx = batch.batch == 0
|
||||
y = batch.y[idx].detach().cpu()
|
||||
y_pred = y_pred[idx].detach().cpu()
|
||||
pos = batch.pos[idx].detach().cpu()
|
||||
|
||||
pos = pos.detach().cpu()
|
||||
tria = Triangulation(pos[:, 0], pos[:, 1])
|
||||
plt.figure(figsize=(18, 5))
|
||||
plt.subplot(1, 3, 1)
|
||||
plt.tricontourf(tria, y.squeeze().numpy(), levels=14)
|
||||
plt.colorbar()
|
||||
plt.title("True temperature")
|
||||
plt.subplot(1, 3, 2)
|
||||
plt.tricontourf(tria, y_pred.squeeze().numpy(), levels=14)
|
||||
plt.colorbar()
|
||||
plt.title("Predicted temperature")
|
||||
plt.subplot(1, 3, 3)
|
||||
plt.tricontourf(tria, torch.abs(y_pred - y).squeeze().numpy(), levels=14)
|
||||
plt.colorbar()
|
||||
plt.title("Error")
|
||||
plt.suptitle("GNO", fontsize=16)
|
||||
name = (
|
||||
f"images/gno_iter_{iteration:04d}.png"
|
||||
if iteration is not None
|
||||
else "gno.png"
|
||||
)
|
||||
plt.savefig(name, dpi=72)
|
||||
plt.close()
|
||||
|
||||
|
||||
class FiniteDifferenceStep(MessagePassing):
|
||||
@@ -9,50 +43,69 @@ class FiniteDifferenceStep(MessagePassing):
|
||||
TODO: add docstring.
|
||||
"""
|
||||
|
||||
def __init__(self, aggr: str = "add", root_weight: float = 1.0):
|
||||
def __init__(self, edge_ch=5, hidden_dim=16, aggr: str = "add"):
|
||||
super().__init__(aggr=aggr)
|
||||
assert (
|
||||
aggr == "add"
|
||||
), "Per somme pesate, l'aggregazione deve essere 'add'."
|
||||
|
||||
self.correction_net = nn.Sequential(
|
||||
nn.Linear(2, 6),
|
||||
nn.Tanh(),
|
||||
nn.Linear(6, 1),
|
||||
nn.Tanh(),
|
||||
self.x_embedding = nn.Sequential(
|
||||
spectral_norm(nn.Linear(1, hidden_dim // 2)),
|
||||
nn.GELU(),
|
||||
spectral_norm(nn.Linear(hidden_dim // 2, hidden_dim)),
|
||||
)
|
||||
|
||||
self.edge_embedding = nn.Sequential(
|
||||
spectral_norm(nn.Linear(edge_ch, hidden_dim // 2)),
|
||||
nn.GELU(),
|
||||
spectral_norm(nn.Linear(hidden_dim // 2, hidden_dim)),
|
||||
)
|
||||
|
||||
self.update_net = nn.Sequential(
|
||||
spectral_norm(nn.Linear(1, 6)),
|
||||
nn.Softplus(),
|
||||
spectral_norm(nn.Linear(6, 1)),
|
||||
nn.Softplus(),
|
||||
spectral_norm(nn.Linear(2 * hidden_dim, hidden_dim)),
|
||||
nn.GELU(),
|
||||
spectral_norm(nn.Linear(hidden_dim, hidden_dim)),
|
||||
nn.GELU(),
|
||||
# spectral_norm(nn.Linear(hidden_dim // 2, 1)),
|
||||
)
|
||||
|
||||
self.message_net = nn.Sequential(
|
||||
spectral_norm(nn.Linear(1, 6)),
|
||||
nn.Softplus(),
|
||||
spectral_norm(nn.Linear(6, 1)),
|
||||
nn.Softplus(),
|
||||
# self.message_net = nn.Sequential(
|
||||
# spectral_norm(nn.Linear(2 * hidden_dim, hidden_dim)),
|
||||
# nn.GELU(),
|
||||
# spectral_norm(nn.Linear(hidden_dim, hidden_dim // 2)),
|
||||
# nn.GELU(),
|
||||
# spectral_norm(nn.Linear(hidden_dim // 2, hidden_dim)),
|
||||
# )
|
||||
|
||||
self.out_net = nn.Sequential(
|
||||
spectral_norm(nn.Linear(hidden_dim, hidden_dim // 2)),
|
||||
nn.GELU(),
|
||||
spectral_norm(nn.Linear(hidden_dim // 2, 1)),
|
||||
)
|
||||
self.p = torch.nn.Parameter(torch.tensor(0.5))
|
||||
# self.a = torch.nn.Parameter(torch.tensor(root_weight))
|
||||
|
||||
def forward(self, x, edge_index, edge_attr, deg):
|
||||
"""
|
||||
TODO: add docstring.
|
||||
"""
|
||||
out = self.propagate(edge_index, x=x, edge_attr=edge_attr, deg=deg)
|
||||
return out
|
||||
x_ = self.x_embedding(x)
|
||||
edge_attr_ = self.edge_embedding(edge_attr)
|
||||
out = self.propagate(edge_index, x=x_, edge_attr=edge_attr_, deg=deg)
|
||||
return self.out_net(x_ + out)
|
||||
|
||||
def message(self, x_j, edge_attr):
|
||||
"""
|
||||
TODO: add docstring.
|
||||
"""
|
||||
# x_in = torch.cat([x_j, edge_attr.view(-1, 1)], dim=-1)
|
||||
# correction = self.correction_net(x_in)
|
||||
# p = torch.sigmoid(self.p)
|
||||
# return (p * edge_attr.view(-1, 1) + (1 - p) * correction) * x_j
|
||||
return edge_attr.view(-1, 1) * x_j
|
||||
# msg_input = torch.cat([x_j, edge_attr], dim=-1)
|
||||
# return self.message_net(msg_input) * edge_attr[:, 3].view(-1, 1)
|
||||
return x_j * edge_attr
|
||||
|
||||
def update(self, aggr_out, x):
|
||||
"""
|
||||
TODO: add docstring.
|
||||
"""
|
||||
update_input = torch.cat([x, aggr_out], dim=-1)
|
||||
return self.update_net(update_input)
|
||||
# return self.update_net(aggr_out)
|
||||
# return aggr_out
|
||||
# h = self.update_net(aggr_out, x)
|
||||
# return h
|
||||
|
||||
def aggregate(self, inputs, index, deg):
|
||||
"""
|
||||
@@ -62,68 +115,12 @@ class FiniteDifferenceStep(MessagePassing):
|
||||
deg = deg + 1e-7
|
||||
return out / deg.view(-1, 1)
|
||||
|
||||
def update(self, aggr_out, x):
|
||||
"""
|
||||
TODO: add docstring.
|
||||
"""
|
||||
return self.update_net(aggr_out)
|
||||
|
||||
# # Da fare:
|
||||
# # - Finire calcolo della loss su ogni step e poi media
|
||||
# # - Test con vari modelli
|
||||
# # - Se non dovesse funzionare, provare ad adeguare il criterio di uscita
|
||||
|
||||
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))
|
||||
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)
|
||||
with torch.no_grad():
|
||||
residual_norm = torch.norm(out - x)
|
||||
if residual_norm < conv_thres:
|
||||
break
|
||||
x = out.detach()
|
||||
return out, _i + 1
|
||||
# # PINN batching:
|
||||
# # - Provare singola condizione
|
||||
# # - Ottimizzatore del secondo ordine (LBFGS)
|
||||
|
||||
Reference in New Issue
Block a user