Implement S2 R2-G24 and Transolver query models with synthetic acceptance

This commit is contained in:
Frank14f committed 2026-11-14 22:45:34 +08:00
1 parent d4e26d48d0
commit bb09e8e164
17 files changed
+1734

No files matched your search

+21
View File
@@ -0,0 +1,21 @@
MIT License
Copyright (c) 2024 THUML @ Tsinghua University
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
@@ -0,0 +1,175 @@
import torch.nn as nn
import torch
from einops import rearrange, repeat
class Physics_Attention_Irregular_Mesh(nn.Module):
## for irregular meshes in 1D, 2D or 3D space
def __init__(self, dim, heads=8, dim_head=64, dropout=0., slice_num=64):
super().__init__()
inner_dim = dim_head * heads
self.dim_head = dim_head
self.heads = heads
self.scale = dim_head ** -0.5
self.softmax = nn.Softmax(dim=-1)
self.dropout = nn.Dropout(dropout)
self.temperature = nn.Parameter(torch.ones([1, heads, 1, 1]) * 0.5)
self.in_project_x = nn.Linear(dim, inner_dim)
self.in_project_fx = nn.Linear(dim, inner_dim)
self.in_project_slice = nn.Linear(dim_head, slice_num)
for l in [self.in_project_slice]:
torch.nn.init.orthogonal_(l.weight) # use a principled initialization
self.to_q = nn.Linear(dim_head, dim_head, bias=False)
self.to_k = nn.Linear(dim_head, dim_head, bias=False)
self.to_v = nn.Linear(dim_head, dim_head, bias=False)
self.to_out = nn.Sequential(
nn.Linear(inner_dim, dim),
nn.Dropout(dropout)
)
def forward(self, x):
# B N C
B, N, C = x.shape
### (1) Slice
fx_mid = self.in_project_fx(x).reshape(B, N, self.heads, self.dim_head) \
.permute(0, 2, 1, 3).contiguous() # B H N C
x_mid = self.in_project_x(x).reshape(B, N, self.heads, self.dim_head) \
.permute(0, 2, 1, 3).contiguous() # B H N C
slice_weights = self.softmax(self.in_project_slice(x_mid) / self.temperature) # B H N G
slice_norm = slice_weights.sum(2) # B H G
slice_token = torch.einsum("bhnc,bhng->bhgc", fx_mid, slice_weights)
slice_token = slice_token / ((slice_norm + 1e-5)[:, :, :, None].repeat(1, 1, 1, self.dim_head))
### (2) Attention among slice tokens
q_slice_token = self.to_q(slice_token)
k_slice_token = self.to_k(slice_token)
v_slice_token = self.to_v(slice_token)
dots = torch.matmul(q_slice_token, k_slice_token.transpose(-1, -2)) * self.scale
attn = self.softmax(dots)
attn = self.dropout(attn)
out_slice_token = torch.matmul(attn, v_slice_token) # B H G D
### (3) Deslice
out_x = torch.einsum("bhgc,bhng->bhnc", out_slice_token, slice_weights)
out_x = rearrange(out_x, 'b h n d -> b n (h d)')
return self.to_out(out_x)
class Physics_Attention_Structured_Mesh_2D(nn.Module):
## for structured mesh in 2D space
def __init__(self, dim, heads=8, dim_head=64, dropout=0., slice_num=64, H=101, W=31, kernel=3): # kernel=3):
super().__init__()
inner_dim = dim_head * heads
self.dim_head = dim_head
self.heads = heads
self.scale = dim_head ** -0.5
self.softmax = nn.Softmax(dim=-1)
self.dropout = nn.Dropout(dropout)
self.temperature = nn.Parameter(torch.ones([1, heads, 1, 1]) * 0.5)
self.H = H
self.W = W
self.in_project_x = nn.Conv2d(dim, inner_dim, kernel, 1, kernel // 2)
self.in_project_fx = nn.Conv2d(dim, inner_dim, kernel, 1, kernel // 2)
self.in_project_slice = nn.Linear(dim_head, slice_num)
for l in [self.in_project_slice]:
torch.nn.init.orthogonal_(l.weight) # use a principled initialization
self.to_q = nn.Linear(dim_head, dim_head, bias=False)
self.to_k = nn.Linear(dim_head, dim_head, bias=False)
self.to_v = nn.Linear(dim_head, dim_head, bias=False)
self.to_out = nn.Sequential(
nn.Linear(inner_dim, dim),
nn.Dropout(dropout)
)
def forward(self, x):
# B N C
B, N, C = x.shape
x = x.reshape(B, self.H, self.W, C).contiguous().permute(0, 3, 1, 2).contiguous() # B C H W
### (1) Slice
fx_mid = self.in_project_fx(x).permute(0, 2, 3, 1).contiguous().reshape(B, N, self.heads, self.dim_head) \
.permute(0, 2, 1, 3).contiguous() # B H N C
x_mid = self.in_project_x(x).permute(0, 2, 3, 1).contiguous().reshape(B, N, self.heads, self.dim_head) \
.permute(0, 2, 1, 3).contiguous() # B H N G
slice_weights = self.softmax(
self.in_project_slice(x_mid) / torch.clamp(self.temperature, min=0.1, max=5)) # B H N G
slice_norm = slice_weights.sum(2) # B H G
slice_token = torch.einsum("bhnc,bhng->bhgc", fx_mid, slice_weights)
slice_token = slice_token / ((slice_norm + 1e-5)[:, :, :, None].repeat(1, 1, 1, self.dim_head))
### (2) Attention among slice tokens
q_slice_token = self.to_q(slice_token)
k_slice_token = self.to_k(slice_token)
v_slice_token = self.to_v(slice_token)
dots = torch.matmul(q_slice_token, k_slice_token.transpose(-1, -2)) * self.scale
attn = self.softmax(dots)
attn = self.dropout(attn)
out_slice_token = torch.matmul(attn, v_slice_token) # B H G D
### (3) Deslice
out_x = torch.einsum("bhgc,bhng->bhnc", out_slice_token, slice_weights)
out_x = rearrange(out_x, 'b h n d -> b n (h d)')
return self.to_out(out_x)
class Physics_Attention_Structured_Mesh_3D(nn.Module):
## for structured mesh in 3D space
def __init__(self, dim, heads=8, dim_head=64, dropout=0., slice_num=32, H=32, W=32, D=32, kernel=3):
super().__init__()
inner_dim = dim_head * heads
self.dim_head = dim_head
self.heads = heads
self.scale = dim_head ** -0.5
self.softmax = nn.Softmax(dim=-1)
self.dropout = nn.Dropout(dropout)
self.temperature = nn.Parameter(torch.ones([1, heads, 1, 1]) * 0.5)
self.H = H
self.W = W
self.D = D
self.in_project_x = nn.Conv3d(dim, inner_dim, kernel, 1, kernel // 2)
self.in_project_fx = nn.Conv3d(dim, inner_dim, kernel, 1, kernel // 2)
self.in_project_slice = nn.Linear(dim_head, slice_num)
for l in [self.in_project_slice]:
torch.nn.init.orthogonal_(l.weight) # use a principled initialization
self.to_q = nn.Linear(dim_head, dim_head, bias=False)
self.to_k = nn.Linear(dim_head, dim_head, bias=False)
self.to_v = nn.Linear(dim_head, dim_head, bias=False)
self.to_out = nn.Sequential(
nn.Linear(inner_dim, dim),
nn.Dropout(dropout)
)
def forward(self, x):
# B N C
B, N, C = x.shape
x = x.reshape(B, self.H, self.W, self.D, C).contiguous().permute(0, 4, 1, 2, 3).contiguous() # B C H W
### (1) Slice
fx_mid = self.in_project_fx(x).permute(0, 2, 3, 4, 1).contiguous().reshape(B, N, self.heads, self.dim_head) \
.permute(0, 2, 1, 3).contiguous() # B H N C
x_mid = self.in_project_x(x).permute(0, 2, 3, 4, 1).contiguous().reshape(B, N, self.heads, self.dim_head) \
.permute(0, 2, 1, 3).contiguous() # B H N G
slice_weights = self.softmax(
self.in_project_slice(x_mid) / torch.clamp(self.temperature, min=0.1, max=5)) # B H N G
slice_norm = slice_weights.sum(2) # B H G
slice_token = torch.einsum("bhnc,bhng->bhgc", fx_mid, slice_weights)
slice_token = slice_token / ((slice_norm + 1e-5)[:, :, :, None].repeat(1, 1, 1, self.dim_head))
### (2) Attention among slice tokens
q_slice_token = self.to_q(slice_token)
k_slice_token = self.to_k(slice_token)
v_slice_token = self.to_v(slice_token)
dots = torch.matmul(q_slice_token, k_slice_token.transpose(-1, -2)) * self.scale
attn = self.softmax(dots)
attn = self.dropout(attn)
out_slice_token = torch.matmul(attn, v_slice_token) # B H G D
### (3) Deslice
out_x = torch.einsum("bhgc,bhng->bhnc", out_slice_token, slice_weights)
out_x = rearrange(out_x, 'b h n d -> b n (h d)')
return self.to_out(out_x)
@@ -0,0 +1,158 @@
import torch
import torch.nn as nn
from timm.models.layers import trunc_normal_
from model.Embedding import timestep_embedding
import numpy as np
from model.Physics_Attention import Physics_Attention_Irregular_Mesh
ACTIVATION = {'gelu': nn.GELU, 'tanh': nn.Tanh, 'sigmoid': nn.Sigmoid, 'relu': nn.ReLU, 'leaky_relu': nn.LeakyReLU(0.1),
'softplus': nn.Softplus, 'ELU': nn.ELU, 'silu': nn.SiLU}
class MLP(nn.Module):
def __init__(self, n_input, n_hidden, n_output, n_layers=1, act='gelu', res=True):
super(MLP, self).__init__()
if act in ACTIVATION.keys():
act = ACTIVATION[act]
else:
raise NotImplementedError
self.n_input = n_input
self.n_hidden = n_hidden
self.n_output = n_output
self.n_layers = n_layers
self.res = res
self.linear_pre = nn.Sequential(nn.Linear(n_input, n_hidden), act())
self.linear_post = nn.Linear(n_hidden, n_output)
self.linears = nn.ModuleList([nn.Sequential(nn.Linear(n_hidden, n_hidden), act()) for _ in range(n_layers)])
def forward(self, x):
x = self.linear_pre(x)
for i in range(self.n_layers):
if self.res:
x = self.linears[i](x) + x
else:
x = self.linears[i](x)
x = self.linear_post(x)
return x
class Transolver_block(nn.Module):
"""Transformer encoder block."""
def __init__(
self,
num_heads: int,
hidden_dim: int,
dropout: float,
act='gelu',
mlp_ratio=4,
last_layer=False,
out_dim=1,
slice_num=32,
):
super().__init__()
self.last_layer = last_layer
self.ln_1 = nn.LayerNorm(hidden_dim)
self.Attn = Physics_Attention_Irregular_Mesh(hidden_dim, heads=num_heads, dim_head=hidden_dim // num_heads,
dropout=dropout, slice_num=slice_num)
self.ln_2 = nn.LayerNorm(hidden_dim)
self.mlp = MLP(hidden_dim, hidden_dim * mlp_ratio, hidden_dim, n_layers=0, res=False, act=act)
if self.last_layer:
self.ln_3 = nn.LayerNorm(hidden_dim)
self.mlp2 = nn.Linear(hidden_dim, out_dim)
def forward(self, fx):
fx = self.Attn(self.ln_1(fx)) + fx
fx = self.mlp(self.ln_2(fx)) + fx
if self.last_layer:
return self.mlp2(self.ln_3(fx))
else:
return fx
class Model(nn.Module):
def __init__(self,
space_dim=1,
n_layers=5,
n_hidden=256,
dropout=0.0,
n_head=8,
Time_Input=False,
act='gelu',
mlp_ratio=1,
fun_dim=1,
out_dim=1,
slice_num=32,
ref=8,
unified_pos=False
):
super(Model, self).__init__()
self.__name__ = 'Transolver_1D'
self.ref = ref
self.unified_pos = unified_pos
self.Time_Input = Time_Input
self.n_hidden = n_hidden
self.space_dim = space_dim
if self.unified_pos:
self.preprocess = MLP(fun_dim + self.ref * self.ref, n_hidden * 2, n_hidden, n_layers=0, res=False, act=act)
else:
self.preprocess = MLP(fun_dim + space_dim, n_hidden * 2, n_hidden, n_layers=0, res=False, act=act)
if Time_Input:
self.time_fc = nn.Sequential(nn.Linear(n_hidden, n_hidden), nn.SiLU(), nn.Linear(n_hidden, n_hidden))
self.blocks = nn.ModuleList([Transolver_block(num_heads=n_head, hidden_dim=n_hidden,
dropout=dropout,
act=act,
mlp_ratio=mlp_ratio,
out_dim=out_dim,
slice_num=slice_num,
last_layer=(_ == n_layers - 1))
for _ in range(n_layers)])
self.initialize_weights()
self.placeholder = nn.Parameter((1 / (n_hidden)) * torch.rand(n_hidden, dtype=torch.float))
def initialize_weights(self):
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=0.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, (nn.LayerNorm, nn.BatchNorm1d)):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def get_grid(self, x, batchsize=1):
# x: B N 2
# grid_ref
gridx = torch.tensor(np.linspace(0, 1, self.ref), dtype=torch.float)
gridx = gridx.reshape(1, self.ref, 1, 1).repeat([batchsize, 1, self.ref, 1])
gridy = torch.tensor(np.linspace(0, 1, self.ref), dtype=torch.float)
gridy = gridy.reshape(1, 1, self.ref, 1).repeat([batchsize, self.ref, 1, 1])
grid_ref = torch.cat((gridx, gridy), dim=-1).cuda().reshape(batchsize, self.ref * self.ref, 2) # B H W 8 8 2
pos = torch.sqrt(torch.sum((x[:, :, None, :] - grid_ref[:, None, :, :]) ** 2, dim=-1)). \
reshape(batchsize, x.shape[1], self.ref * self.ref).contiguous()
return pos
def forward(self, x, fx, T=None):
if self.unified_pos:
x = self.get_grid(x, x.shape[0])
if fx is not None:
fx = torch.cat((x, fx), -1)
fx = self.preprocess(fx)
else:
fx = self.preprocess(x)
fx = fx + self.placeholder[None, None, :]
if T is not None:
Time_emb = timestep_embedding(T, self.n_hidden).repeat(1, x.shape[1], 1)
Time_emb = self.time_fc(Time_emb)
fx = fx + Time_emb
for block in self.blocks:
fx = block(fx)
return fx
+9
View File
@@ -0,0 +1,9 @@
{
"revision": "75e0f67643806a81cd1d3f6adc88dd8c02416fe7",
"repository": "https://github.com/thuml/Transolver",
"files": {
"LICENSE": "2c919cd03fa823bf7eefc00a957ff8324c865cd22aa5285e563dc4b558084f25",
"PDE-Solving-StandardBenchmark/model/Transolver_Irregular_Mesh.py": "5259759f7ecc6a954c8442ac8359990b9f242f1437c452c69e5e0f03549b5311",
"PDE-Solving-StandardBenchmark/model/Physics_Attention.py": "4fc710f6f96788fb59cb19fab8d64873d3d16e9e182f12a544d451c74c548cc7"
}
}