File size: 4,601 Bytes
ff0fadf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 | import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from timm.layers import trunc_normal_
from onescience.modules.attention.flashattention import FlashAttention
from onescience.modules.mlp.MLP import StandardMLP
from onescience.modules.embedding import timestep_embedding, unified_pos_embedding
from einops import rearrange, repeat
class Transformer_block(nn.Module):
"""
Transformer encoder block.
包含多头自注意力和前馈神经网络，最后一层可选是否直接输出预测维度。
"""
def __init__(
self,
num_heads: int,
hidden_dim: int,
dropout: float,
act="gelu",
mlp_ratio=4,
last_layer=False,
out_dim=1,
):
super().__init__()
self.last_layer = last_layer
self.ln_1 = nn.LayerNorm(hidden_dim)
self.Attn = FlashAttention(
dim=hidden_dim,
heads=num_heads,
dim_head=hidden_dim // num_heads,
dropout=dropout,
)
self.ln_2 = nn.LayerNorm(hidden_dim)
self.mlp = StandardMLP(
input_dim=hidden_dim,
output_dim=hidden_dim,
hidden_dims=[hidden_dim * mlp_ratio],
activation=act,
use_bias=True
)
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):
"""
标准的 Transformer 模型架构。
通过堆叠 Transformer_block 处理物理场数据。
"""
def __init__(self, args, device):
super(Model, self).__init__()
self.__name__ = "Transformer"
self.args = args
## embedding
if (
args.unified_pos and args.geotype != "unstructured"
): # only for structured mesh
self.pos = unified_pos_embedding(args.shapelist, args.ref, device=device)
input_dim = args.fun_dim + args.ref ** len(args.shapelist)
else:
input_dim = args.fun_dim + args.space_dim
self.preprocess = StandardMLP(
input_dim=input_dim,
output_dim=args.n_hidden,
hidden_dims=[args.n_hidden * 2],
activation=args.act,
use_bias=True
)
if args.time_input:
self.time_fc = nn.Sequential(
nn.Linear(args.n_hidden, args.n_hidden),
nn.SiLU(),
nn.Linear(args.n_hidden, args.n_hidden),
)
self.blocks = nn.ModuleList(
[
Transformer_block(
num_heads=args.n_heads,
hidden_dim=args.n_hidden,
dropout=args.dropout,
act=args.act,
mlp_ratio=args.mlp_ratio,
out_dim=args.out_dim,
last_layer=(_ == args.n_layers - 1),
)
for _ in range(args.n_layers)
]
)
self.placeholder = nn.Parameter(
(1 / (args.n_hidden)) * torch.rand(args.n_hidden, dtype=torch.float)
)
self.initialize_weights()
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 forward(self, x, fx, T=None, geo=None):
if self.args.unified_pos:
x = self.pos.repeat(x.shape[0], 1, 1)
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.args.n_hidden)
Time_emb = self.time_fc(Time_emb)
if Time_emb.ndim == 2:
Time_emb = Time_emb.unsqueeze(1)
fx = fx + Time_emb
for block in self.blocks:
fx = block(fx)
return fx
|