File size: 11,773 Bytes
e611d1f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_scatter import scatter
from torch_geometric.nn import InstanceNorm
class EGNNLayer(nn.Module):
"""
EGNN layer with optional feed forward network and batch normalization.
Args:
input_nf: Number of input node features
output_nf: Number of output node features
hidden_nf: Number of hidden features
edges_in_d: Number of input edge features
act_fn: Activation function
residual: Whether to use residual connections
attention: Whether to use attention mechanism for edge features
normalize: Whether to normalize coordinates
coords_agg: Aggregation method for coordinates (mean, sum, max, min)
tanh: Whether to use tanh activation for coordinate updates
dropout: Dropout rate
ffn: Whether to use feed forward network
batch_norm: Whether to use batch normalization
"""
def __init__(self, input_nf, output_nf, hidden_nf,
edges_in_d=0, act_fn=nn.SiLU(),
residual=True, attention=False, normalize=False,
coords_agg='mean', tanh=False, dropout=0.0,
ffn=False, batch_norm=True):
super().__init__()
self.input_nf = input_nf
self.output_nf = output_nf
self.hidden_nf = hidden_nf
self.residual = residual
self.attention = attention
self.normalize = normalize
self.coords_agg = coords_agg
self.tanh = tanh
self.epsilon = 1e-8
self.dropout = dropout
self.ffn = ffn
self.batch_norm = batch_norm
# Edge MLP
in_edge = input_nf*2 + 1 + edges_in_d
self.edge_mlp = nn.Sequential(
nn.Linear(in_edge, hidden_nf),
act_fn, nn.Dropout(dropout),
nn.Linear(hidden_nf, hidden_nf),
act_fn, nn.Dropout(dropout),
)
if attention:
self.att_mlp = nn.Sequential(nn.Linear(hidden_nf,1), nn.Sigmoid())
# Coord MLP
layer = nn.Linear(hidden_nf,1, bias=False)
nn.init.xavier_uniform_(layer.weight, gain=0.001)
coord_blocks = [nn.Linear(hidden_nf, hidden_nf), act_fn,
nn.Dropout(dropout), layer]
if tanh: coord_blocks.append(nn.Tanh())
self.coord_mlp = nn.Sequential(*coord_blocks)
# Node MLP
self.node_mlp = nn.Sequential(
nn.Linear(hidden_nf + input_nf, hidden_nf),
act_fn, nn.Dropout(dropout),
nn.Linear(hidden_nf, output_nf),
)
# per-graph normalization
if batch_norm:
self.norm_node = InstanceNorm(output_nf, affine=True)
self.norm_coord = InstanceNorm(3, affine=True)
# FFN
if ffn:
self.ff1 = nn.Linear(output_nf, output_nf*2)
self.ff2 = nn.Linear(output_nf*2, output_nf)
self.act_ff = act_fn
self.drop_ff = nn.Dropout(dropout)
if batch_norm:
self.norm_ff1 = InstanceNorm(output_nf, affine=True)
self.norm_ff2 = InstanceNorm(output_nf, affine=True)
def coord2radial(self, edge_index, coord):
row, col = edge_index
diff = coord[row] - coord[col]
dist2 = (diff**2).sum(dim=-1, keepdim=True)
# Clamp distance to prevent extreme values
dist2 = torch.clamp(dist2, min=self.epsilon, max=100.0)
if self.normalize:
norm = (dist2.sqrt().detach() + self.epsilon)
diff = diff / norm
# Check for NaN/Inf in normalized diff
diff = torch.where(torch.isfinite(diff), diff, torch.zeros_like(diff))
return dist2, diff
def _ff_block(self, x):
"""Feed Forward block.
"""
x = self.drop_ff(self.act_ff(self.ff1(x)))
return self.ff2(x)
def forward(self, h, coord, edge_index, batch, edge_attr=None, node_attr=None):
row, col = edge_index
radial, coord_diff = self.coord2radial(edge_index, coord)
# -- edge features --
e_in = [h[row], h[col], radial]
if edge_attr is not None: e_in.append(edge_attr)
e = torch.cat(e_in, dim=-1)
e = self.edge_mlp(e)
if self.attention:
att = self.att_mlp(e)
e = e * att
# -- coordinate update --
coord_update = self.coord_mlp(e) # [E,1]
# Clamp coordinate updates to prevent explosion
coord_update = torch.clamp(coord_update, -1.0, 1.0)
trans = coord_diff * coord_update # [E,3]
# Check for NaN/Inf in coordinate updates
trans = torch.where(torch.isfinite(trans), trans, torch.zeros_like(trans))
agg_coord = scatter(trans, row, dim=0,
dim_size=coord.size(0),
reduce=self.coords_agg)
coord = coord + agg_coord
# Check for NaN/Inf in final coordinates
coord = torch.where(torch.isfinite(coord), coord, torch.zeros_like(coord))
if self.batch_norm:
coord = self.norm_coord(coord, batch)
# -- node update --
agg_node = scatter(e, row, dim=0,
dim_size=h.size(0), reduce='sum')
x_in = torch.cat([h, agg_node], dim=-1)
if node_attr is not None:
x_in = torch.cat([x_in, node_attr], dim=-1)
h_new = self.node_mlp(x_in)
if self.batch_norm:
h_new = self.norm_node(h_new, batch)
if self.residual and h_new.shape[-1] == h.shape[-1]:
h_new = h + h_new
# -- optional FFN --
if self.ffn:
if self.batch_norm:
h_new = self.norm_ff1(h_new, batch)
h_new = h_new + self._ff_block(h_new)
if self.batch_norm:
h_new = self.norm_ff2(h_new, batch)
return h_new, coord, e
class EGNNLayer2(nn.Module):
"""
EGNN layer with optional feed forward network and batch normalization.
Args:
input_nf: Number of input node features
output_nf: Number of output node features
hidden_nf: Number of hidden features
edges_in_d: Number of input edge features
act_fn: Activation function
residual: Whether to use residual connections
attention: Whether to use attention mechanism for edge features
normalize: Whether to normalize coordinates
coords_agg: Aggregation method for coordinates (mean, sum, max, min)
tanh: Whether to use tanh activation for coordinate updates
dropout: Dropout rate
ffn: Whether to use feed forward network
batch_norm: Whether to use batch normalization
"""
def __init__(self, input_nf, output_nf, hidden_nf,
edges_in_d=0, act_fn=nn.SiLU(),
residual=True, attention=False, normalize=False,
coords_agg='mean', tanh=False, dropout=0.0,
ffn=False, batch_norm=True):
super().__init__()
self.input_nf = input_nf
self.output_nf = output_nf
self.hidden_nf = hidden_nf
self.residual = residual
self.attention = attention
self.normalize = normalize
self.coords_agg = coords_agg
self.tanh = tanh
self.epsilon = 1e-8
self.dropout = dropout
self.ffn = ffn
self.batch_norm = batch_norm
# Edge MLP
in_edge = input_nf*2 + 1 + edges_in_d
self.edge_mlp = nn.Sequential(
nn.Linear(in_edge, hidden_nf),
act_fn, nn.Dropout(dropout),
nn.Linear(hidden_nf, hidden_nf),
act_fn, nn.Dropout(dropout),
)
if attention:
self.att_mlp = nn.Sequential(nn.Linear(hidden_nf,1), nn.Sigmoid())
# Coord MLP
layer = nn.Linear(hidden_nf,1, bias=False)
nn.init.xavier_uniform_(layer.weight, gain=0.001)
coord_blocks = [nn.Linear(hidden_nf, hidden_nf), act_fn,
nn.Dropout(dropout), layer]
if tanh: coord_blocks.append(nn.Tanh())
self.coord_mlp = nn.Sequential(*coord_blocks)
# Node MLP
self.node_mlp = nn.Sequential(
nn.Linear(hidden_nf + input_nf, hidden_nf),
act_fn, nn.Dropout(dropout),
nn.Linear(hidden_nf, output_nf),
)
# per-graph normalization
if batch_norm:
self.norm_node = InstanceNorm(output_nf, affine=True)
self.norm_coord = InstanceNorm(3, affine=True)
# FFN
if ffn:
self.ff1 = nn.Linear(output_nf, output_nf*2)
self.ff2 = nn.Linear(output_nf*2, output_nf)
self.act_ff = act_fn
self.drop_ff = nn.Dropout(dropout)
if batch_norm:
self.norm_ff1 = InstanceNorm(output_nf, affine=True)
self.norm_ff2 = InstanceNorm(output_nf, affine=True)
def coord2radial(self, edge_index, coord):
row, col = edge_index
diff = coord[row] - coord[col]
dist2 = (diff**2).sum(dim=-1, keepdim=True)
# Clamp distance to prevent extreme values
dist2 = torch.clamp(dist2, min=self.epsilon, max=100.0)
if self.normalize:
norm = (dist2.sqrt().detach() + self.epsilon)
diff = diff / norm
# Check for NaN/Inf in normalized diff
diff = torch.where(torch.isfinite(diff), diff, torch.zeros_like(diff))
return dist2, diff
def _ff_block(self, x):
"""Feed Forward block.
"""
x = self.drop_ff(self.act_ff(self.ff1(x)))
return self.ff2(x)
def forward(self, h, coord, edge_index, batch, edge_attr=None, node_attr=None):
row, col = edge_index
radial, coord_diff = self.coord2radial(edge_index, coord)
# -- edge features --
e_in = [h[row], h[col], radial]
if edge_attr is not None: e_in.append(edge_attr)
e = torch.cat(e_in, dim=-1)
e = self.edge_mlp(e)
if self.attention:
att = self.att_mlp(e)
e = e * att
# -- coordinate update --
coord_update = self.coord_mlp(e) # [E,1]
# Clamp coordinate updates to prevent explosion
coord_update = torch.clamp(coord_update, -1.0, 1.0)
trans = coord_diff * coord_update # [E,3]
# Check for NaN/Inf in coordinate updates
trans = torch.where(torch.isfinite(trans), trans, torch.zeros_like(trans))
agg_coord = scatter(trans, row, dim=0,
dim_size=coord.size(0),
reduce=self.coords_agg)
coord = coord + agg_coord
# Check for NaN/Inf in final coordinates
coord = torch.where(torch.isfinite(coord), coord, torch.zeros_like(coord))
if self.batch_norm:
coord = self.norm_coord(coord, batch)
# -- node update --
agg_node = scatter(e, row, dim=0,
dim_size=h.size(0), reduce='sum')
x_in = torch.cat([h, agg_node], dim=-1)
if node_attr is not None:
x_in = torch.cat([x_in, node_attr], dim=-1)
h_new = self.node_mlp(x_in)
if self.batch_norm:
h_new = self.norm_node(h_new, batch)
if self.residual and h_new.shape[-1] == h.shape[-1]:
h_new = h + h_new
# -- optional FFN --
if self.ffn:
if self.batch_norm:
h_new = self.norm_ff1(h_new, batch)
h_new = h_new + self._ff_block(h_new)
if self.batch_norm:
h_new = self.norm_ff2(h_new, batch)
return h_new, coord, e |