File size: 7,309 Bytes
78b3cac 8c463a9 78b3cac f00cec9 78b3cac f00cec9 78b3cac |
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 |
import torch
import torch.nn as nn
from .layers import Attention, MLP
from .conditions import TimestepEmbedder, ConditionEmbedder
# from .diffusion_utils import PlaceHolder
#### graph utils
class PlaceHolder:
def __init__(self, X, E, y):
self.X = X
self.E = E
self.y = y
def type_as(self, x: torch.Tensor, categorical: bool = False):
"""Changes the device and dtype of X, E, y."""
self.X = self.X.type_as(x)
self.E = self.E.type_as(x)
if categorical:
self.y = self.y.type_as(x)
return self
def mask(self, node_mask, collapse=False):
x_mask = node_mask.unsqueeze(-1) # bs, n, 1
e_mask1 = x_mask.unsqueeze(2) # bs, n, 1, 1
e_mask2 = x_mask.unsqueeze(1) # bs, 1, n, 1
if collapse:
self.X = torch.argmax(self.X, dim=-1)
self.E = torch.argmax(self.E, dim=-1)
self.X[node_mask == 0] = -1
self.E[(e_mask1 * e_mask2).squeeze(-1) == 0] = -1
else:
self.X = self.X * x_mask
self.E = self.E * e_mask1 * e_mask2
assert torch.allclose(self.E, torch.transpose(self.E, 1, 2))
return self
def modulate(x, shift, scale):
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
class Transformer(nn.Module):
def __init__(
self,
max_n_nodes,
hidden_size=384,
depth=12,
num_heads=16,
mlp_ratio=4.0,
drop_condition=0.1,
Xdim=118,
Edim=5,
ydim=5,
):
super().__init__()
self.num_heads = num_heads
self.ydim = ydim
self.x_embedder = nn.Sequential(
nn.Linear(Xdim + max_n_nodes * Edim, hidden_size, bias=False),
nn.LayerNorm(hidden_size)
)
self.t_embedder = TimestepEmbedder(hidden_size)
self.y_embedder = ConditionEmbedder(ydim, hidden_size, drop_condition)
self.blocks = nn.ModuleList(
[
Block(hidden_size, num_heads, mlp_ratio=mlp_ratio)
for _ in range(depth)
]
)
self.output_layer = OutputLayer(
max_n_nodes=max_n_nodes,
hidden_size=hidden_size,
atom_type=Xdim,
bond_type=Edim,
mlp_ratio=mlp_ratio,
num_heads=num_heads,
)
self.initialize_weights()
def initialize_weights(self):
# Initialize transformer layers:
def _basic_init(module):
if isinstance(module, nn.Linear):
torch.nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.constant_(module.bias, 0)
def _constant_init(module, i):
if isinstance(module, nn.Linear):
nn.init.constant_(module.weight, i)
if module.bias is not None:
nn.init.constant_(module.bias, i)
self.apply(_basic_init)
for block in self.blocks:
_constant_init(block.adaLN_modulation[0], 0)
_constant_init(self.output_layer.adaLN_modulation[0], 0)
def disable_grads(self):
"""
Disable gradients for all parameters in the model.
"""
for param in self.parameters():
param.requires_grad = False
def print_trainable_parameters(self):
print("Trainable parameters:")
for name, param in self.named_parameters():
if param.requires_grad:
print(f"{name}: {param.size()}")
# Calculate and print the total number of trainable parameters
total_params = sum(p.numel() for p in self.parameters() if p.requires_grad)
print(f"\nTotal trainable parameters: {total_params}")
def forward(self, X_in, E_in, node_mask, y_in, t, unconditioned):
bs, n, _ = X_in.size()
X = torch.cat([X_in, E_in.reshape(bs, n, -1)], dim=-1)
X = self.x_embedder(X)
c1 = self.t_embedder(t)
c2 = self.y_embedder(y_in, self.training, unconditioned)
c = c1 + c2
for i, block in enumerate(self.blocks):
X = block(X, c, node_mask)
# X: B * N * dx, E: B * N * N * de
X, E = self.output_layer(X, X_in, E_in, c, t, node_mask)
return PlaceHolder(X=X, E=E, y=None).mask(node_mask)
class Block(nn.Module):
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, **block_kwargs):
super().__init__()
self.attn_norm = nn.LayerNorm(hidden_size, eps=1e-05, elementwise_affine=False)
self.mlp_norm = nn.LayerNorm(hidden_size, eps=1e-05, elementwise_affine=False)
self.attn = Attention(
hidden_size, num_heads=num_heads, qkv_bias=False, qk_norm=True, **block_kwargs
)
self.mlp = MLP(
in_features=hidden_size,
hidden_features=int(hidden_size * mlp_ratio),
)
self.adaLN_modulation = nn.Sequential(
nn.Linear(hidden_size, hidden_size, bias=True),
nn.SiLU(),
nn.Linear(hidden_size, 6 * hidden_size, bias=True),
nn.Softsign()
)
def forward(self, x, c, node_mask):
(
shift_msa,
scale_msa,
gate_msa,
shift_mlp,
scale_mlp,
gate_mlp,
) = self.adaLN_modulation(c).chunk(6, dim=1)
x = x + gate_msa.unsqueeze(1) * modulate(self.attn_norm(self.attn(x, node_mask=node_mask)), shift_msa, scale_msa)
x = x + gate_mlp.unsqueeze(1) * modulate(self.mlp_norm(self.mlp(x)), shift_mlp, scale_mlp)
return x
class OutputLayer(nn.Module):
def __init__(self, max_n_nodes, hidden_size, atom_type, bond_type, mlp_ratio, num_heads=None):
super().__init__()
self.atom_type = atom_type
self.bond_type = bond_type
final_size = atom_type + max_n_nodes * bond_type
self.xedecoder = MLP(in_features=hidden_size,
out_features=final_size, drop=0)
self.norm_final = nn.LayerNorm(final_size, eps=1e-05, elementwise_affine=False)
self.adaLN_modulation = nn.Sequential(
nn.Linear(hidden_size, hidden_size, bias=True),
nn.SiLU(),
nn.Linear(hidden_size, 2 * final_size, bias=True)
)
def forward(self, x, x_in, e_in, c, t, node_mask):
x_all = self.xedecoder(x)
B, N, D = x_all.size()
shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
x_all = modulate(self.norm_final(x_all), shift, scale)
atom_out = x_all[:, :, :self.atom_type]
atom_out = x_in + atom_out
bond_out = x_all[:, :, self.atom_type:].reshape(B, N, N, self.bond_type)
bond_out = e_in + bond_out
##### standardize adj_out
edge_mask = (~node_mask)[:, :, None] & (~node_mask)[:, None, :]
diag_mask = (
torch.eye(N, dtype=torch.bool)
.unsqueeze(0)
.expand(B, -1, -1)
.type_as(edge_mask)
)
bond_out.masked_fill_(edge_mask[:, :, :, None], 0)
bond_out.masked_fill_(diag_mask[:, :, :, None], 0)
bond_out = 1 / 2 * (bond_out + torch.transpose(bond_out, 1, 2))
return atom_out, bond_out |