Spaces:
Sleeping
Sleeping
File size: 12,217 Bytes
2f54ec8 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 |
import hydra
import torch
import torch.nn.functional as F
from einops.layers.torch import Rearrange
from utils.pooling import HomogeneousAggregator
import torch.nn as nn
class RelationalTransformer(nn.Module):
def __init__(
self,
d_node,
d_edge,
d_attn_hid,
d_node_hid,
d_edge_hid,
d_out_hid,
d_out,
n_layers,
n_heads,
layer_layout,
graph_constructor,
dropout=0.0,
node_update_type="rt",
disable_edge_updates=False,
use_cls_token=False,
pooling_method="cat",
pooling_layer_idx="last",
rev_edge_features=False,
modulate_v=True,
use_ln=True,
tfixit_init=False,
):
super().__init__()
assert use_cls_token == (pooling_method == "cls_token")
self.pooling_method = pooling_method
self.pooling_layer_idx = pooling_layer_idx
self.rev_edge_features = rev_edge_features
self.nodes_per_layer = layer_layout
self.construct_graph = hydra.utils.instantiate(
graph_constructor,
d_node=d_node,
d_edge=d_edge,
layer_layout=layer_layout,
rev_edge_features=rev_edge_features,
)
self.use_cls_token = use_cls_token
if use_cls_token:
self.cls_token = nn.Parameter(torch.randn(d_node))
self.layers = nn.ModuleList(
[
torch.jit.script(
RTLayer(
d_node,
d_edge,
d_attn_hid,
d_node_hid,
d_edge_hid,
n_heads,
dropout,
node_update_type=node_update_type,
disable_edge_updates=(
(disable_edge_updates or (i == n_layers - 1))
and pooling_method != "mean_edge"
and pooling_layer_idx != "all"
),
modulate_v=modulate_v,
use_ln=use_ln,
tfixit_init=tfixit_init,
n_layers=n_layers,
)
)
for i in range(n_layers)
]
)
if pooling_method != "cls_token":
self.pool = HomogeneousAggregator(
pooling_method,
pooling_layer_idx,
layer_layout,
)
self.num_graph_features = (
layer_layout[-1] * d_node
if pooling_method == "cat" and pooling_layer_idx == "last"
else d_edge if pooling_method in ("mean_edge", "max_edge") else d_node
)
self.proj_out = nn.Sequential(
nn.Linear(self.num_graph_features, d_out_hid),
nn.ReLU(),
# nn.Linear(d_out_hid, d_out_hid),
# nn.ReLU(),
nn.Linear(d_out_hid, d_out),
)
self.final_features = (None,None,None,None)
def forward(self, inputs):
attn_weights = None
node_features, edge_features, mask = self.construct_graph(inputs)
if self.use_cls_token:
node_features = torch.cat(
[
# repeat(self.cls_token, "d -> b 1 d", b=node_features.size(0)),
self.cls_token.unsqueeze(0).expand(node_features.size(0), 1, -1),
node_features,
],
dim=1,
)
edge_features = F.pad(edge_features, (0, 0, 1, 0, 1, 0), value=0)
for layer in self.layers:
node_features, edge_features, attn_weights = layer(node_features, edge_features, mask)
if self.pooling_method == "cls_token":
graph_features = node_features[:, 0]
else:
graph_features = self.pool(node_features, edge_features)
self.final_features = (graph_features, node_features, edge_features, attn_weights)
return self.proj_out(graph_features)
class RTLayer(nn.Module):
def __init__(
self,
d_node,
d_edge,
d_attn_hid,
d_node_hid,
d_edge_hid,
n_heads,
dropout,
node_update_type="rt",
disable_edge_updates=False,
modulate_v=True,
use_ln=True,
tfixit_init=False,
n_layers=None,
):
super().__init__()
self.node_update_type = node_update_type
self.disable_edge_updates = disable_edge_updates
self.use_ln = use_ln
self.n_layers = n_layers
self.self_attn = torch.jit.script(
RTAttention(
d_node,
d_edge,
d_attn_hid,
n_heads,
modulate_v=modulate_v,
use_ln=use_ln,
)
)
# self.self_attn = RTAttention(d_hid, d_hid, d_hid, n_heads)
self.lin0 = Linear(d_node, d_node)
self.dropout0 = nn.Dropout(dropout)
if use_ln:
self.node_ln0 = nn.LayerNorm(d_node)
self.node_ln1 = nn.LayerNorm(d_node)
else:
self.node_ln0 = nn.Identity()
self.node_ln1 = nn.Identity()
act_fn = nn.GELU
self.node_mlp = nn.Sequential(
Linear(d_node, d_node_hid, bias=False),
act_fn(),
Linear(d_node_hid, d_node),
nn.Dropout(dropout),
)
if not self.disable_edge_updates:
self.edge_updates = EdgeLayer(
d_node=d_node,
d_edge=d_edge,
d_edge_hid=d_edge_hid,
dropout=dropout,
act_fn=act_fn,
use_ln=use_ln,
)
else:
self.edge_updates = NoEdgeLayer()
if tfixit_init:
self.fixit_init()
def fixit_init(self):
temp_state_dict = self.state_dict()
n_layers = self.n_layers
for name, param in self.named_parameters():
if "weight" in name:
if name.split(".")[0] in ["node_mlp", "edge_mlp0", "edge_mlp1"]:
temp_state_dict[name] = (0.67 * (n_layers) ** (-1.0 / 4.0)) * param
elif name.split(".")[0] in ["self_attn"]:
temp_state_dict[name] = (0.67 * (n_layers) ** (-1.0 / 4.0)) * (
param * (2**0.5)
)
self.load_state_dict(temp_state_dict)
def node_updates(self, node_features, edge_features, mask):
out = self.self_attn(node_features, edge_features, mask)
attn_out, attn_weights = out
node_features = self.node_ln0(
node_features
+ self.dropout0(
self.lin0(attn_out)
)
)
node_features = self.node_ln1(node_features + self.node_mlp(node_features))
return node_features, attn_weights
def forward(self, node_features, edge_features, mask):
node_features, attn_weights = self.node_updates(node_features, edge_features, mask)
edge_features = self.edge_updates(node_features, edge_features, mask)
return node_features, edge_features, attn_weights
class EdgeLayer(nn.Module):
def __init__(
self,
*,
d_node,
d_edge,
d_edge_hid,
dropout,
act_fn,
use_ln=True,
) -> None:
super().__init__()
self.edge_mlp0 = EdgeMLP(
d_edge=d_edge,
d_node=d_node,
d_edge_hid=d_edge_hid,
act_fn=act_fn,
dropout=dropout,
)
self.edge_mlp1 = nn.Sequential(
Linear(d_edge, d_edge_hid, bias=False),
act_fn(),
Linear(d_edge_hid, d_edge),
nn.Dropout(dropout),
)
if use_ln:
self.eln0 = nn.LayerNorm(d_edge)
self.eln1 = nn.LayerNorm(d_edge)
else:
self.eln0 = nn.Identity()
self.eln1 = nn.Identity()
def forward(self, node_features, edge_features, mask):
edge_features = self.eln0(
edge_features + self.edge_mlp0(node_features, edge_features)
)
edge_features = self.eln1(edge_features + self.edge_mlp1(edge_features))
return edge_features
class NoEdgeLayer(nn.Module):
def forward(self, node_features, edge_features, mask):
return edge_features
class EdgeMLP(nn.Module):
def __init__(self, *, d_node, d_edge, d_edge_hid, act_fn, dropout):
super().__init__()
self.reverse_edge = Rearrange("b n m d -> b m n d")
self.lin0_e = Linear(2 * d_edge, d_edge_hid)
self.lin0_s = Linear(d_node, d_edge_hid)
self.lin0_t = Linear(d_node, d_edge_hid)
self.act = act_fn()
self.lin1 = Linear(d_edge_hid, d_edge)
self.drop = nn.Dropout(dropout)
def forward(self, node_features, edge_features):
source_nodes = (
self.lin0_s(node_features)
.unsqueeze(-2)
.expand(-1, -1, node_features.size(-2), -1)
)
target_nodes = (
self.lin0_t(node_features)
.unsqueeze(-3)
.expand(-1, node_features.size(-2), -1, -1)
)
# reversed_edge_features = self.reverse_edge(edge_features)
edge_features = self.lin0_e(
torch.cat([edge_features, self.reverse_edge(edge_features)], dim=-1)
)
edge_features = edge_features + source_nodes + target_nodes
edge_features = self.act(edge_features)
edge_features = self.lin1(edge_features)
edge_features = self.drop(edge_features)
return edge_features
class RTAttention(nn.Module):
def __init__(self, d_node, d_edge, d_hid, n_heads, modulate_v=None, use_ln=True):
super().__init__()
self.n_heads = n_heads
self.d_node = d_node
self.d_edge = d_edge
self.d_hid = d_hid
self.use_ln = use_ln
self.modulate_v = modulate_v
self.scale = 1 / (d_hid**0.5)
self.split_head_node = Rearrange("b n (h d) -> b h n d", h=n_heads)
self.split_head_edge = Rearrange("b n m (h d) -> b h n m d", h=n_heads)
self.cat_head_node = Rearrange("... h n d -> ... n (h d)", h=n_heads)
self.qkv_node = Linear(d_node, 3 * d_hid, bias=False)
self.edge_factor = 4 if modulate_v else 3
self.qkv_edge = Linear(d_edge, self.edge_factor * d_hid, bias=False)
self.proj_out = Linear(d_hid, d_node)
def forward(self, node_features, edge_features, mask):
qkv_node = self.qkv_node(node_features)
# qkv_node = rearrange(qkv_node, "b n (h d) -> b h n d", h=self.n_heads)
qkv_node = self.split_head_node(qkv_node)
q_node, k_node, v_node = torch.chunk(qkv_node, 3, dim=-1)
qkv_edge = self.qkv_edge(edge_features)
# qkv_edge = rearrange(qkv_edge, "b n m (h d) -> b h n m d", h=self.n_heads)
qkv_edge = self.split_head_edge(qkv_edge)
qkv_edge = torch.chunk(qkv_edge, self.edge_factor, dim=-1)
# q_edge, k_edge, v_edge, q_edge_b, k_edge_b, v_edge_b = torch.chunk(
# qkv_edge, 6, dim=-1
# )
# qkv_edge = [item.masked_fill(mask.unsqueeze(1) == 0, 0) for item in qkv_edge]
q = q_node.unsqueeze(-2) + qkv_edge[0] # + q_edge_b
k = k_node.unsqueeze(-3) + qkv_edge[1] # + k_edge_b
if self.modulate_v:
v = v_node.unsqueeze(-3) * qkv_edge[3] + qkv_edge[2]
else:
v = v_node.unsqueeze(-3) + qkv_edge[2]
dots = self.scale * torch.einsum("b h i j d, b h i j d -> b h i j", q, k)
# dots.masked_fill_(mask.unsqueeze(1).squeeze(-1) == 0, -1e-9)
attn = F.softmax(dots, dim=-1)
out = torch.einsum("b h i j, b h i j d -> b h i d", attn, v)
out = self.cat_head_node(out)
return self.proj_out(out), attn
def Linear(in_features, out_features, bias=True):
m = nn.Linear(in_features, out_features, bias)
nn.init.xavier_uniform_(m.weight) # , gain=1 / math.sqrt(2))
if bias:
nn.init.constant_(m.bias, 0.0)
return m
|