-
Notifications
You must be signed in to change notification settings - Fork 20
/
models_graph_classification_ogb_original.py
268 lines (207 loc) · 10.7 KB
/
models_graph_classification_ogb_original.py
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
import copy
import torch
import torch.nn as nn
import torch.nn.functional as F
from graph_filters.GSN_edge_sparse_ogb import GSN_edge_sparse_ogb
from graph_filters.MPNN_edge_sparse_ogb import MPNN_edge_sparse_ogb
from models_misc import mlp, choose_activation
from utils_graph_learning import global_add_pool_sparse, global_mean_pool_sparse, DiscreteEmbedding
from ogb.graphproppred.mol_encoder import AtomEncoder, BondEncoder
class GNN_OGB(torch.nn.Module):
def __init__(self,
in_features,
out_features,
encoder_ids,
d_in_id,
in_edge_features=None,
d_in_node_encoder=None,
d_in_edge_encoder=None,
encoder_degrees=None,
d_degree=None,
**kwargs):
super(GNN_OGB, self).__init__()
seed = kwargs['seed']
#-------------- Initializations
self.model_name = kwargs['model_name']
self.readout = kwargs['readout'] if kwargs['readout'] is not None else 'sum'
self.dropout_features = kwargs['dropout_features']
self.bn = kwargs['bn']
self.final_projection = kwargs['final_projection']
self.residual = kwargs['residual']
self.inject_ids = kwargs['inject_ids']
self.vn = kwargs['vn']
id_scope = kwargs['id_scope']
d_msg = kwargs['d_msg']
d_out = kwargs['d_out']
d_h = kwargs['d_h']
aggr = kwargs['aggr'] if kwargs['aggr'] is not None else 'add'
flow = kwargs['flow'] if kwargs['flow'] is not None else 'target_to_source'
msg_kind = kwargs['msg_kind'] if kwargs['msg_kind'] is not None else 'general'
train_eps = kwargs['train_eps'] if kwargs['train_eps'] is not None else [False for _ in range(len(d_out))]
activation_mlp = kwargs['activation_mlp']
bn_mlp = kwargs['bn_mlp']
jk_mlp = kwargs['jk_mlp']
degree_embedding = kwargs['degree_embedding'] if kwargs['degree_as_tag'][0] else 'None'
degree_as_tag = kwargs['degree_as_tag']
retain_features = kwargs['retain_features']
encoders_kwargs = {'seed':seed,
'activation_mlp': activation_mlp,
'bn_mlp': bn_mlp,
'aggr': kwargs['multi_embedding_aggr'],
'features_scope': kwargs['features_scope']}
#-------------- Input node embedding
self.input_node_encoder = DiscreteEmbedding(kwargs['input_node_encoder'],
in_features,
d_in_node_encoder,
kwargs['d_out_node_encoder'],
**encoders_kwargs)
d_in = self.input_node_encoder.d_out
#-------------- Virtual node embedding
if self.vn:
vn_encoder_kwargs = copy.deepcopy(encoders_kwargs)
vn_encoder_kwargs['init'] = 'zeros'
self.vn_encoder = DiscreteEmbedding(kwargs['input_vn_encoder'],
1,
[1],
kwargs['d_out_vn_encoder'],
**vn_encoder_kwargs)
d_in_vn = self.vn_encoder.d_out
#-------------- Edge embedding (for each GNN layer)
self.edge_encoder = []
d_ef = []
for i in range(len(d_out)):
edge_encoder_layer = DiscreteEmbedding(kwargs['edge_encoder'],
in_edge_features,
d_in_edge_encoder,
kwargs['d_out_edge_encoder'][i],
**encoders_kwargs)
self.edge_encoder.append(edge_encoder_layer)
d_ef.append(edge_encoder_layer.d_out)
self.edge_encoder = nn.ModuleList(self.edge_encoder)
# -------------- Identifier embedding (for each GNN layer)
self.id_encoder = []
d_id = []
num_id_encoders = len(d_out) if kwargs['inject_ids'] else 1
for i in range(num_id_encoders):
id_encoder_layer = DiscreteEmbedding(kwargs['id_embedding'],
len(d_in_id),
d_in_id,
kwargs['d_out_id_embedding'],
**encoders_kwargs)
self.id_encoder.append(id_encoder_layer)
d_id.append(id_encoder_layer.d_out)
self.id_encoder = nn.ModuleList(self.id_encoder)
#-------------- Degree embedding
self.degree_encoder = DiscreteEmbedding(degree_embedding,
1,
d_degree,
kwargs['d_out_degree_embedding'],
**encoders_kwargs)
d_degree = self.degree_encoder.d_out
#-------------- GNN layers w/ bn
self.conv = []
self.batch_norms = []
self.mlp_vn = []
for i in range(len(d_out)):
if i > 0 and self.vn:
#-------------- vn msg function
mlp_vn_temp = mlp(d_in_vn, kwargs['d_out_vn'][i-1], d_h[i], seed, activation_mlp, bn_mlp)
self.mlp_vn.append(mlp_vn_temp)
d_in_vn= kwargs['d_out_vn'][i-1]
kwargs_filter = {
'd_in': d_in,
'd_degree': d_degree,
'degree_as_tag': degree_as_tag[i],
'retain_features': retain_features[i],
'd_msg': d_msg[i],
'd_up': d_out[i],
'd_h': d_h[i],
'seed': seed,
'activation_name': activation_mlp,
'bn': bn_mlp,
'aggr': aggr,
'msg_kind': msg_kind,
'eps': 0,
'train_eps': train_eps[i],
'flow': flow,
'd_ef': d_ef[i],
'edge_embedding': kwargs['edge_encoder'],
'id_embedding': kwargs['id_embedding'],
'extend_dims': kwargs['extend_dims']}
use_ids = ((i > 0 and kwargs['inject_ids']) or (i == 0)) and (self.model_name == 'GSN_edge_sparse_ogb')
if use_ids:
filter_fn = GSN_edge_sparse_ogb
kwargs_filter['d_id'] = d_id[i] if self.inject_ids else d_id[0]
kwargs_filter['id_scope'] = id_scope
else:
filter_fn = MPNN_edge_sparse_ogb
self.conv.append(filter_fn(**kwargs_filter))
bn_layer = nn.BatchNorm1d(d_out[i]) if self.bn[i] else None
self.batch_norms.append(bn_layer)
d_in = d_out[i]
self.conv = nn.ModuleList(self.conv)
self.batch_norms = nn.ModuleList(self.batch_norms)
if kwargs['vn']:
self.mlp_vn = nn.ModuleList(self.mlp_vn)
#-------------- Readout
if self.readout == 'sum':
self.global_pool = global_add_pool_sparse
elif self.readout == 'mean':
self.global_pool = global_mean_pool_sparse
else:
raise ValueError("Invalid graph pooling type.")
#-------------- Virtual node aggregation operator
if self.vn:
if kwargs['vn_pooling'] == 'sum':
self.global_vn_pool = global_add_pool_sparse
elif kwargs['vn_pooling'] == 'mean':
self.global_vn_pool = global_mean_pool_sparse
else:
raise ValueError("Invalid graph virtual node pooling type.")
self.lin_proj = nn.Linear(d_out[-1], out_features)
#-------------- Activation fn (same across the network)
self.activation = choose_activation(kwargs['activation'])
return
def forward(self, data, return_intermediate=False):
#-------------- Code adopted from https://github.com/snap-stanford/ogb/tree/master/examples/graphproppred/mol.
#-------------- Modified accordingly to allow for the existence of structural identifiers
kwargs = {}
kwargs['degrees'] = self.degree_encoder(data.degrees)
#-------------- edge index, initial node features enmbedding, initial vn embedding
edge_index = data.edge_index
if self.vn:
vn_embedding = self.vn_encoder(torch.zeros(data.batch[-1].item() + 1).to(edge_index.dtype).to(edge_index.device))
x = self.input_node_encoder(data.x)
x_interm = [x]
for i in range(0, len(self.conv)):
#-------------- encode ids (different for each layer)
kwargs['identifiers'] = self.id_encoder[i](data.identifiers) if self.inject_ids else self.id_encoder[0](data.identifiers)
#-------------- edge features embedding (different for each layer)
if hasattr(data, 'edge_features'):
kwargs['edge_features'] = self.edge_encoder[i](data.edge_features)
else:
kwargs['edge_features'] = None
if self.vn:
x_interm[i] = x_interm[i] + vn_embedding[data.batch]
x = self.conv[i](x_interm[i], edge_index, **kwargs)
x = self.batch_norms[i](x) if self.bn[i] else x
if i == len(self.conv) - 1:
x = F.dropout(x, self.dropout_features[i], training = self.training)
else:
x = F.dropout(self.activation(x), self.dropout_features[i], training = self.training)
if self.residual:
x += x_interm[-1]
x_interm.append(x)
if i < len(self.conv) - 1 and self.vn:
vn_embedding_temp = self.global_vn_pool(x_interm[i], data.batch) + vn_embedding
vn_embedding = self.mlp_vn[i](vn_embedding_temp)
if self.residual:
vn_embedding = vn_embedding + F.dropout(self.activation(vn_embedding), self.dropout_features[i], training = self.training)
else:
vn_embedding = F.dropout(self.activation(vn_embedding), self.dropout_features[i], training = self.training)
prediction = 0
for i in range(0,len(self.conv)+1):
if self.final_projection[i]:
prediction += x_interm[i]
x_global = self.global_pool(prediction, data.batch)
return self.lin_proj(x_global)