stcrpy 1.0.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- examples/__init__.py +0 -0
- examples/egnn.py +425 -0
- stcrpy/__init__.py +5 -0
- stcrpy/tcr_datasets/__init__.py +0 -0
- stcrpy/tcr_datasets/tcr_graph_dataset.py +499 -0
- stcrpy/tcr_datasets/tcr_selector.py +0 -0
- stcrpy/tcr_datasets/tcr_structure_dataset.py +0 -0
- stcrpy/tcr_datasets/utils.py +350 -0
- stcrpy/tcr_formats/__init__.py +0 -0
- stcrpy/tcr_formats/tcr_formats.py +114 -0
- stcrpy/tcr_formats/tcr_haddock.py +556 -0
- stcrpy/tcr_geometry/TCRCoM.py +350 -0
- stcrpy/tcr_geometry/TCRCoM_LICENCE +168 -0
- stcrpy/tcr_geometry/TCRDock.py +261 -0
- stcrpy/tcr_geometry/TCRGeom.py +450 -0
- stcrpy/tcr_geometry/TCRGeomFiltering.py +273 -0
- stcrpy/tcr_geometry/__init__.py +0 -0
- stcrpy/tcr_geometry/reference_data/__init__.py +0 -0
- stcrpy/tcr_geometry/reference_data/dock_reference_1_imgt_numbered.pdb +6549 -0
- stcrpy/tcr_geometry/reference_data/dock_reference_2_imgt_numbered.pdb +6495 -0
- stcrpy/tcr_geometry/reference_data/reference_A.pdb +31 -0
- stcrpy/tcr_geometry/reference_data/reference_B.pdb +31 -0
- stcrpy/tcr_geometry/reference_data/reference_D.pdb +31 -0
- stcrpy/tcr_geometry/reference_data/reference_G.pdb +31 -0
- stcrpy/tcr_geometry/reference_data/reference_data.py +104 -0
- stcrpy/tcr_interactions/PLIPParser.py +147 -0
- stcrpy/tcr_interactions/TCRInteractionProfiler.py +433 -0
- stcrpy/tcr_interactions/TCRpMHC_PLIP_Model_Parser.py +133 -0
- stcrpy/tcr_interactions/__init__.py +0 -0
- stcrpy/tcr_interactions/utils.py +170 -0
- stcrpy/tcr_methods/__init__.py +0 -0
- stcrpy/tcr_methods/tcr_batch_operations.py +223 -0
- stcrpy/tcr_methods/tcr_methods.py +150 -0
- stcrpy/tcr_methods/tcr_reformatting.py +18 -0
- stcrpy/tcr_metrics/__init__.py +2 -0
- stcrpy/tcr_metrics/constants.py +39 -0
- stcrpy/tcr_metrics/tcr_interface_rmsd.py +237 -0
- stcrpy/tcr_metrics/tcr_rmsd.py +179 -0
- stcrpy/tcr_ml/__init__.py +0 -0
- stcrpy/tcr_ml/geometry_predictor.py +3 -0
- stcrpy/tcr_processing/AGchain.py +89 -0
- stcrpy/tcr_processing/Chemical_components.py +48915 -0
- stcrpy/tcr_processing/Entity.py +301 -0
- stcrpy/tcr_processing/Fragment.py +58 -0
- stcrpy/tcr_processing/Holder.py +24 -0
- stcrpy/tcr_processing/MHC.py +449 -0
- stcrpy/tcr_processing/MHCchain.py +149 -0
- stcrpy/tcr_processing/Model.py +37 -0
- stcrpy/tcr_processing/Select.py +145 -0
- stcrpy/tcr_processing/TCR.py +532 -0
- stcrpy/tcr_processing/TCRIO.py +47 -0
- stcrpy/tcr_processing/TCRParser.py +1230 -0
- stcrpy/tcr_processing/TCRStructure.py +148 -0
- stcrpy/tcr_processing/TCRchain.py +160 -0
- stcrpy/tcr_processing/__init__.py +3 -0
- stcrpy/tcr_processing/annotate.py +480 -0
- stcrpy/tcr_processing/utils/__init__.py +0 -0
- stcrpy/tcr_processing/utils/common.py +67 -0
- stcrpy/tcr_processing/utils/constants.py +367 -0
- stcrpy/tcr_processing/utils/region_definitions.py +782 -0
- stcrpy/utils/__init__.py +0 -0
- stcrpy/utils/error_stream.py +12 -0
- stcrpy-1.0.0.dist-info/METADATA +173 -0
- stcrpy-1.0.0.dist-info/RECORD +68 -0
- stcrpy-1.0.0.dist-info/WHEEL +5 -0
- stcrpy-1.0.0.dist-info/licenses/LICENCE +28 -0
- stcrpy-1.0.0.dist-info/licenses/stcrpy/tcr_geometry/TCRCoM_LICENCE +168 -0
- stcrpy-1.0.0.dist-info/top_level.txt +2 -0
examples/__init__.py
ADDED
|
File without changes
|
examples/egnn.py
ADDED
|
@@ -0,0 +1,425 @@
|
|
|
1
|
+
# Code adapted from https://github.com/lucidrains/egnn-pytorch/blob/main/egnn_pytorch/egnn_pytorch_geometric.py
|
|
2
|
+
|
|
3
|
+
|
|
4
|
+
# MIT License
|
|
5
|
+
|
|
6
|
+
# Copyright (c) 2021 Phil Wang, Eric Alcaide
|
|
7
|
+
|
|
8
|
+
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
9
|
+
# of this software and associated documentation files (the "Software"), to deal
|
|
10
|
+
# in the Software without restriction, including without limitation the rights
|
|
11
|
+
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
12
|
+
# copies of the Software, and to permit persons to whom the Software is
|
|
13
|
+
# furnished to do so, subject to the following conditions:
|
|
14
|
+
|
|
15
|
+
# The above copyright notice and this permission notice shall be included in all
|
|
16
|
+
# copies or substantial portions of the Software.
|
|
17
|
+
|
|
18
|
+
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
19
|
+
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
20
|
+
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
21
|
+
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
22
|
+
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
23
|
+
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
24
|
+
# SOFTWARE.
|
|
25
|
+
|
|
26
|
+
|
|
27
|
+
import torch
|
|
28
|
+
from torch import nn, einsum, broadcast_tensors
|
|
29
|
+
import torch.nn.functional as F
|
|
30
|
+
|
|
31
|
+
from einops import rearrange, repeat
|
|
32
|
+
from einops.layers.torch import Rearrange
|
|
33
|
+
|
|
34
|
+
# helper functions
|
|
35
|
+
|
|
36
|
+
|
|
37
|
+
def exists(val):
|
|
38
|
+
return val is not None
|
|
39
|
+
|
|
40
|
+
|
|
41
|
+
def safe_div(num, den, eps=1e-8):
|
|
42
|
+
res = num.div(den.clamp(min=eps))
|
|
43
|
+
res.masked_fill_(den == 0, 0.0)
|
|
44
|
+
return res
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
def batched_index_select(values, indices, dim=1):
|
|
48
|
+
value_dims = values.shape[(dim + 1) :]
|
|
49
|
+
values_shape, indices_shape = map(lambda t: list(t.shape), (values, indices))
|
|
50
|
+
indices = indices[(..., *((None,) * len(value_dims)))]
|
|
51
|
+
indices = indices.expand(*((-1,) * len(indices_shape)), *value_dims)
|
|
52
|
+
value_expand_len = len(indices_shape) - (dim + 1)
|
|
53
|
+
values = values[(*((slice(None),) * dim), *((None,) * value_expand_len), ...)]
|
|
54
|
+
|
|
55
|
+
value_expand_shape = [-1] * len(values.shape)
|
|
56
|
+
expand_slice = slice(dim, (dim + value_expand_len))
|
|
57
|
+
value_expand_shape[expand_slice] = indices.shape[expand_slice]
|
|
58
|
+
values = values.expand(*value_expand_shape)
|
|
59
|
+
|
|
60
|
+
dim += value_expand_len
|
|
61
|
+
return values.gather(dim, indices)
|
|
62
|
+
|
|
63
|
+
|
|
64
|
+
def fourier_encode_dist(x, num_encodings=4, include_self=True):
|
|
65
|
+
x = x.unsqueeze(-1)
|
|
66
|
+
device, dtype, orig_x = x.device, x.dtype, x
|
|
67
|
+
scales = 2 ** torch.arange(num_encodings, device=device, dtype=dtype)
|
|
68
|
+
x = x / scales
|
|
69
|
+
x = torch.cat([x.sin(), x.cos()], dim=-1)
|
|
70
|
+
x = torch.cat((x, orig_x), dim=-1) if include_self else x
|
|
71
|
+
return x
|
|
72
|
+
|
|
73
|
+
|
|
74
|
+
def embedd_token(x, dims, layers):
|
|
75
|
+
stop_concat = -len(dims)
|
|
76
|
+
to_embedd = x[:, stop_concat:].long()
|
|
77
|
+
for i, emb_layer in enumerate(layers):
|
|
78
|
+
# the portion corresponding to `to_embedd` part gets dropped
|
|
79
|
+
x = torch.cat([x[:, :stop_concat], emb_layer(to_embedd[:, i])], dim=-1)
|
|
80
|
+
stop_concat = x.shape[-1]
|
|
81
|
+
return x
|
|
82
|
+
|
|
83
|
+
|
|
84
|
+
# swish activation fallback
|
|
85
|
+
|
|
86
|
+
|
|
87
|
+
class Swish_(nn.Module):
|
|
88
|
+
def forward(self, x):
|
|
89
|
+
return x * x.sigmoid()
|
|
90
|
+
|
|
91
|
+
|
|
92
|
+
SiLU = nn.SiLU if hasattr(nn, "SiLU") else Swish_
|
|
93
|
+
|
|
94
|
+
# helper classes
|
|
95
|
+
|
|
96
|
+
# this follows the same strategy for normalization as done in SE3 Transformers
|
|
97
|
+
# https://github.com/lucidrains/se3-transformer-pytorch/blob/main/se3_transformer_pytorch/se3_transformer_pytorch.py#L95
|
|
98
|
+
|
|
99
|
+
|
|
100
|
+
class CoorsNorm(nn.Module):
|
|
101
|
+
def __init__(self, eps=1e-8, scale_init=1.0):
|
|
102
|
+
super().__init__()
|
|
103
|
+
self.eps = eps
|
|
104
|
+
scale = torch.zeros(1).fill_(scale_init)
|
|
105
|
+
self.scale = nn.Parameter(scale)
|
|
106
|
+
|
|
107
|
+
def forward(self, coors):
|
|
108
|
+
norm = coors.norm(dim=-1, keepdim=True)
|
|
109
|
+
normed_coors = coors / norm.clamp(min=self.eps)
|
|
110
|
+
return normed_coors * self.scale
|
|
111
|
+
|
|
112
|
+
|
|
113
|
+
# global linear attention
|
|
114
|
+
|
|
115
|
+
|
|
116
|
+
class Attention(nn.Module):
|
|
117
|
+
def __init__(self, dim, heads=8, dim_head=64):
|
|
118
|
+
super().__init__()
|
|
119
|
+
inner_dim = heads * dim_head
|
|
120
|
+
self.heads = heads
|
|
121
|
+
self.scale = dim_head**-0.5
|
|
122
|
+
|
|
123
|
+
self.to_q = nn.Linear(dim, inner_dim, bias=False)
|
|
124
|
+
self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
|
|
125
|
+
self.to_out = nn.Linear(inner_dim, dim)
|
|
126
|
+
|
|
127
|
+
def forward(self, x, context, mask=None):
|
|
128
|
+
h = self.heads
|
|
129
|
+
|
|
130
|
+
q = self.to_q(x)
|
|
131
|
+
kv = self.to_kv(context).chunk(2, dim=-1)
|
|
132
|
+
|
|
133
|
+
q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, *kv))
|
|
134
|
+
dots = einsum("b h i d, b h j d -> b h i j", q, k) * self.scale
|
|
135
|
+
|
|
136
|
+
if exists(mask):
|
|
137
|
+
mask_value = -torch.finfo(dots.dtype).max
|
|
138
|
+
mask = rearrange(mask, "b n -> b () () n")
|
|
139
|
+
dots.masked_fill_(~mask, mask_value)
|
|
140
|
+
|
|
141
|
+
attn = dots.softmax(dim=-1)
|
|
142
|
+
out = einsum("b h i j, b h j d -> b h i d", attn, v)
|
|
143
|
+
|
|
144
|
+
out = rearrange(out, "b h n d -> b n (h d)", h=h)
|
|
145
|
+
return self.to_out(out)
|
|
146
|
+
|
|
147
|
+
|
|
148
|
+
class GlobalLinearAttention(nn.Module):
|
|
149
|
+
def __init__(self, *, dim, heads=8, dim_head=64):
|
|
150
|
+
super().__init__()
|
|
151
|
+
self.norm_seq = nn.LayerNorm(dim)
|
|
152
|
+
self.norm_queries = nn.LayerNorm(dim)
|
|
153
|
+
self.attn1 = Attention(dim, heads, dim_head)
|
|
154
|
+
self.attn2 = Attention(dim, heads, dim_head)
|
|
155
|
+
|
|
156
|
+
self.ff = nn.Sequential(
|
|
157
|
+
nn.LayerNorm(dim),
|
|
158
|
+
nn.Linear(dim, dim * 4),
|
|
159
|
+
nn.GELU(),
|
|
160
|
+
nn.Linear(dim * 4, dim),
|
|
161
|
+
)
|
|
162
|
+
|
|
163
|
+
def forward(self, x, queries, mask=None):
|
|
164
|
+
res_x, res_queries = x, queries
|
|
165
|
+
x, queries = self.norm_seq(x), self.norm_queries(queries)
|
|
166
|
+
|
|
167
|
+
induced = self.attn1(queries, x, mask=mask)
|
|
168
|
+
out = self.attn2(x, induced)
|
|
169
|
+
|
|
170
|
+
x = out + res_x
|
|
171
|
+
queries = induced + res_queries
|
|
172
|
+
|
|
173
|
+
x = self.ff(x) + x
|
|
174
|
+
return x, queries
|
|
175
|
+
|
|
176
|
+
|
|
177
|
+
# classes
|
|
178
|
+
|
|
179
|
+
|
|
180
|
+
class EGNN(nn.Module):
|
|
181
|
+
def __init__(
|
|
182
|
+
self,
|
|
183
|
+
dim,
|
|
184
|
+
edge_dim=0,
|
|
185
|
+
m_dim=16,
|
|
186
|
+
fourier_features=0,
|
|
187
|
+
num_nearest_neighbors=0,
|
|
188
|
+
dropout=0.0,
|
|
189
|
+
init_eps=1e-3,
|
|
190
|
+
norm_feats=False,
|
|
191
|
+
norm_coors=False,
|
|
192
|
+
norm_coors_scale_init=1e-2,
|
|
193
|
+
update_feats=True,
|
|
194
|
+
update_coors=True,
|
|
195
|
+
only_sparse_neighbors=False,
|
|
196
|
+
valid_radius=float("inf"),
|
|
197
|
+
m_pool_method="sum",
|
|
198
|
+
soft_edges=False,
|
|
199
|
+
coor_weights_clamp_value=None,
|
|
200
|
+
return_edges=False,
|
|
201
|
+
):
|
|
202
|
+
super().__init__()
|
|
203
|
+
assert m_pool_method in {
|
|
204
|
+
"sum",
|
|
205
|
+
"mean",
|
|
206
|
+
}, "pool method must be either sum or mean"
|
|
207
|
+
assert (
|
|
208
|
+
update_feats or update_coors
|
|
209
|
+
), "you must update either features, coordinates, or both"
|
|
210
|
+
|
|
211
|
+
self.fourier_features = fourier_features
|
|
212
|
+
|
|
213
|
+
edge_input_dim = (fourier_features * 2) + (dim * 2) + edge_dim + 1
|
|
214
|
+
dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
|
|
215
|
+
|
|
216
|
+
self.edge_mlp = nn.Sequential(
|
|
217
|
+
nn.Linear(edge_input_dim, edge_input_dim * 2),
|
|
218
|
+
dropout,
|
|
219
|
+
SiLU(),
|
|
220
|
+
nn.Linear(edge_input_dim * 2, m_dim),
|
|
221
|
+
SiLU(),
|
|
222
|
+
)
|
|
223
|
+
|
|
224
|
+
self.edge_gate = (
|
|
225
|
+
nn.Sequential(nn.Linear(m_dim, 1), nn.Sigmoid()) if soft_edges else None
|
|
226
|
+
)
|
|
227
|
+
|
|
228
|
+
self.node_norm = nn.LayerNorm(dim) if norm_feats else nn.Identity()
|
|
229
|
+
self.coors_norm = (
|
|
230
|
+
CoorsNorm(scale_init=norm_coors_scale_init) if norm_coors else nn.Identity()
|
|
231
|
+
)
|
|
232
|
+
|
|
233
|
+
self.m_pool_method = m_pool_method
|
|
234
|
+
|
|
235
|
+
self.node_mlp = (
|
|
236
|
+
nn.Sequential(
|
|
237
|
+
nn.Linear(dim + m_dim, dim * 2),
|
|
238
|
+
dropout,
|
|
239
|
+
SiLU(),
|
|
240
|
+
nn.Linear(dim * 2, dim),
|
|
241
|
+
)
|
|
242
|
+
if update_feats
|
|
243
|
+
else None
|
|
244
|
+
)
|
|
245
|
+
|
|
246
|
+
self.coors_mlp = (
|
|
247
|
+
nn.Sequential(
|
|
248
|
+
nn.Linear(m_dim, m_dim * 4), dropout, SiLU(), nn.Linear(m_dim * 4, 1)
|
|
249
|
+
)
|
|
250
|
+
if update_coors
|
|
251
|
+
else None
|
|
252
|
+
)
|
|
253
|
+
|
|
254
|
+
self.num_nearest_neighbors = num_nearest_neighbors
|
|
255
|
+
self.only_sparse_neighbors = only_sparse_neighbors
|
|
256
|
+
self.valid_radius = valid_radius
|
|
257
|
+
|
|
258
|
+
self.coor_weights_clamp_value = coor_weights_clamp_value
|
|
259
|
+
|
|
260
|
+
self.init_eps = init_eps
|
|
261
|
+
self.apply(self.init_)
|
|
262
|
+
|
|
263
|
+
self.return_edges = return_edges
|
|
264
|
+
|
|
265
|
+
def init_(self, module):
|
|
266
|
+
if type(module) in {nn.Linear}:
|
|
267
|
+
# seems to be needed to keep the network from exploding to NaN with greater depths
|
|
268
|
+
nn.init.normal_(module.weight, std=self.init_eps)
|
|
269
|
+
|
|
270
|
+
def forward(self, feats, coors, edges=None, mask=None, adj_mat=None):
|
|
271
|
+
(
|
|
272
|
+
b,
|
|
273
|
+
n,
|
|
274
|
+
d,
|
|
275
|
+
device,
|
|
276
|
+
fourier_features,
|
|
277
|
+
num_nearest,
|
|
278
|
+
valid_radius,
|
|
279
|
+
only_sparse_neighbors,
|
|
280
|
+
) = (
|
|
281
|
+
*feats.shape,
|
|
282
|
+
feats.device,
|
|
283
|
+
self.fourier_features,
|
|
284
|
+
self.num_nearest_neighbors,
|
|
285
|
+
self.valid_radius,
|
|
286
|
+
self.only_sparse_neighbors,
|
|
287
|
+
)
|
|
288
|
+
|
|
289
|
+
if exists(mask):
|
|
290
|
+
num_nodes = mask.sum(dim=-1)
|
|
291
|
+
|
|
292
|
+
use_nearest = num_nearest > 0 or only_sparse_neighbors
|
|
293
|
+
|
|
294
|
+
rel_coors = rearrange(coors, "b i d -> b i () d") - rearrange(
|
|
295
|
+
coors, "b j d -> b () j d"
|
|
296
|
+
)
|
|
297
|
+
rel_dist = (rel_coors**2).sum(dim=-1, keepdim=True)
|
|
298
|
+
|
|
299
|
+
i = j = n
|
|
300
|
+
|
|
301
|
+
if use_nearest:
|
|
302
|
+
ranking = rel_dist[..., 0].clone()
|
|
303
|
+
|
|
304
|
+
if exists(mask):
|
|
305
|
+
rank_mask = mask[:, :, None] * mask[:, None, :]
|
|
306
|
+
ranking.masked_fill_(~rank_mask, 1e5)
|
|
307
|
+
|
|
308
|
+
if exists(adj_mat):
|
|
309
|
+
if len(adj_mat.shape) == 2:
|
|
310
|
+
adj_mat = repeat(adj_mat.clone(), "i j -> b i j", b=b)
|
|
311
|
+
|
|
312
|
+
if only_sparse_neighbors:
|
|
313
|
+
num_nearest = int(adj_mat.float().sum(dim=-1).max().item())
|
|
314
|
+
valid_radius = 0
|
|
315
|
+
|
|
316
|
+
self_mask = rearrange(
|
|
317
|
+
torch.eye(n, device=device, dtype=torch.bool), "i j -> () i j"
|
|
318
|
+
)
|
|
319
|
+
|
|
320
|
+
adj_mat = adj_mat.masked_fill(self_mask, False)
|
|
321
|
+
ranking.masked_fill_(self_mask, -1.0)
|
|
322
|
+
ranking.masked_fill_(adj_mat, 0.0)
|
|
323
|
+
|
|
324
|
+
nbhd_ranking, nbhd_indices = ranking.topk(
|
|
325
|
+
num_nearest, dim=-1, largest=False
|
|
326
|
+
)
|
|
327
|
+
|
|
328
|
+
nbhd_mask = nbhd_ranking <= valid_radius
|
|
329
|
+
|
|
330
|
+
rel_coors = batched_index_select(rel_coors, nbhd_indices, dim=2)
|
|
331
|
+
rel_dist = batched_index_select(rel_dist, nbhd_indices, dim=2)
|
|
332
|
+
|
|
333
|
+
if exists(edges):
|
|
334
|
+
edges = batched_index_select(edges, nbhd_indices, dim=2)
|
|
335
|
+
|
|
336
|
+
j = num_nearest
|
|
337
|
+
|
|
338
|
+
if fourier_features > 0:
|
|
339
|
+
rel_dist = fourier_encode_dist(rel_dist, num_encodings=fourier_features)
|
|
340
|
+
rel_dist = rearrange(rel_dist, "b i j () d -> b i j d")
|
|
341
|
+
|
|
342
|
+
if use_nearest:
|
|
343
|
+
feats_j = batched_index_select(feats, nbhd_indices, dim=1)
|
|
344
|
+
else:
|
|
345
|
+
feats_j = rearrange(feats, "b j d -> b () j d")
|
|
346
|
+
|
|
347
|
+
feats_i = rearrange(feats, "b i d -> b i () d")
|
|
348
|
+
feats_i, feats_j = broadcast_tensors(feats_i, feats_j)
|
|
349
|
+
|
|
350
|
+
edge_input = torch.cat((feats_i, feats_j, rel_dist), dim=-1)
|
|
351
|
+
|
|
352
|
+
if exists(edges):
|
|
353
|
+
edge_input = torch.cat((edge_input, edges), dim=-1)
|
|
354
|
+
|
|
355
|
+
m_ij = self.edge_mlp(edge_input)
|
|
356
|
+
|
|
357
|
+
if exists(self.edge_gate):
|
|
358
|
+
m_ij = m_ij * self.edge_gate(m_ij)
|
|
359
|
+
|
|
360
|
+
if exists(mask):
|
|
361
|
+
mask_i = rearrange(mask, "b i -> b i ()")
|
|
362
|
+
|
|
363
|
+
if use_nearest:
|
|
364
|
+
mask_j = batched_index_select(mask, nbhd_indices, dim=1)
|
|
365
|
+
mask = (mask_i * mask_j) & nbhd_mask
|
|
366
|
+
else:
|
|
367
|
+
mask_j = rearrange(mask, "b j -> b () j")
|
|
368
|
+
mask = mask_i * mask_j
|
|
369
|
+
|
|
370
|
+
if exists(self.coors_mlp):
|
|
371
|
+
coor_weights = self.coors_mlp(m_ij)
|
|
372
|
+
coor_weights = rearrange(coor_weights, "b i j () -> b i j")
|
|
373
|
+
|
|
374
|
+
rel_coors = self.coors_norm(rel_coors)
|
|
375
|
+
|
|
376
|
+
if exists(mask):
|
|
377
|
+
coor_weights.masked_fill_(~mask, 0.0)
|
|
378
|
+
|
|
379
|
+
if exists(self.coor_weights_clamp_value):
|
|
380
|
+
clamp_value = self.coor_weights_clamp_value
|
|
381
|
+
coor_weights.clamp_(min=-clamp_value, max=clamp_value)
|
|
382
|
+
|
|
383
|
+
coors_out = (
|
|
384
|
+
einsum("b i j, b i j c -> b i c", coor_weights, rel_coors) + coors
|
|
385
|
+
)
|
|
386
|
+
else:
|
|
387
|
+
coors_out = coors
|
|
388
|
+
|
|
389
|
+
if exists(self.node_mlp):
|
|
390
|
+
if exists(mask):
|
|
391
|
+
m_ij_mask = rearrange(mask, "... -> ... ()")
|
|
392
|
+
m_ij = m_ij.masked_fill(~m_ij_mask, 0.0)
|
|
393
|
+
|
|
394
|
+
if self.m_pool_method == "mean":
|
|
395
|
+
if exists(mask):
|
|
396
|
+
# masked mean
|
|
397
|
+
mask_sum = m_ij_mask.sum(dim=-2)
|
|
398
|
+
m_i = safe_div(m_ij.sum(dim=-2), mask_sum)
|
|
399
|
+
else:
|
|
400
|
+
m_i = m_ij.mean(dim=-2)
|
|
401
|
+
|
|
402
|
+
elif self.m_pool_method == "sum":
|
|
403
|
+
m_i = m_ij.sum(dim=-2)
|
|
404
|
+
|
|
405
|
+
normed_feats = self.node_norm(feats)
|
|
406
|
+
node_mlp_input = torch.cat((normed_feats, m_i), dim=-1)
|
|
407
|
+
node_out = self.node_mlp(node_mlp_input) + feats
|
|
408
|
+
else:
|
|
409
|
+
node_out = feats
|
|
410
|
+
|
|
411
|
+
if self.return_edges:
|
|
412
|
+
if exists(num_nearest):
|
|
413
|
+
num_neighbours = num_nearest
|
|
414
|
+
else:
|
|
415
|
+
num_neighbours = n
|
|
416
|
+
edges_out = torch.zeros(
|
|
417
|
+
(b, n, n, m_ij.shape[-1])
|
|
418
|
+
) # initialise full edge matrix
|
|
419
|
+
edges_out.scatter_(
|
|
420
|
+
2, nbhd_indices.unsqueeze(-1).expand(-1, -1, -1, m_ij.shape[-1]), m_ij
|
|
421
|
+
)
|
|
422
|
+
# assert torch.stack([edges_out[0, i, nbhd_indices[0, i, idx]] == m_ij[0, i, idx] for idx in range(num_neighbours) for i in range(n)]).all() # check assignemnt has worked, comment out at runtime
|
|
423
|
+
return node_out, coors_out, edges_out
|
|
424
|
+
|
|
425
|
+
return node_out, coors_out
|
stcrpy/__init__.py
ADDED
|
File without changes
|