-
Notifications
You must be signed in to change notification settings - Fork 30
/
_wacsf.py
366 lines (323 loc) · 19.7 KB
/
_wacsf.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
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
363
364
365
366
import numpy as np
import keras as ks
from keras import ops
from keras.layers import Layer, Multiply, Subtract
from kgcnn.layers.gather import GatherNodesOutgoing, GatherNodes, GatherNodesIngoing
from kgcnn.layers.geom import NodeDistanceEuclidean, NodePosition
from kgcnn.layers.aggr import AggregateLocalEdges
from kgcnn.layers.modules import ExpandDims
# Parameters for Weighted Atom-Centered Symmetry Functions (wACSF).
# We set the unoptimized C default values as default for all 118 atomic species.
# They are taken from https://aip.scitation.org/doi/suppl/10.1063/1.5019667/suppl_file/si.pdf .
# We chose the 22:10 radial:angular as default values. Other combinations ranging from 0:32 to 32:0 .
radial_eta_mu_defaults = np.array([[
[4.5000000, 7.5000000, 8.0], [4.5000000, 7.1667000, 8.0], [4.5000000, 6.8333000, 8.0], [4.5000000, 6.5000000, 8.0],
[4.5000000, 6.1667000, 8.0], [4.5000000, 5.8333000, 8.0], [4.5000000, 5.5000000, 8.0], [4.5000000, 5.1667000, 8.0],
[4.5000000, 4.8333000, 8.0], [4.5000000, 4.5000000, 8.0], [4.5000000, 4.1667000, 8.0], [4.5000000, 3.8333000, 8.0],
[4.5000000, 3.5000000, 8.0], [4.5000000, 3.1667000, 8.0], [4.5000000, 2.8333000, 8.0], [4.5000000, 2.5000000, 8.0],
[4.5000000, 2.1667000, 8.0], [4.5000000, 1.8333000, 8.0], [4.5000000, 1.5000000, 8.0], [4.5000000, 1.1667000, 8.0],
[4.5000000, 0.8333000, 8.0], [4.5000000, 0.5000000, 8.0]
]] * 118)
angular_eta_mu_lambda_zeta_defaults = np.array(
[[[0.03306120, 0.0, -1.0, 1.0, 8.00], [0.03306120, 0.0, 1.0, 1.0, 8.00], [0.04986150, 0.0, -1.0, 1.0, 8.00],
[0.04986150, 0.0, 1.0, 1.0, 8.00], [0.08367770, 0.0, -1.0, 1.0, 8.00], [0.08367770, 0.0, 1.0, 1.0, 8.00],
[0.16857440, 0.0, -1.0, 1.0, 8.00], [0.16857440, 0.0, 1.0, 1.0, 8.00], [0.50000000, 0.0, -1.0, 1.0, 8.00],
[0.50000000, 0.0, 1.0, 1.0, 8.00]]] * 118)
# Update optimized parameters for H, C, F, O, N .
radial_eta_mu_defaults[1] = np.array(
[[4.3814240, 2.4432533, 8.0], [4.4149216, 6.1667000, 8.0], [4.4493159, 3.8332928, 8.0], [4.4815889, 3.1667000, 8.0],
[4.4865336, 3.5000000, 8.0], [4.4937252, 4.8332601, 8.0], [4.4986009, 5.5010689, 8.0], [4.4997975, 2.1558150, 8.0],
[4.4999291, 7.4956238, 8.0], [4.4999905, 6.4964721, 8.0], [4.5000000, 5.2730567, 8.0], [4.5000011, 1.1606611, 8.0],
[4.5000039, 0.4918029, 8.0], [4.5000398, 0.8333002, 8.0], [4.5002241, 1.8334858, 8.0], [4.5036121, 2.9532907, 8.0],
[4.5059680, 5.8330954, 8.0], [4.5062245, 6.8333000, 8.0], [4.5064492, 7.1666408, 8.0], [4.5675834, 4.5046412, 8.0],
[4.5899630, 4.1663934, 8.0], [4.6604619, 1.4989046, 8.0]])
angular_eta_mu_lambda_zeta_defaults[1] = np.array(
[[0.01919876, 0.0, -1.0, 1.0, 8.00], [0.03306120, 0.0, 1.0, 1.0, 8.00], [0.03447591, 0.0, -1.0, 2.0, 8.00],
[0.04986150, 0.0, -1.0, 1.0, 8.00], [0.08235996, 0.0, -1.0, 1.0, 8.00], [0.08367770, 0.0, 1.0, 2.0, 8.00],
[0.16857440, 0.0, -1.0, 2.0, 8.00], [0.16953019, 0.0, 1.0, 2.0, 8.00], [0.50000000, 0.0, -1.0, 1.0, 8.00],
[0.50000000, 0.0, 1.0, 1.0, 8.00]]
)
radial_eta_mu_defaults[6] = np.array(
[[4.4534562, 3.8338779, 8.0], [4.4804536, 5.1667000, 8.0], [4.4918531, 4.8333000, 8.0], [4.4991551, 0.8333000, 8.0],
[4.4998245, 2.4997843, 8.0], [4.4999043, 1.5993354, 8.0], [4.4999230, 5.8324479, 8.0], [4.4999997, 6.5117668, 8.0],
[4.5000000, 6.1860418, 8.0], [4.5000000, 3.1584848, 8.0], [4.5000001, 4.4991770, 8.0], [4.5000132, 6.8333000, 8.0],
[4.5001099, 7.4999458, 8.0], [4.5002911, 2.1638163, 8.0], [4.5004121, 0.3038147, 8.0], [4.5014183, 5.4361520, 8.0],
[4.5023686, 1.8302079, 8.0], [4.5027053, 3.5002020, 8.0], [4.5074706, 7.2656992, 8.0], [4.5104731, 1.1645345, 8.0],
[4.5515436, 4.1666210, 8.0], [4.5906934, 2.8328539, 8.0]])
angular_eta_mu_lambda_zeta_defaults[6] = np.array(
[[0.03392104, 0.0, 1.0, 2.0, 8.00], [0.04232205, 0.0, -1.0, 2.0, 8.00], [0.04663133, 0.0, -1.0, 2.0, 8.00],
[0.05097753, 0.0, -1.0, 1.0, 8.00], [0.08367770, 0.0, -1.0, 3.0, 8.00], [0.08368552, 0.0, -1.0, 1.0, 8.00],
[0.16857440, 0.0, 1.0, 2.0, 8.00], [0.25965214, 0.0, 1.0, 2.0, 8.00], [0.47868498, 0.0, -1.0, 1.0, 8.00],
[0.56008922, 0.0, 1.0, 1.0, 8.00]]
)
radial_eta_mu_defaults[9] = np.array(
[[4.3755129, 4.3891938, 8.0], [4.4739340, 6.5000000, 8.0], [4.4934064, 1.1667000, 8.0], [4.4955168, 3.1421155, 8.0],
[4.4982145, 6.1665305, 8.0], [4.4986470, 0.8333000, 8.0], [4.4989965, 7.4999930, 8.0], [4.4999944, 4.1666929, 8.0],
[4.4999998, 6.8333000, 8.0], [4.4999998, 2.8332546, 8.0], [4.4999999, 5.8424471, 8.0], [4.5000000, 5.4954750, 8.0],
[4.5000000, 4.8332314, 8.0], [4.5000000, 2.1675890, 8.0], [4.5000000, 1.8332281, 8.0], [4.5000016, 3.8334246, 8.0],
[4.5000047, 7.1667013, 8.0], [4.5002559, 0.5866000, 8.0], [4.5004160, 5.1667000, 8.0], [4.5165951, 1.5000000, 8.0],
[4.5200686, 3.5000000, 8.0], [4.5333816, 2.5065631, 8.0]])
angular_eta_mu_lambda_zeta_defaults[9] = np.array(
[[0.03302291, 0.0, -1.0, 1.0, 8.00], [0.03306120, 0.0, 1.0, 1.0, 8.00], [0.04986182, 0.0, -1.0, 1.0, 8.00],
[0.05024516, 0.0, 1.0, 1.0, 8.00], [0.08367770, 0.0, -1.0, 2.0, 8.00], [0.08367770, 0.0, 1.0, 1.0, 8.00],
[0.16857440, 0.0, 1.0, 2.0, 8.00], [0.16857440, 0.0, 1.0, 1.0, 8.00], [0.49854417, 0.0, 1.0, 1.0, 8.00],
[0.50003188, 0.0, 1.0, 1.0, 8.00]]
)
radial_eta_mu_defaults[8] = np.array(
[[4.3810310, 6.1634839, 8.0], [4.4613793, 4.7025864, 8.0], [4.4708866, 0.8337341, 8.0], [4.4741510, 7.1654091, 8.0],
[4.4934208, 0.4909963, 8.0], [4.4999979, 6.7895512, 8.0], [4.4999984, 2.5548726, 8.0], [4.4999996, 5.1188661, 8.0],
[4.4999999, 6.4941667, 8.0], [4.5000000, 7.5269793, 8.0], [4.5000000, 3.5008886, 8.0], [4.5000000, 3.1673548, 8.0],
[4.5000000, 3.0664616, 8.0], [4.5000000, 1.1771551, 8.0], [4.5000010, 5.9628537, 8.0], [4.5000305, 1.6101757, 8.0],
[4.5000505, 3.8352953, 8.0], [4.5000877, 5.2259600, 8.0], [4.5003116, 2.1667069, 8.0], [4.5007395, 1.8264699, 8.0],
[4.5037122, 4.1666850, 8.0], [4.5108571, 4.7474998, 8.0]])
angular_eta_mu_lambda_zeta_defaults[8] = np.array(
[[0.03306120, 0.0, -1.0, 1.0, 8.00], [0.04970837, 0.0, -1.0, 1.0, 8.00], [0.04986150, 0.0, 1.0, 1.0, 8.00],
[0.05680303, 0.0, -1.0, 1.0, 8.00], [0.08366889, 0.0, 1.0, 1.0, 8.00], [0.08367770, 0.0, -1.0, 1.0, 8.00],
[0.16857440, 0.0, 1.0, 2.0, 8.00], [0.16857440, 0.0, 1.0, 1.0, 8.00], [0.50000000, 0.0, -1.0, 1.0, 8.00],
[0.50000000, 0.0, 1.0, 1.0, 8.00]]
)
radial_eta_mu_defaults[7] = np.array(
[[4.4148209, 6.8472192, 8.0], [4.4481457, 1.0317668, 8.0], [4.4608631, 0.4581865, 8.0], [4.4797608, 3.9799903, 8.0],
[4.4951823, 0.8337448, 8.0], [4.4987030, 7.4573738, 8.0], [4.4998898, 6.9779219, 8.0], [4.4999884, 2.1676806, 8.0],
[4.4999920, 3.1667000, 8.0], [4.5000000, 6.1662980, 8.0], [4.5000000, 5.8328781, 8.0], [4.5000000, 5.4757644, 8.0],
[4.5000000, 5.1666981, 8.0], [4.5000000, 3.3518840, 8.0], [4.5000000, 2.4982784, 8.0], [4.5001975, 4.8204812, 8.0],
[4.5004709, 6.5050056, 8.0], [4.5016337, 4.1667000, 8.0], [4.5077984, 1.4994658, 8.0], [4.5228543, 2.8333000, 8.0],
[4.5780852, 4.5000000, 8.0], [4.5897686, 1.8340302, 8.0]])
angular_eta_mu_lambda_zeta_defaults[7] = np.array(
[[0.03306120, 0.0, -1.0, 1.0, 8.00], [0.03306120, 0.0, 1.0, 1.0, 8.00], [0.04907151, 0.0, -1.0, 2.0, 8.00],
[0.07102913, 0.0, 1.0, 1.0, 8.00], [0.07116163, 0.0, 1.0, 1.0, 8.00], [0.08384128, 0.0, -1.0, 1.0, 8.00],
[0.16857391, 0.0, -1.0, 2.0, 8.00], [0.16902220, 0.0, 1.0, 1.0, 8.00], [0.50000000, 0.0, -1.0, 1.0, 8.00],
[0.50000000, 0.0, 1.0, 1.0, 8.00]]
)
class wACSFRad(Layer):
r"""Weighted atom-centered symmetry functions (wACSF) for high-dimensional neural network potentials (HDNNPs).
From `Gastegger et al. (2017) <https://arxiv.org/abs/1712.05861>`__ .
Default values can be found in `<https://aip.scitation.org/doi/suppl/10.1063/1.5019667>`_ .
This layer implements the radial part :math:`W_{i}^{rad}` :
.. math::
W_{i}^{rad} = \sum_{j \neq i} \; g(Z_j) \; e^{−\eta \, (r_{ij} − \mu)^{2} } \; f_{ij}
Here, for each atom type there is a set of parameters :math:`\eta` and :math:`\mu` .
The cutoff function :math:`f_ij = f_c(r_{ij})` is given by:
.. math::
f_c(r_{ij}) = 0.5 [\cos{\frac{\pi r_{ij}}{R_c}} + 1]
The cutoff radius is implemented as a float and can not be changed dependent on the atom type.
Not that the parameters :math:`\eta` and :math:`\mu` can be made trainable.
"""
def __init__(self, eta_mu: list = None, cutoff: float = 8.0, add_eps: bool = False,
use_external_weights: bool = False, param_constraint=None, param_regularizer=None,
param_initializer="zeros", **kwargs):
super(wACSFRad, self).__init__(**kwargs)
self.cutoff = cutoff
self.add_eps = add_eps
if eta_mu is None:
eta_mu = radial_eta_mu_defaults[:, :, :2]
self.eta_mu = np.array(eta_mu, dtype="float").tolist()
self.use_external_weights = use_external_weights
self.lazy_mult = Multiply()
self.layer_pos = NodePosition()
self.layer_gather_out = GatherNodesOutgoing()
self.layer_gather_in = GatherNodesIngoing()
self.layer_exp_dims = ExpandDims(axis=1)
self.layer_dist = NodeDistanceEuclidean(add_eps=add_eps)
self.pool_sum = AggregateLocalEdges(pooling_method="sum")
# We can do this in init since weights do not depend on input shape.
self.param_initializer = ks.initializers.deserialize(param_initializer)
self.param_regularizer = ks.regularizers.deserialize(param_regularizer)
self.param_constraint = ks.constraints.deserialize(param_constraint)
np_eta_mu = np.array(eta_mu)
weight_shape = np_eta_mu.shape
self._weight_eta_mu = self.add_weight(
name="eta_mu",
shape=weight_shape,
initializer=self.param_initializer,
regularizer=self.param_regularizer,
constraint=self.param_constraint,
dtype=self.dtype, trainable=False
)
self.set_weights([np_eta_mu])
@staticmethod
def _compute_fc(inputs, cutoff: float):
fc = ops.clip(inputs, -cutoff, cutoff)
fc = (ops.cos(fc * np.pi / cutoff) + 1.0) * 0.5
# fc = tf.where(tf.abs(inputs) < self.cutoff, fc, tf.zeros_like(fc))
return fc
def _compute_gaussian_expansion(self, inputs):
rij, zi = inputs
params = ops.take(self._weight_eta_mu, zi, axis=0)
# eta, mu = ops.split(params, [1, 1], axis=-1)
# eta, mu = ops.squeeze(eta, axis=-1), ops.squeeze(mu, axis=-1)
eta, mu = ops.take(params, 0, axis=-1), ops.take(params, 1, axis=-1)
arg = ops.square(rij - mu) * eta
return ops.exp(-arg)
def build(self, input_shape):
super(wACSFRad, self).build(input_shape)
def call(self, inputs, mask=None, **kwargs):
r"""Forward pass.
Args:
inputs: [z, xyz, ij] or [z, xyz, ij, w] if `use_external_weights`
- z (tf.RaggedTensor): Atomic numbers of shape ([N], )
- xyz (tf.RaggedTensor): Node coordinates of shape ([N], 3)
- ij (tf.RaggedTensor): Edge indices referring to nodes of shape (2, [M])
- w (tf.RaggedTensor): Edge weight tensor of shape ([M], 1)
mask: Boolean mask for inputs. Not used. Defaults to None.
Returns:
Tensor: Atomic representation of shape `(None, units)` .
"""
if self.use_external_weights:
z, xyz, eij, w = inputs
z = ops.cast(z, dtype=eij.dtype)
else:
z, xyz, eij = inputs
z = ops.cast(z, dtype=eij.dtype)
zj = self.layer_gather_out([z, eij], **kwargs)
w = self.layer_exp_dims(zj, **kwargs)
w = ops.cast(w, dtype=self.dtype)
xi, xj = self.layer_pos([xyz, eij], **kwargs)
rij = self.layer_dist([xi, xj], **kwargs)
fc = self._compute_fc(rij, cutoff=self.cutoff)
zi = self.layer_gather_in([z, eij], **kwargs)
gij = self._compute_gaussian_expansion([rij, zi])
rep = self.lazy_mult([gij, fc, w], **kwargs)
return self.pool_sum([xyz, rep, eij], **kwargs)
def get_config(self):
config = super(wACSFRad, self).get_config()
config.update({"eta_mu": self.eta_mu, "cutoff": self.cutoff, "use_external_weights": self.use_external_weights,
"add_eps": self.add_eps,
"param_constraint": ks.constraints.serialize(self.param_constraint),
"param_regularizer": ks.regularizers.serialize(self.param_regularizer),
"param_initializer": ks.initializers.serialize(self.param_initializer)
})
return config
class wACSFAng(Layer):
r"""Weighted atom-centered symmetry functions (wACSF) for high-dimensional neural network potentials (HDNNPs).
From `Gastegger et al. (2017) <https://arxiv.org/abs/1712.05861>`__ .
Default values can be found in `<https://aip.scitation.org/doi/suppl/10.1063/1.5019667>`_ .
This layer implements the angular part :math:`W_{i}^{ang}` :
.. math::
W_{i}^{ang} = 2^{1−\zeta} \; \sum_{j\neq i} \sum_{k\neq j,i} \; h(Z_j , Z_k) \;
(1 + \lambda\, \cos{\theta_{ijk}})^\zeta\;
\times \; e^{−\eta (r_{ij} −\mu )^2} \; e^{−\eta (r_{ik} −\mu )^2} \; e^{−\eta (r_{jk} −\mu )^2} \;
\times \; f_{ij} \; f_{ik} \; f_{jk}
Here, for each atom type there is a set of parameters :math:`\eta` , :math:`\mu` , :math:`\lambda`
and :math:`\zeta`.
The cutoff function :math:`f_ij = f_c(r_{ij})` is given by:
.. math::
f_c(r_{ij}) = 0.5 [\cos{\frac{\pi r_{ij}}{R_c}} + 1]
The cutoff radius is implemented as a float and can not be changed dependent on the atom type.
Not that the parameters :math:`\eta` to :math:`\lambda` can be made trainable.
"""
def __init__(self, eta_mu_lambda_zeta: list = None, cutoff: float = 8.0,
add_eps: bool = False, use_external_weights: bool = False, param_initializer="zeros",
param_regularizer=None, param_constraint=None, **kwargs):
super(wACSFAng, self).__init__(**kwargs)
self.add_eps = add_eps
self.cutoff = cutoff
if eta_mu_lambda_zeta is None:
eta_mu_lambda_zeta = angular_eta_mu_lambda_zeta_defaults[:, :, :4]
self.eta_mu_lambda_zeta = np.array(eta_mu_lambda_zeta, dtype="float").tolist()
self.use_external_weights = use_external_weights
self.lazy_mult = Multiply()
self.layer_pos = NodePosition(selection_index=[0, 1, 2])
self.layer_dist = NodeDistanceEuclidean(add_eps=add_eps)
self.pool_sum = AggregateLocalEdges(pooling_method="sum")
self.lazy_sub = Subtract()
self.layer_gather_in = GatherNodesIngoing()
self.layer_gather_1 = GatherNodes(split_indices=[1], concat_axis=None)
self.layer_gather_2 = GatherNodes(split_indices=[2], concat_axis=None)
self.layer_exp_dims = ExpandDims(axis=1)
# We can do this in init since weights do not depend on input shape.
self.param_initializer = ks.initializers.deserialize(param_initializer)
self.param_regularizer = ks.regularizers.deserialize(param_regularizer)
self.param_constraint = ks.constraints.deserialize(param_constraint)
np_eta_mu_lambda_zeta = np.array(eta_mu_lambda_zeta)
weight_shape = np_eta_mu_lambda_zeta.shape
self._weight_eta_mu_lambda_zeta = self.add_weight(
name="eta_mu_lambda_zeta",
shape=weight_shape,
initializer=self.param_initializer,
regularizer=self.param_regularizer,
constraint=self.param_constraint,
dtype=self.dtype, trainable=False
)
self.set_weights([np_eta_mu_lambda_zeta])
@staticmethod
def _compute_fc(inputs, cutoff: float):
fc = ops.clip(inputs, -cutoff, cutoff)
fc = (ops.cos(fc * np.pi / cutoff) + 1.0) * 0.5
# fc = tf.where(tf.abs(inputs) < self.cutoff, fc, tf.zeros_like(fc))
return fc
def _compute_gaussian_expansion(self, inputs: list):
rij, zi = inputs
params = ops.take(self._weight_eta_mu_lambda_zeta, zi, axis=0)
eta, mu = ops.take(params, 0, axis=-1), ops.take(params, 1, axis=-1)
arg = ops.square(rij - mu) * eta
return ops.exp(-arg)
def _compute_pow_cos_angle_(self, inputs: list):
vij, vik, rij, rik, zi = inputs
params = ops.take(self._weight_eta_mu_lambda_zeta, zi, axis=0)
lamda, zeta = ops.take(params, 2, axis=-1), ops.take(params, 3, axis=-1)
cos_theta = ops.sum(vij * vik, axis=-1, keepdims=True) / rij / rik
cos_term = cos_theta * lamda + 1.0
cos_term = ops.power(cos_term, zeta)
return cos_term
def _compute_pow_scale(self, inputs: list):
to_scale, zi = inputs
params = ops.take(self._weight_eta_mu_lambda_zeta, zi, axis=0)
zeta = ops.take(params, 3, axis=-1)
scale = ops.ones_like(to_scale) * 2.0
scaled = ops.power(scale, 1.0 - zeta) * to_scale
return scaled
def build(self, input_shape):
super(wACSFAng, self).build(input_shape)
def call(self, inputs, mask=None, **kwargs):
r"""Forward pass.
Args:
inputs (list): [z, xyz, ijk] or [z, xyz, ijk, w] if `use_external_weights`
- z (Tensor): Atomic numbers of shape ([N])
- xyz (Tensor): Node coordinates of shape ([N], 3)
- ijk (Tensor): Angle indices referring to nodes of shape (3, [M])
- w (Tensor): Angle weight tensor of shape ([M], 1)
mask (list): Boolean mask for inputs. Not used. Defaults to None.
Returns:
Tensor: Atomic representation of shape `(None, units)` .
"""
if self.use_external_weights:
z, xyz, ijk, w = inputs
z = ops.cast(z, dtype=ijk.dtype)
w = ops.cast(w, dtype=self.dtype)
else:
z, xyz, ijk = inputs
z = ops.cast(z, dtype=ijk.dtype)
w1 = self.layer_gather_1([z, ijk], **kwargs)[0]
w2 = self.layer_gather_2([z, ijk], **kwargs)[0]
w1 = ops.cast(w1, dtype=self.dtype)
w2 = ops.cast(w2, dtype=self.dtype)
w = self.lazy_mult([w1, w2], **kwargs)
w = self.layer_exp_dims(w, **kwargs)
xi, xj, xk = self.layer_pos([xyz, ijk])
rij = self.layer_dist([xi, xj], **kwargs)
rik = self.layer_dist([xi, xk], **kwargs)
rjk = self.layer_dist([xj, xk], **kwargs)
fij = self._compute_fc(rij, cutoff=self.cutoff)
fik = self._compute_fc(rik, cutoff=self.cutoff)
fjk = self._compute_fc(rjk, cutoff=self.cutoff)
zi = self.layer_gather_in([z, ijk], **kwargs)
gij = self._compute_gaussian_expansion([rij, zi])
gik = self._compute_gaussian_expansion([rik, zi])
gjk = self._compute_gaussian_expansion([rjk, zi])
vij = self.lazy_sub([xi, xj], **kwargs)
vik = self.lazy_sub([xi, xk], **kwargs)
pow_cos_theta = self._compute_pow_cos_angle_([vij, vik, rij, rik, zi])
rep = self.lazy_mult([pow_cos_theta, gij, gik, gjk, fij, fik, fjk, w], **kwargs)
pool_ang = self.pool_sum([xyz, rep, ijk], **kwargs)
return self._compute_pow_scale([pool_ang, z])
def get_config(self):
config = super(wACSFAng, self).get_config()
config.update({"eta_mu_lambda_zeta": self.eta_mu_lambda_zeta, "cutoff": self.cutoff,
"use_external_weights": self.use_external_weights,
"add_eps": self.add_eps,
"param_constraint": ks.constraints.serialize(self.param_constraint),
"param_regularizer": ks.regularizers.serialize(self.param_regularizer),
"param_initializer": ks.initializers.serialize(self.param_initializer)
})
return config