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 | class MolGNNForce:
def __init__(self,
G,
n_layers=(3, 2),
sizes=[(40, 20, 20), (20, 10)],
nn=1,
sigma=162.13039087945623,
mu=117.41975505778706,
seed=12345):
""" Constructor for MolGNNForce
Parameters
----------
G: TopGraph object
The topological graph object, created using dmff.sgnn.graph.TopGraph
n_layers: int tuple, optional
Number of hidden layers before and after message passing
default = (3, 2)
sizes: [tuple, tuple], optional
sizes (numbers of hidden neurons) of the network before and after message passing
default = [(40, 20, 20), (20, 10)]
nn: int, optional
size of the subgraphs, i.e., how many neighbors to include around the central bond
default = 1
sigma: float, optional
final scaling factor of the energy.
default = 162.13039087945623
mu: float, optional
a constant shift
the final total energy would be ${(E_{NN} + \mu) * \sigma}
seed: int: optional
random seed used in network initialization
default = 12345
"""
self.nn = nn
self.G = G
self.G.get_all_subgraphs(nn, typify=True)
self.G.prepare_subgraph_feature_calc()
params = OrderedDict()
key = jax.random.PRNGKey(seed)
params['w'] = jax.random.uniform(key)
self.n_layers = n_layers
self.sizes = sizes
dim_in = G.n_features
initializer = jax.nn.initializers.he_uniform()
for i_nn, n_layers in enumerate(n_layers):
nn_name = 'fc%d' % i_nn
params[nn_name + '.weight'] = []
params[nn_name + '.bias'] = []
for i_layer in range(n_layers):
layer_name = nn_name + '.' + '%d' % i_layer
dim_out = sizes[i_nn][i_layer]
# params[nn_name+'.weight'].append(jnp.array(np.random.random((dim_out, dim_in))))
# params[nn_name+'.bias'].append(jnp.array(np.random.random(dim_out)))
key, subkey = jax.random.split(key)
params[nn_name + '.weight'].append(
initializer(subkey, (dim_out, dim_in)))
params[nn_name + '.bias'].append(jnp.zeros(dim_out))
dim_in = dim_out
key, subkey = jax.random.split(key)
params['fc_final.weight'] = jnp.array(initializer(subkey, (1, dim_in)))
key, subkey = jax.random.split(key)
params['fc_final.bias'] = jax.random.uniform(subkey)
self.params = params
self.sigma = sigma
self.mu = mu
# generate the forward functions
@jit_condition(static_argnums=3)
def forward(positions, box, params, nn):
features = self.G.calc_subgraph_features(positions, box)
@jit_condition(static_argnums=())
@partial(vmap, in_axes=(0, None), out_axes=(0))
@partial(vmap, in_axes=(0, None), out_axes=(0))
def fc0(f_in, params):
f = f_in
for i in range(self.n_layers[0]):
f = jnp.tanh(params['fc0.weight'][i].dot(f) +
params['fc0.bias'][i])
return f
@jit_condition(static_argnums=())
@partial(vmap, in_axes=(0, None), out_axes=(0))
def fc1(f_in, params):
f = f_in
for i in range(self.n_layers[1]):
f = jnp.tanh(params['fc1.weight'][i].dot(f) +
params['fc1.bias'][i])
return f
@jit_condition(static_argnums=())
@partial(vmap, in_axes=(0, None), out_axes=(0))
def fc_final(f_in, params):
return params['fc_final.weight'].dot(
f_in) + params['fc_final.bias']
# @jit_condition(static_argnums=(3))
@partial(vmap, in_axes=(0, 0, None, None), out_axes=(0))
def message_pass(f_in, nb_connect, w, nn):
if nn == 0:
return f_in[0]
elif nn == 1:
nb_connect0 = nb_connect[0:MAX_VALENCE - 1]
nb_connect1 = nb_connect[MAX_VALENCE - 1:2 *
(MAX_VALENCE - 1)]
nb0 = jnp.sum(nb_connect0)
nb1 = jnp.sum(nb_connect1)
f = f_in[0] * (1 - jnp.heaviside(nb0, 0)*w - jnp.heaviside(nb1, 0)*w) + \
w * nb_connect0.dot(f_in[1:MAX_VALENCE, :]) / jnp.piecewise(nb0, [nb0<1e-5, nb0>=1e-5], [lambda x: jnp.array(1e-5), lambda x: x]) + \
w * nb_connect1.dot(f_in[MAX_VALENCE:2*MAX_VALENCE-1, :])/ jnp.piecewise(nb1, [nb1<1e-5, nb1>=1e-5], [lambda x: jnp.array(1e-5), lambda x: x])
return f
features = fc0(features, params)
features = message_pass(features, self.G.nb_connect, params['w'],
self.G.nn)
features = fc1(features, params)
energies = fc_final(features, params)
return self.G.weights.dot(energies)[0] * self.sigma + self.mu
self.forward = partial(forward, nn=self.G.nn)
self.batch_forward = vmap(self.forward,
in_axes=(0, 0, None),
out_axes=(0))
# provide the get_energy function, to be consistent with the other parts of DMFF
self.get_energy = self.forward
return
def load_params(self, ifn):
""" Load the network parameters from saved file
Parameters
----------
ifn: string
the input file name
"""
with open(ifn, 'rb') as ifile:
params = pickle.load(ifile)
for k in params.keys():
params[k] = jnp.array(params[k])
# transform format
keys = list(params.keys())
for i_nn in [0, 1]:
nn_name = 'fc%d' % i_nn
keys_weight = []
keys_bias = []
for k in keys:
if re.search(nn_name + '.[0-9]+.weight', k) is not None:
keys_weight.append(k)
elif re.search(nn_name + '.[0-9]+.bias', k) is not None:
keys_bias.append(k)
if len(keys_weight) != self.n_layers[i_nn] or len(
keys_bias) != self.n_layers[i_nn]:
sys.exit(
'Error while loading GNN params, inconsistent inputs with the GNN structure, check your input!'
)
params['%s.weight' % nn_name] = []
params['%s.bias' % nn_name] = []
for i_layer in range(self.n_layers[i_nn]):
k_w = '%s.%d.weight' % (nn_name, i_layer)
k_b = '%s.%d.bias' % (nn_name, i_layer)
params['%s.weight' % nn_name].append(params.pop(k_w, None))
params['%s.bias' % nn_name].append(params.pop(k_b, None))
# params[nn_name]
self.params = params
return
def save_params(self, ofn):
""" Save the network parameters to a pickle file
Parameters
----------
ofn: string
the output file name
"""
# transform format
params = {}
params['w'] = self.params['w']
params['fc_final.weight'] = self.params['fc_final.weight']
params['fc_final.bias'] = self.params['fc_final.bias']
for i_nn in range(2):
nn_name = 'fc%d' % i_nn
for i_layer in range(self.n_layers[i_nn]):
params[nn_name + '.%d.weight' %
i_layer] = self.params[nn_name + '.weight'][i_layer]
params[nn_name +
'.%d.bias' % i_layer] = self.params[nn_name +
'.bias'][i_layer]
with open(ofn, 'wb') as ofile:
pickle.dump(params, ofile)
return
|