-
Notifications
You must be signed in to change notification settings - Fork 31
/
grads_utils.py
196 lines (154 loc) · 6.75 KB
/
grads_utils.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
# Copyright 2019 PIQuIL - All Rights Reserved.
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
# http://www.apache.org/licenses/LICENSE-2.0
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
from qucumber.utils import cplx
from qucumber.utils import training_statistics as ts
class PosGradsUtils:
def __init__(self, nn_state):
self.nn_state = nn_state
def compute_numerical_kl(self, target_psi, vis):
return ts.KL(self.nn_state, target_psi, vis)
def compute_numerical_NLL(self, data, vis):
return ts.NLL(self.nn_state, data, vis)
def algorithmic_gradKL(self, target_psi, vis, **kwargs):
Z = self.nn_state.compute_normalization(vis)
grad_KL = torch.zeros(
self.nn_state.rbm_am.num_pars,
dtype=torch.double,
device=self.nn_state.device,
)
for i in range(len(vis)):
grad_KL += ((target_psi[0, i]) ** 2) * self.nn_state.gradient(vis[i])
grad_KL -= self.nn_state.probability(vis[i], Z) * self.nn_state.gradient(
vis[i]
)
return [grad_KL]
def algorithmic_gradNLL(self, data, k, **kwargs):
return self.nn_state.compute_batch_gradients(k, data, data)
def numeric_gradKL(self, target_psi, param, vis, eps, **kwargs):
num_gradKL = []
for i in range(len(param)):
param[i] += eps
KL_p = self.compute_numerical_kl(target_psi, vis)
param[i] -= 2 * eps
KL_m = self.compute_numerical_kl(target_psi, vis)
param[i] += eps
num_gradKL.append((KL_p - KL_m) / (2 * eps))
return torch.tensor(num_gradKL, dtype=torch.double).to(param)
def numeric_gradNLL(self, param, data, vis, eps, **kwargs):
num_gradNLL = []
for i in range(len(param)):
param[i] += eps
NLL_p = self.compute_numerical_NLL(data, vis)
param[i] -= 2 * eps
NLL_m = self.compute_numerical_NLL(data, vis)
param[i] += eps
num_gradNLL.append((NLL_p - NLL_m) / (2 * eps))
return torch.tensor(num_gradNLL, dtype=torch.double).to(param)
class ComplexGradsUtils:
def __init__(self, nn_state):
self.nn_state = nn_state
def load_target_psi(self, bases, psi_data):
if isinstance(psi_data, torch.Tensor):
psi_data = psi_data.clone().detach().to(dtype=torch.double)
else:
psi_data = torch.tensor(psi_data, dtype=torch.double)
psi_dict = {}
D = int(psi_data.shape[1] / float(len(bases)))
for b in range(len(bases)):
psi = torch.zeros(2, D, dtype=torch.double)
psi[0, ...] = psi_data[0, b * D : (b + 1) * D]
psi[1, ...] = psi_data[1, b * D : (b + 1) * D]
psi_dict[bases[b]] = psi
return psi_dict
def transform_bases(self, bases_data):
bases = []
for i in range(len(bases_data)):
tmp = ""
for j in range(len(bases_data[i])):
if bases_data[i][j] != " ":
tmp += bases_data[i][j]
bases.append(tmp)
return bases
def rotate_psi(self, basis, unitary_dict, vis):
return ts.rotate_psi(self.nn_state, basis, vis, unitary_dict)
def compute_numerical_NLL(self, data_samples, data_bases, vis):
return ts.NLL(self.nn_state, data_samples, vis, bases=data_bases)
def compute_numerical_kl(self, psi_dict, vis, bases):
return ts.KL(self.nn_state, psi_dict, vis, bases=bases)
def algorithmic_gradNLL(self, data_samples, data_bases, k, **kwargs):
return self.nn_state.compute_batch_gradients(
k, data_samples, data_samples, data_bases
)
def numeric_gradNLL(self, data_samples, data_bases, param, vis, eps, **kwargs):
num_gradNLL = []
for i in range(len(param)):
param[i] += eps
NLL_p = self.compute_numerical_NLL(data_samples, data_bases, vis)
param[i] -= 2 * eps
NLL_m = self.compute_numerical_NLL(data_samples, data_bases, vis)
param[i] += eps
num_gradNLL.append((NLL_p - NLL_m) / (2 * eps))
return torch.tensor(num_gradNLL, dtype=torch.double).to(param)
def numeric_gradKL(self, param, psi_dict, vis, bases, eps, **kwargs):
num_gradKL = []
for i in range(len(param)):
param[i] += eps
KL_p = self.compute_numerical_kl(psi_dict, vis, bases)
param[i] -= 2 * eps
KL_m = self.compute_numerical_kl(psi_dict, vis, bases)
param[i] += eps
num_gradKL.append((KL_p - KL_m) / (2 * eps))
return torch.tensor(num_gradKL).to(param)
def algorithmic_gradKL(self, psi_dict, vis, unitary_dict, bases, **kwargs):
grad_KL = [
torch.zeros(
self.nn_state.rbm_am.num_pars,
dtype=torch.double,
device=self.nn_state.device,
),
torch.zeros(
self.nn_state.rbm_ph.num_pars,
dtype=torch.double,
device=self.nn_state.device,
),
]
Z = self.nn_state.compute_normalization(vis).to(device=self.nn_state.device)
for i in range(len(vis)):
grad_KL[0] += (
cplx.norm_sqr(psi_dict[bases[0]][:, i])
* self.nn_state.rbm_am.effective_energy_gradient(vis[i])
/ float(len(bases))
)
grad_KL[0] -= (
self.nn_state.probability(vis[i], Z)
* self.nn_state.rbm_am.effective_energy_gradient(vis[i])
/ float(len(bases))
)
for b in range(1, len(bases)):
for i in range(len(vis)):
rotated_grad = self.nn_state.gradient(bases[b], vis[i])
grad_KL[0] += (
cplx.norm_sqr(psi_dict[bases[b]][:, i])
* rotated_grad[0]
/ float(len(bases))
)
grad_KL[1] += (
cplx.norm_sqr(psi_dict[bases[b]][:, i])
* rotated_grad[1]
/ float(len(bases))
)
grad_KL[0] -= (
self.nn_state.probability(vis[i], Z)
* self.nn_state.rbm_am.effective_energy_gradient(vis[i])
/ float(len(bases))
)
return grad_KL