-
Notifications
You must be signed in to change notification settings - Fork 5
/
MPSNew.py
285 lines (233 loc) · 11.2 KB
/
MPSNew.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
import tensorflow as tf
import numpy as np
from collections import namedtuple
#TODO: REALLY needs some documentation
class MPS(object):
'''
Class variables:
input_size: int
d_matrix: int
d_feature: int
d_output: int
nodes: tf.TensorArray
'''
def __init__(self, d_matrix, d_feature, d_output, input_size):
# structure parameters
self.input_size = input_size
self.d_matrix = d_matrix
self.d_feature = d_feature
self.d_output = d_output
# Initialise the nodes, input and output
self._setup_nodes()
def predict(self, feature):
'''
feature must be numpy array of dtype float32
'''
phi = tf.placeholder(tf.float32, shape=[self.input_size, None, self.d_feature])
f = self._predict(phi)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print(sess.run(f, {phi: feature}))
# ================
# hidden functions
# ================
def _setup_nodes(self):
self.nodes = tf.TensorArray(tf.float32, size = 0, dynamic_size= True,
clear_after_read= False, infer_shape= False)
# First node
self.nodes = self.nodes.write(0, self._make_random_normal([self.d_feature, self.d_matrix]))
# The Second node with output leg attached
self.nodes = self.nodes.write(1, self._make_random_normal([self.d_output, self.d_feature, self.d_matrix, self.d_matrix]))
# The rest of the matrix nodes
for i in range(self.input_size - 3):
self.nodes = self.nodes.write(i+2, self._make_random_normal([self.d_feature, self.d_matrix, self.d_matrix]))
# Last node
self.nodes = self.nodes.write(self.input_size-1, self._make_random_normal([self.d_feature, self.d_matrix]))
def _make_random_normal(self, shape, mean=0, stddev=1):
return tf.Variable(tf.random_normal(shape, mean=mean, stddev=stddev))
def _predict(self, phi):
# Read in phi
self.phi = phi
# Read in the nodes
node1 = self.nodes.read(0)
node1.set_shape([self.d_feature, None])
node2 = self.nodes.read(1)
node2.set_shape([self.d_output, self.d_feature, None, None])
nodelast = self.nodes.read(self.input_size-1)
nodelast.set_shape([self.d_feature, None])
# Calculate C1
C1 = tf.einsum('ni,tn->ti', node1, phi[0])
contracted_node2 = tf.einsum('lnij,tn->tlij', node2, phi[1])
C1 = tf.einsum('ti,tlij->tlj', C1, contracted_node2)
# Calculate C2
C2 = tf.einsum('mi,tm->ti', nodelast, phi[self.input_size-1])
#counter = tf.Variable(2, dtype=tf.int32)
counter = 2
cond = lambda c, b: tf.less(c, self.input_size-1)
_, C1 = tf.while_loop(cond=cond, body=self._chain_multiply, loop_vars=[counter, C1],
shape_invariants=[tf.TensorShape([]), tf.TensorShape([None, self.d_output, None])])
f = tf.einsum('tli,ti->tl', C1, C2)
return f
def _chain_multiply(self, counter, C1):
node = self.nodes.read(counter)
node.set_shape([self.d_feature, None, None])
input_leg = self.phi[counter]
contracted_node = tf.einsum('mij,tm->tij', node, input_leg)
C1 = tf.einsum('tli,tij->tlj', C1, contracted_node)
counter = counter + 1
return [counter, C1]
class MPSOptimizer(object):
def __init__(self, MPSNetwork, m, grad_func, rate_of_change = None):
self.MPS = MPSNetwork
if rate_of_change is None:
self.rate_of_change = 0.1
else:
self.rate_of_change = rate_of_change
self.m = m
self.grad_func = grad_func
self._phi = tf.placeholder(tf.float32, shape=[input_size, None, self.MPS.d_feature])
self._delta = tf.placeholder(tf.float32, shape=[None, self.MPS.d_output])
self._setup_optimization()
self._setup_training_graph()
def train(self, phi, delta):
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
end_results = sess.run(self.C1, {self._phi: phi, self._delta: delta})
print(end_results)
writer = tf.summary.FileWriter("output", sess.graph)
writer.close()
#self.MPS.nodes = end_results[-1]
def _setup_optimization(self):
phi = self._phi
nodes = self.MPS.nodes
n1 = nodes.read(0)
n1.set_shape([None,None])
nlast = nodes.read(nodes.size() - 1)
nlast.set_shape([None,None])
C2s = tf.TensorArray(tf.float32, size=self.MPS.input_size-3, infer_shape=False)
self.C1 = tf.einsum('ni,tn->ti', n1, phi[0])
C2 = tf.einsum('mi,tm->ti', nlast, phi[-1])
C2s = C2s.write(self.MPS.input_size-4, C2)
cond = lambda counter,b,c: tf.less(counter, self.MPS.input_size-4)
_, _, self.C2s = tf.while_loop(cond=cond, body=self._find_C2, loop_vars=[0, C2, C2s],
shape_invariants=[tf.TensorShape([]), tf.TensorShape([None, None]),
tf.TensorShape(None)])
def _setup_training_graph(self):
"""
Sets up graph needed to train the MPS, only for going to the right once.
:return: nothing
"""
# First sweep
n1 = self.MPS.nodes.read(1)
n1.set_shape([None, None, None, None])
C1s, n1 = self._one_sweep(n1, self.C1, self.C2s)
# Second sweep
C2 = self.C2s.read(self.MPS.input_size-4)
C2.set_shape([None, None])
self.C2s, _ = self._one_sweep(n1, C2, C1s)
def _one_sweep(self, n1, C1, C2s):
C1s = tf.TensorArray(tf.float32, size=self.MPS.input_size-3, infer_shape=False)
C1s = C1s.write(0, C1)
updated_nodes = self._make_new_nodes()
wrapped = [0, C1, C1s, C2s, updated_nodes, n1]
cond = lambda counter, b, c, d, e, f: tf.less(counter, self.MPS.input_size - 3)
_, C1, C1s, C2s, self.MPS.nodes, n1 = tf.while_loop(cond=cond, body=self._update, loop_vars=wrapped,
parallel_iterations = 1,
shape_invariants=[tf.TensorShape([]),
tf.TensorShape([None, None]),
tf.TensorShape(None),
tf.TensorShape(None),
tf.TensorShape(None),
tf.TensorShape([None, None, None, None])])
n1 = tf.transpose(n1, perm=[0, 1, 3, 2])
self.MPS.nodes = self.MPS.nodes.write(1, n1)
return (C1s, n1)
def _make_new_nodes(self):
new_nodes = tf.TensorArray(tf.float32, size=self.MPS.input_size, dynamic_size=True, infer_shape=False)
new_nodes = new_nodes.write(0, self.MPS.nodes.read(self.MPS.input_size-1))
new_nodes = new_nodes.write(self.MPS.input_size, self.MPS.nodes.read(0))
return new_nodes
def _find_C2(self, counter, prev_C2, C2s):
loc2 = self.MPS.input_size - 2 - counter
node2 = self.MPS.nodes.read(loc2)
node2.set_shape([self.MPS.d_feature, None, None])
contracted_node2 = tf.einsum('mij,tm->tij', node2, self._phi[loc2]) # CHECK einsum
updated_counter = counter + 1
new_C2 = tf.einsum('tij,tj->ti', contracted_node2, prev_C2)
C2s = C2s.write(self.MPS.input_size-5-counter, new_C2)
return [updated_counter, new_C2, C2s]
def _update(self, counter, C1, C1s, C2s, updated_nodes, previous_node):
# Read in the notes
n1 = previous_node
n2 = self.MPS.nodes.read(counter+2)
n1.set_shape([self.MPS.d_output, self.MPS.d_feature, None, None])
n2.set_shape([self.MPS.d_feature, None, None])
# Calculate the bond
bond = tf.einsum('lmij,njk->lmnik', n1, n2)
# Calculate the C matrix
C2 = C2s.read(counter)
C2.set_shape([None,None])
C = tf.einsum('ti,tk,tm,tn->tmnik', C1, C2, self._phi[counter], self._phi[counter+1])
# Update the bond
f = tf.einsum('lmnik,tmnik->tl', bond, C)
gradient = tf.einsum('tl,tmnik->lmnik', self._delta - f, C)
delta_bond = self.rate_of_change * gradient
updated_bond = tf.add(bond, delta_bond)
# Decompose the bond
aj, aj1 = self._bond_decomposition(updated_bond, m)
# Transpose the values and add to the new variables
aj = tf.transpose(aj, perm=[0, 2, 1])
updated_nodes = updated_nodes.write(self.MPS.input_size-3-counter, aj)
contracted_aj = tf.einsum('mij,tm->tij', aj, self._phi[counter])
C1 = tf.einsum('tij,ti->tj', contracted_aj, C1)
C1s = C1s.write(counter+1, C1)
updated_counter = counter + 1
return [updated_counter, C1, C1s, C2s, updated_nodes, aj1]
def _bond_decomposition(self, bond, m):
"""
Decomposes bond, so that the next step can be done.
:param bond:
:param m:
:return:
"""
bond_reshaped = tf.transpose(bond, perm=[3, 1, 2, 0, 4])
dims = tf.shape(bond_reshaped)
l_dim = dims[0] * dims[1]
r_dim = dims[2] * dims[3] * dims[4]
bond_flattened = tf.reshape(bond_reshaped, [l_dim, r_dim])
s, u, v = tf.svd(bond_flattened)
# filtered_s = s[:,-1]
filtered_s = s
s_size = tf.size(filtered_s)
s_im = tf.reshape(tf.diag(filtered_s), [s_size, s_size, 1])
v_im = tf.reshape(v, [s_size, r_dim, 1])
u_im = tf.reshape(u, [l_dim, s_size, 1])
s_im_cropped = tf.image.resize_image_with_crop_or_pad(s_im, m, m)
v_im_cropped = tf.image.resize_image_with_crop_or_pad(v_im, m, r_dim)
u_im_cropped = tf.image.resize_image_with_crop_or_pad(u_im, l_dim, m)
s_mat = tf.reshape(s_im_cropped, [m, m])
v_mat = tf.reshape(v_im_cropped, [m, r_dim])
a_prime_j_mixed = tf.reshape(u_im_cropped, [dims[0], dims[1], m])
sv = tf.matmul(s_mat, v_mat)
a_prime_j1_mixed = tf.reshape(sv, [m, dims[2], dims[3], dims[4]])
a_prime_j = tf.transpose(a_prime_j_mixed, perm=[1, 0, 2])
a_prime_j1 = tf.transpose(a_prime_j1_mixed, perm=[2, 1, 0, 3])
return (a_prime_j, a_prime_j1)
if __name__ == '__main__':
# Model parameters
input_size = 4
d_feature = 2
d_matrix = 5
d_output = 6
rate_of_change = 0.2
batch_size = 10
m = 5
# Make up input and output
phi = np.random.normal(size=(input_size, batch_size, d_feature)).astype(np.float32)
delta = []
for i in range(batch_size):
delta.append([0,1,0, 0, 0, 0])
# Initialise the model
network = MPS(d_matrix, d_feature, d_output, input_size)
optimizer = MPSOptimizer(network, m, None, rate_of_change = rate_of_change)
optimizer.train(phi, delta)