-
Notifications
You must be signed in to change notification settings - Fork 1
/
vqar.py
208 lines (166 loc) · 6.16 KB
/
vqar.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
# Toshiaki Koike-Akino, 2022
# QHack 2022, variational quantum autregressive (VQAR)
# train/test vqc.py given 2D-data
import pennylane as qml
import pennylane.numpy as np
import argparse
import sklearn
from sklearn import model_selection
from tqdm import tqdm
import vqc
# args
def get_args():
parser = argparse.ArgumentParser('vqar')
# general args
parser.add_argument('--verb', action='store_true', help='verbose')
# vqar args
add_args(parser)
return parser.parse_args()
# add args
def add_args(parser):
# simulator args
sim_args(parser)
# optimizer args
opt_args(parser)
# QML args
vqc.vqc_args(parser)
parser.set_defaults(skip=True)
# sim args
def sim_args(parser):
parser = parser.add_argument_group('sim')
parser.add_argument('--split', default=0.9, type=float, help='train/test split ratio')
parser.add_argument('--seed', default=1, type=int, help='random seed')
parser.add_argument('--epoch', default=100, type=int, help='training epoch')
# opt args
def opt_args(parser):
parser = parser.add_argument_group('opt')
parser.add_argument('--opt', default='AdamOptimizer',
choices=['AdamOptimizer', 'AdagradOptimizer', 'GradientDescentOptimizer',
'MomentumOptimizer', 'NesterovMomentumOptimizer', 'RMSPropOptimizer', 'QNGOptimizer',
#'ShotAdaptiveOptimizer', 'LieAlgebraOptimizer', 'RotosolveOptimizer', 'RotoselectOptimizer',
],
help='optimizer. (some are not tested)')
parser.add_argument('--lr', default=0.01, type=float, help='learning rate (stepsize)')
# random seed
def seeding(args, verb=False):
if args.seed > 0: # reseed if seed > 0
if verb: print('# seeding', args.seed)
np.random.seed(args.seed)
else: # random
args.seed = None
# delay line unfolding: x[n], x[n-1], x[n-2], x[n-M], ...
def delay_line(data, memory=1):
X = list()
for k in range(memory + 1):
X.append(np.roll(data, k, axis=0))
X = np.hstack(X) # data:[N, D] -> X:[N, (M+1) D]
X = X[memory:] # truncate rolled-up head
#print(data)
#print(X)
return X
# optimizer selction
def get_opt(args):
opt = getattr(qml, args.opt)(stepsize=args.lr)
return opt
# model
def get_model(args, verb=False):
if args.model == None:
model = vqc.VQC(dev=args.dev, ansatz=args.ansatz, obs=args.obs,
dim=args.dim, layer=args.layer, memory=args.memory,
skip=args.skip, reup=args.reup,
prescale=args.prescale, postscale=args.postscale, scalebias=args.scalebias,
verb=args.verb)
else:
model = vqc.load_model(fname=args.model, verb=verb)
if verb: print('# model:', model)
return model
# test
def testing(model, xtest, ytest):
loss_ave = 0
for x, y in tqdm(zip(xtest, ytest), leave=True, desc='test'):
# predict
yhat = model(x)
# MSE loss
loss = np.mean((y - yhat)**2)
# loss average
loss_ave += loss.item() / len(ytest)
return loss_ave
# main process
def main(args, data):
if args.verb: print('#', args)
# seed
seeding(args, verb=args.verb)
# data shape [N, D]
args.sample, args.dim = data.shape # orignal data shape
# unfold
data = delay_line(data, args.memory)
if args.verb: print('# unfolded', data.shape)
# train/test split
train, test = sklearn.model_selection.train_test_split(data, train_size=args.split, shuffle=False)
if args.verb: print('# train/test split', args.split, train.shape, test.shape)
# model
model = get_model(args, verb=args.verb)
#model.draw()
# optimizer
opt = get_opt(args)
if args.verb: print('# optimizer', opt)
# cost
def cost(weights, scales, x, y):
model.weights = weights
model.scales = scales
yhat = model(x)
#yhat = x * np.sum(weights)
loss = np.mean((y - yhat)**2)
#print('x, y, yhat, loss:', x._value, y._value, yhat._value, loss._value)
#print(weights._value)
return loss
# initial weights
weights = model.weights
if args.verb: print('#weights:', weights.shape, weights)
scales = model.scales
if args.verb: print('#scales:', scales.shape, scales)
# train loop
for epoch in tqdm(range(args.epoch), leave=True, desc='epoch'):
# permutation index of train data
perm = np.random.permutation(len(train))
#if args.verb: print('# perm', perm)
# target: y, input: x
ytrain, xtrain = train[perm, :args.dim], train[perm, args.dim:]
ytest, xtest = test[:, :args.dim], test[:, args.dim:]
#if args.verb: print('# yx-train yx-test', ytrain.shape, xtrain.shape, ytest.shape, xtest.shape)
# batch loop
loss_ave = 0
for x, y in tqdm(zip(xtrain, ytrain), leave=True, desc='train'):
#print('pre-weights', model.weights.shape)
params, loss = opt.step_and_cost(cost, weights, scales, x, y)
#if args.verb: print('step:', params[0].shape, params[1].shape, params[2].shape, params[3].shape, loss)
# update
weights = params[0]
scales = params[1]
model.weights = weights
model.scales = scales
# loss average
loss_ave += loss.item() / len(perm)
#print(model.weights[0])
# test
loss_test = testing(model, xtest, ytest)
# results
results = (loss_ave, loss_test, # MSE
loss_ave**0.5, loss_test**0.5, # RMSE
10 * np.log10(loss_ave), 10 * np.log10(loss_test)) # MSE in dB
print(epoch, *results)
# save models
vqc.save_model(model, fname=f'model{args.memory}.p', verb=args.verb)
model.draw()
# optimized model
return model, results
# example use
if __name__ == '__main__':
# args
args = get_args()
# toy data: Wiener process
data = np.random.randn(100, 2) # 100-sample of 2-dim data
data = np.cumsum(data, axis=0)
# main train/test
model, results = main(args, data)
print(*results)