-
Notifications
You must be signed in to change notification settings - Fork 12
/
Copy pathtrain_network.py
168 lines (160 loc) · 8.04 KB
/
train_network.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
import argparse
parser = argparse.ArgumentParser(description='Train a neural network to decode a code.',
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog='''\
''')
parser.add_argument('dist', type=int,
help='the distance of the code')
parser.add_argument('out', type=str,
help='the name of the output file (used as a prefix for the log file as well)')
parser.add_argument('--trainset', type=str,
help='the name of the training set file (generated by `generate_training_data.py`); if not specified --onthefly is assumed')
parser.add_argument('--onthefly', type=int, nargs=2, default=[2000000, 50000],
help='generate the training set on the fly, specify training and validation size (default: %(default)s)')
parser.add_argument('--prob', type=float, default=0.9,
help='the probability of no error on the physical qubit when generating training data (considered only if --onthefly is present) (default: %(default)s)')
parser.add_argument('--load', type=str, default='',
help='the file from which to load a pretrained model weights (optional, requires correct hyperparameters)')
parser.add_argument('--eval', action='store_true',
help='if present, calculate the fraction of successful corrections based on sampling the NN using the validation set')
parser.add_argument('--giveup', type=int, default=1000000,
help='after how many samples to give up decoding a given test error vector (considered only if --eval is present) (default: %(default)s)')
parser.add_argument('--batch', type=int, default=512,
help='the batch size (default: %(default)s)')
parser.add_argument('--epochs', type=int, default=20,
help='the number of epochs (default: %(default)s)')
parser.add_argument('--learningrate', type=float, default=0.002,
help='the learning rate (default: %(default)s)')
parser.add_argument('--hact', type=str, default='tanh',
help='the activation for hidden layers (default: %(default)s)')
parser.add_argument('--act', type=str, default='sigmoid',
help='the activation for the output layer (default: %(default)s)')
parser.add_argument('--loss', type=str, default='binary_crossentropy',
help='the loss to be optimized (default: %(default)s)')
parser.add_argument('--layers', type=float, default=[4, 4, 4], nargs='+',
help='the list of sizes of the hidden layers (as a factor of the output layer) (default: %(default)s)')
parser.add_argument('--normcentererr', action='store_true',
help='if present, zero-center and normalize the error training and evaluation data (should be done in advance for pregenerated data)')
parser.add_argument('--normcenterstab', action='store_true',
help='if present, zero-center and normalize the stabilizer training and evaluation data (should be done in advance for pregenerated data)')
parser.add_argument('--batchnorm', type=float, default=0,
help='if nonzero, batchnormalize each layer with the given momentum (default: %(default)s)')
parser.add_argument('--Zstab', action='store_true',
help='if present, include the Z stabilizer in the neural network')
parser.add_argument('--Xstab', action='store_true',
help='if present, include the X stabilizer in the neural network')
args = parser.parse_args()
print(args)
from neural import create_model, data_generator, do_normcenterstab, undo_normcentererr, smart_sample
from codes import ToricCode
import numpy as np
import tqdm
if args.trainset:
f = np.load(args.trainset)
x_test = []
y_test = []
if args.Zstab:
x_test.append(f['arr_4'])
y_test.append(f['arr_5'])
if args.Xstab:
x_test.append(f['arr_6'])
y_test.append(f['arr_7'])
x_test = np.hstack(x_test)
y_test = np.hstack(y_test)
model = create_model(L=args.dist,
hidden_sizes=args.layers,
hidden_act=args.hact,
act=args.act,
loss=args.loss,
Z=args.Zstab, X=args.Xstab,
learning_rate=args.learningrate,
normcentererr_p=args.prob if args.normcentererr else None,
batchnorm=args.batchnorm
)
L = args.dist
code = ToricCode(L)
out_dimZ = 2*L**2 * args.Zstab
out_dimX = 2*L**2 * args.Xstab
in_dim = L**2 * (args.Xstab+args.Zstab)
H = code.H(args.Zstab,args.Xstab)
if args.load:
model.load_weights(args.load)
if args.epochs:
if args.trainset:
raise NotImplementedError("This is still using the OLD keras API. Update it!")
x_train = []
y_train = []
if args.Zstab:
x_train.append(f['arr_0'])
y_train.append(f['arr_1'])
if args.Xstab:
x_train.append(f['arr_2'])
y_train.append(f['arr_3'])
x_train = np.hstack(x_train)
y_train = np.hstack(y_train)
hist = model.fit(x_train, y_train,
nb_epoch=args.epochs,
batch_size=args.batch,
validation_data=(x_test, y_test)
)
else:
dat = data_generator(H, out_dimZ, out_dimX, in_dim, args.prob, args.batch,
normcenterstab=args.normcenterstab, normcentererr=args.normcentererr)
val = data_generator(H, out_dimZ, out_dimX, in_dim, args.prob, args.batch,
normcenterstab=args.normcenterstab, normcentererr=args.normcentererr)
hist = model.fit_generator(dat, args.onthefly[0]//args.batch, args.epochs,
validation_data=val, validation_steps=args.onthefly[1]//args.batch)
model.save_weights(args.out)
with open(args.out+'.log', 'w') as f:
f.write(str((hist.params, hist.history)))
if args.eval:
E = ToricCode(L).E(args.Zstab,args.Xstab)
both = args.Zstab and args.Xstab
if both:
Hz = ToricCode(L).H(True, False)
Hx = ToricCode(L).H(False, True)
outlen = 2*L**2*(args.Zstab+args.Xstab)
inlen = L**2*(args.Zstab+args.Xstab)
c = cz = cx = 0
if args.trainset:
stabflipgen = zip(x_test, y_test)
size = len(y_test)
else:
size = args.onthefly[1]
stabflipgen = data_generator(H, out_dimZ, out_dimX, in_dim, args.prob, batch_size=1, size=size)
full_log = np.zeros((size, E.shape[0]+args.Zstab+args.Xstab), dtype=int)
for i, (stab, flips) in tqdm.tqdm(enumerate(stabflipgen), total=size):
if args.normcenterstab:
stab_normed = do_normcenterstab(stab, args.prob)
pred = model.predict(stab_normed).ravel()
else:
pred = model.predict(stab).ravel()
if args.normcentererr:
pred = undo_normcentererr(pred, args.prob)
stab = stab.ravel()
flips = flips.ravel()
sample = pred>np.random.uniform(size=outlen)
if both:
attemptsZ = smart_sample(Hz, stab[:inlen//2], pred[:outlen//2], sample[:outlen//2], args.giveup)
attemptsX = smart_sample(Hx, stab[inlen//2:], pred[outlen//2:], sample[outlen//2:], args.giveup)
else:
attempts = smart_sample(H, stab, pred, sample, args.giveup)
errors = E.dot((sample+flips)%2)%2
if np.any(errors) or np.any(stab!=H.dot(sample)%2):
c += 1
if both:
cz += np.any(errors[:len(errors)//2])
cx += np.any(errors[len(errors)//2:])
if both:
full_log[i,:-2] = errors
full_log[i,-2] = attemptsZ
full_log[i,-1] = attemptsX
else:
full_log[i,:-1] = errors
full_log[i,-1] = attempts
with open(args.out+'.eval', 'w') as f:
if both:
f.write(str(((1-c/size),(1-cz/size),(1-cx/size))))
else:
f.write(str(((1-c/size),)))
np.savetxt(args.out+'.eval.log', full_log, fmt='%d')