Files
MP-SPDZ/Programs/Source/mnist_full_D.mpc
Marcel Keller 2813c0ef0f Maintenance.
2023-08-14 18:29:46 +10:00

110 lines
2.4 KiB
Plaintext

# this trains network D from SecureNN
# see https://github.com/csiro-mlai/mnist-mpc for data preparation
import ml
import math
import re
import util
program.options_from_args()
sfix.set_precision_from_args(program, True)
MultiArray.disable_index_checks()
if 'profile' in program.args:
print('Compiling for profiling')
N = 1000
n_test = 100
elif 'debug' in program.args:
N = 100
n_test = 100
elif 'debug1000' in program.args:
N = 1000
n_test = 1000
elif 'debug5000' in program.args:
N = 5000
n_test = 5000
else:
N = 60000
n_test = 10000
n_examples = N
n_features = 28 ** 2
try:
n_epochs = int(program.args[1])
except:
n_epochs = 100
try:
batch_size = int(program.args[2])
except:
batch_size = min(N, 128)
assert batch_size <= N
ml.Layer.back_batch_size = batch_size
try:
ml.set_n_threads(int(program.args[3]))
except:
pass
if program.options.ring:
assert sfix.f * 4 == int(program.options.ring)
if 'stride1' in program.args:
stride = (1, 1)
else:
stride = (2, 2)
if 'valid' in program.args:
padding = 'VALID'
inner_dim = (28 - 4) // stride[0]
else:
padding = 'SAME'
inner_dim = 28 // stride[0]
layers = [
ml.FixConv2d([N, 28, 28, 1], (5, 5, 5, 1), (5,),
[N, inner_dim, inner_dim, 5], stride, padding),
ml.Relu([N, inner_dim, inner_dim, 5]),
]
if 'maxpool' in program.args:
layers += [ml.MaxPool((N, inner_dim, inner_dim, 5))]
inner_dim //= 2
n_inner = inner_dim ** 2 * 5
dropout = 'dropout' in program.args
if '1dense' in program.args:
if dropout:
layers += [ml.Dropout(N, n_inner)]
layers += [ml.Dense(N, n_inner, 10),]
elif '2dense' in program.args:
if dropout:
layers += [ml.Dropout(N, n_inner)]
layers += [
ml.Dense(N, n_inner, 100),
ml.Relu([N, 100]),
ml.Dense(N, 100, 10),
]
if dropout or 'dropout1' in program.args:
layers.insert(-1, ml.Dropout(N, 100))
else:
raise Exception('need to specify number of dense layers')
layers += [ml.MultiOutput(N, 10)]
Y = sint.Matrix(n_test, 10)
X = sfix.Matrix(n_test, n_features)
if not ('no_acc' in program.args and 'no_loss' in program.args):
layers[-1].Y.input_from(0)
layers[0].X.input_from(0)
Y.input_from(0)
X.input_from(0)
optim = ml.Optimizer.from_args(program, layers)
optim.run_by_args(program, n_epochs, batch_size, X, Y)