N = 60000
x_train, x_test = np.split(mnist['data'], [N])
y_train, y_test = np.split(mnist['target'], [N])
N_test = y_test.size
# Prepare multi-layer perceptron model, defined in net.py
if args.net == 'simple':
model = L.Classifier(net.MnistMLP(784, n_units, 10))
if args.gpu >= 0:
cuda.get_device(args.gpu).use()
model.to_gpu()
xp = np if args.gpu < 0 else cuda.cupy
elif args.net == 'parallel':
cuda.check_cuda_available()
model = L.Classifier(net.MnistMLPParallel(784, n_units, 10))
xp = cuda.cupy
# Setup optimizer
optimizer = optimizers.Adam()
optimizer.setup(model)
# Init/Resume
if args.initmodel:
print('Load model from', args.initmodel)
serializers.load_npz(args.initmodel, model)
if args.resume:
print('Load optimizer state from', args.resume)
serializers.load_npz(args.resume, optimizer)
# Learning loop
#temp_1_test = np.where(y_test==1)[0]
#y_test = y_test[temp_1_test]
#x_test = x_test[temp_1_test]
N = x_train.shape[0]
N_test = y_test.size
# Prepare multi-layer perceptron model, defined in net.py
if args.net == 'simple':
model = L.Classifier(net.MnistMLP(61, n_units, 3))
if args.gpu >= 0:
cuda.get_device(args.gpu).use()
model.to_gpu()
xp = np if args.gpu < 0 else cuda.cupy
elif args.net == 'parallel':
cuda.check_cuda_available()
model = L.Classifier(net.MnistMLPParallel(61, n_units, 3))
xp = cuda.cupy
# Setup optimizer
if 'opt' in args:
#Todo can also pass arguments to each optimizer, see https://github.com/mitmul/chainer-cifar10/blob/master/train.py#L62
if args.opt == 'MomentumSGD':
optimizer = optimizers.MomentumSGD(lr=args.lr, momentum=0.9)
elif args.opt == 'AdaGrad':
optimizer = optimizers.AdaGrad(lr=args.lr)
elif args.opt == 'Adam':
optimizer = optimizers.Adam(alpha=args.alpha)
else:
optimizer = optimizers.Adam()
optimizer.setup(model)
def train(trainL, trainA, testL, testA):
trainLoss = []
trainAccuracy = []
testLoss = []
testAccuracy = []
l1w = []
l2w = []
l3w = []
print('load MNIST dataset')
mnist = data.load_mnist_data()
mnist['data'] = mnist['data'].astype(np.float32)
mnist['data'] /= 255
mnist['target'] = mnist['target'].astype(np.int32)
N = 1000
lsizes = [784, 50, 50, 10]
x_train, x_test = np.split(mnist['data'], [N])
y_train, y_test = np.split(mnist['target'], [N])
N_test = y_test.size
# Prepare multi-layer perceptron model, defined in net.py
if args.net == 'simple':
#model = net.MnistMLP(lsizes)
model = net.MnistMLP(layer_sizes=lsizes)
if args.gpu >= 0:
cuda.get_device(args.gpu).use()
model.to_gpu()
xp = np if args.gpu < 0 else cuda.cupy
elif args.net == 'parallel':
cuda.check_cuda_available()
model = L.Classifier(net.MnistMLPParallel(784, n_units, 10))
xp = cuda.cupy
# Setup optimizer
optimizer = optimizers.Adam()
optimizer.setup(model)
# Init/Resume
if args.initmodel:
print('Load model from', args.initmodel)
serializers.load_npz(args.initmodel, model)
if args.resume:
print('Load optimizer state from', args.resume)
serializers.load_npz(args.resume, optimizer)
# Pretrain loop
print("start pretrain")
epo = p_epoch
for j in six.moves.range(1, len(lsizes)):
if j == len(lsizes) - 1:
model.setfinetuning()
print("start finetuning")
epo = n_epoch
for epoch in six.moves.range(1, epo + 1):
print('layer ', j, 'p_epoch ', epoch)
perm = np.random.permutation(N)
sum_accuracy = 0
sum_loss = 0
for i in six.moves.range(0, N, batchsize):
x = chainer.Variable(xp.asarray(x_train[perm[i:i +
batchsize]]))
t = chainer.Variable(xp.asarray(y_train[perm[i:i +
batchsize]]))
optimizer.update(model, x, t, j)
sum_loss += float(model.loss.data) * len(t.data)
if not model.pretrain:
sum_accuracy += float(model.accuracy.data) * len(t.data)
if model.pretrain:
print('Pretrain: train mean loss={}'.format(sum_loss / N))
else:
print('Finetune: train mean loss={}, accuracy={}'.format(
sum_loss / N, sum_accuracy / N))
trainLoss.append(sum_loss / N)
trainAccuracy.append(sum_accuracy / N)
# evaluation
sum_accuracy = 0
sum_loss = 0
model.train = False
for i in six.moves.range(0, N_test, batchsize):
x = chainer.Variable(xp.asarray(x_test[i:i + batchsize]),
volatile='on')
t = chainer.Variable(xp.asarray(y_test[i:i + batchsize]),
volatile='on')
loss = model(x, t, j)
sum_loss += float(loss.data) * len(t.data)
if not model.pretrain:
sum_accuracy += float(model.accuracy.data) * len(t.data)
if model.pretrain:
print('Pretrain: test mean loss={}'.format(sum_loss / N_test))
else:
print('Finetune: test mean loss={}, accuracy={}'.format(
sum_loss / N_test, sum_accuracy / N_test))
testLoss.append(sum_loss / N_test)
testAccuracy.append(sum_accuracy / N_test)
model.train = True
# Save the model and the optimizer
savecsv(trainLoss, trainL)
savecsv(trainAccuracy, trainA)
savecsv(testLoss, testL)
savecsv(testAccuracy, testA)
print('save the model')
serializers.save_npz('mlp.model', model)
print('save the optimizer')
serializers.save_npz('mlp.state', optimizer)
# Split dataset to data_train(80%) and data_test(20%)
N = 320
data_train, data_test = np.split(dataset, [N])
target_train, target_test = np.split(targetset, [N])
N_test = target_test.size
# Prepare multi-layer perceptron model, defined in net.py
if args.net == 'simple':
model = L.Classifier(net.MnistMLP(16128, n_units, 2))
if args.gpu >= 0:
cuda.get_device(args.gpu).use()
model.to_gpu()
xp = np if args.gpu < 0 else cuda.cupy
elif args.net == 'parallel':
cuda.check_cuda_available()
model = L.Classifier(net.MnistMLPParallel(16128, n_units, 2))
xp = cuda.cupy
# Setup optimizer
optimizer = optimizers.Adam()
optimizer.setup(model)
# Init/Resume
if args.initmodel:
print('Load model from', args.initmodel)
serializers.load_hdf5(args.initmodel, model)
if args.resume:
print('Load optimizer state from', args.resume)
serializers.load_hdf5(args.resume, optimizer)
# Learning loop