cifar

config/cifar10/base.py

@@ -0,0 +1,17 @@
+import chainer
+from src.model.wide_resnet_new import WideResNet
+from src.dataset.cifar import cifar
+from src.preprocess.preprocess_svhn import PreprocessSVHN
+from src.extension.learning_rate_scheduler import LearningRateScheduler
+from src.extension.learning_rate_scheduler import ExponentialSchedule
+from src.hook.power_iteration import PowerIteration
+
+batchsize = 128
+dataset = cifar()
+epoch = 160
+preprocess = PreprocessSVHN()
+predictor = WideResNet(k=4, n_layer=16, drop=.4)
+lr = 1e-2
+optimizer = chainer.optimizers.NesterovAG(lr)
+extension = [(LearningRateScheduler(ExponentialSchedule(.1, (80, 120))), (1, 'iteration'))]
+hook = [PowerIteration()]

config/cifar10/lmt_001.py

@@ -0,0 +1,10 @@
+from config.cifar10.base import *
+from src.model.classifier import LMTraining
+from src.extension.c_scheduler import CScheduler
+from src.extension.c_scheduler import GoyalSchedule
+
+mode = ['lmt', 'lmt-fc']
+start_c = 1e-5
+end_c = 1e-2
+model = LMTraining(predictor, preprocess, c=start_c)
+extension += [(CScheduler(GoyalSchedule(start_c, end_c, 5)), (1, 'iteration'))]

src/dataset/cifar.py

@@ -0,0 +1,22 @@
+import chainer
+from chainer.datasets import TupleDataset
+
+means = (0.4914, 0.4822, 0.4465)
+sds = (0.2023, 0.1994, 0.2010)
+
+def normalize(X):
+    for i in range(3):
+        X[:,i,:,:] -= means[i]
+        X[:,i,:,:] /= sds[i]
+    return X
+
+def transform(dataset):
+    X = dataset._datasets[0]
+    y = dataset._datasets[1]
+    return TupleDataset(normalize(X), y)
+
+def cifar():
+    cifar10 = chainer.datasets.get_cifar10()
+    train = transform(cifar10[0])
+    test = transform(cifar10[1])
+    return (train, test)