DeepLearningフレームワーク Chainerではまったこと

PFNが提供するChainerですが、バージョンの移り変わりが激しい（最新情報に追従しているのでうれしいこと）ため、仕様がかわってることがあるので、メモ代わりに残します。結構、気づかないとはまりますね。

dropoutなどを入れてる場合は,Evaluatorの挙動に注意

VGGベースのネットワークで下記のようなものを使う場合、Evaluatorからforwordが呼ばれるので、テスト時にdropoutが有効になってしまうので毎回結果が変わる。

テスト時にdropoutを無効にするにはchainer.configで条件分岐を記載すること。

<https://github.com/chainer/chainer/blob/master/examples/cifar/models/VGG.py>よりネットワーク構造部を引用

<https://chainerfan.jimdo.com/chainer-v3-%E3%83%AA%E3%83%95%E3%82%A1%E3%83%AC%E3%83%B3%E3%82%B9%E6%97%A5%E6%9C%AC%E8%AA%9E%E8%A8%B3/core-functionalities/chainer%E3%81%AE%E8%A8%AD%E5%AE%9A/>

上記にchainer.configで条件分岐する例が記載されている。Example参照

class VGG(chainer.Chain):
	"""A VGG-style network for very small images.
	This model is based on the VGG-style model from
	http://torch.ch/blog/2015/07/30/cifar.html
	which is based on the network architecture from the paper:
	https://arxiv.org/pdf/1409.1556v6.pdf
	This model is intended to be used with either RGB or greyscale input
	images that are of size 32x32 pixels, such as those in the CIFAR10
	and CIFAR100 datasets.
	On CIFAR10, it achieves approximately 89% accuracy on the test set with
	no data augmentation.
	On CIFAR100, it achieves approximately 63% accuracy on the test set with
	no data augmentation.
	Args:
	class_labels (int): The number of class labels.
	"""
	def __init__(self, class_labels=10):
	super(VGG, self).__init__()
	with self.init_scope():
	self.block1_1 = Block(64, 3)
	self.block1_2 = Block(64, 3)
	self.block2_1 = Block(128, 3)
	self.block2_2 = Block(128, 3)
	self.block3_1 = Block(256, 3)
	self.block3_2 = Block(256, 3)
	self.block3_3 = Block(256, 3)
	self.block4_1 = Block(512, 3)
	self.block4_2 = Block(512, 3)
	self.block4_3 = Block(512, 3)
	self.block5_1 = Block(512, 3)
	self.block5_2 = Block(512, 3)
	self.block5_3 = Block(512, 3)
	self.fc1 = L.Linear(None, 512, nobias=True)
	self.bn_fc1 = L.BatchNormalization(512)
	self.fc2 = L.Linear(None, class_labels, nobias=True)
	def forward(self, x):
	# 64 channel blocks:
	h = self.block1_1(x)
	h = F.dropout(h, ratio=0.3)
	h = self.block1_2(h)
	h = F.max_pooling_2d(h, ksize=2, stride=2)
	# 128 channel blocks:
	h = self.block2_1(h)
	h = F.dropout(h, ratio=0.4)
	h = self.block2_2(h)
	h = F.max_pooling_2d(h, ksize=2, stride=2)
	# 256 channel blocks:
	h = self.block3_1(h)
	h = F.dropout(h, ratio=0.4)
	h = self.block3_2(h)
	h = F.dropout(h, ratio=0.4)
	h = self.block3_3(h)
	h = F.max_pooling_2d(h, ksize=2, stride=2)
	# 512 channel blocks:
	h = self.block4_1(h)
	h = F.dropout(h, ratio=0.4)
	h = self.block4_2(h)
	h = F.dropout(h, ratio=0.4)
	h = self.block4_3(h)
	h = F.max_pooling_2d(h, ksize=2, stride=2)
	# 512 channel blocks:
	h = self.block5_1(h)
	h = F.dropout(h, ratio=0.4)
	h = self.block5_2(h)
	h = F.dropout(h, ratio=0.4)
	h = self.block5_3(h)
	h = F.max_pooling_2d(h, ksize=2, stride=2)
	h = F.dropout(h, ratio=0.5)
	h = self.fc1(h)
	h = self.bn_fc1(h)
	h = F.relu(h)
	h = F.dropout(h, ratio=0.5)
	return self.fc2(h)