こちらのページはVGG16ではなく自分で何層かのCNNを作成して適用してみた結果についてまとめたものです。
モデルの作成
DeZeroでモデルを作成するには、Modelクラスを継承したクラスで下記を設定する
- __init__()でレイヤーのインスタンスを生成
- forward()でレイヤーインスタンスに入力値を渡し、活性化関数に通す
作成したモデルのインスタンスを生成し予測値を求める
テストモデル
前回まではVGG16を使用してきたが、手始めに4層のCNNと２層の全結合層で6層のモデルを作成して適用してみる
コード
※Google Colaboratory（ランタイムのタイプ= GPU)
※エポック数 8
import glob
from PIL import Image
import math
import numpy as np
import matplotlib.pyplot as plt
import dezero
import dezero.functions as F
import dezero.layers as L
from dezero import DataLoader, optimizers, Parameter, test_mode, Dataset, Model
path = "./drive/MyDrive/Colab Notebooks/data/"
train_path_iso = glob.glob(path + 'train/イソ/*')
train_path_wonyon = glob.glob(path + 'train/ウォニョン/*')
train_path_gaul = glob.glob(path + 'train/ガウル/*')
train_path_yujin = glob.glob(path + 'train/ユジン/*')
train_path_lizu = glob.glob(path + 'train/リズ/*')
train_path_rey = glob.glob(path + 'train/レイ/*')
test_path_iso = glob.glob(path + 'test/イソ/*')
test_path_wonyon = glob.glob(path + 'test/ウォニョン/*')
test_path_gaul = glob.glob(path + 'test/ガウル/*')
test_path_yujin = glob.glob(path + 'test/ユジン/*')
test_path_lizu = glob.glob(path + 'test/リズ/*')
test_path_rey = glob.glob(path + 'test/レイ/*')
train_path, test_path = [], []
train_path.extend(train_path_iso)
train_path.extend(train_path_wonyon)
train_path.extend(train_path_gaul)
train_path.extend(train_path_yujin)
train_path.extend(train_path_lizu)
train_path.extend(train_path_rey)
test_path.extend(test_path_iso)
test_path.extend(test_path_wonyon)
test_path.extend(test_path_gaul)
test_path.extend(test_path_yujin)
test_path.extend(test_path_lizu)
test_path.extend(test_path_rey)
_train_label = [0]*len(train_path_iso) + [1]*len(train_path_wonyon) + [2]*len(train_path_gaul) + [3]*len(train_path_yujin) + [4]*len(train_path_lizu) + [5]*len(train_path_rey)
train_label = np.array(_train_label)
_test_label = [0]*len(test_path_iso) + [1]*len(test_path_wonyon) + [2]*len(test_path_gaul) + [3]*len(test_path_yujin) + [4]*len(test_path_lizu) + [5]*len(test_path_rey)
test_label = np.array(_test_label)
class TestModel(Model):
WEIGHTS_PATH = None
def __init__(self, pretrained=False):
super().__init__()
self.conv1 = L.Conv2d(32, kernel_size=3, stride=1, pad=1)
self.conv2 = L.Conv2d(64, kernel_size=3, stride=1, pad=1)
self.conv3 = L.Conv2d(128, kernel_size=3, stride=1, pad=1)
self.conv4 = L.Conv2d(128, kernel_size=3, stride=1, pad=1)
self.fc5 = L.Linear(512)
self.fc6 = L.Linear(6)
if pretrained:
weights_path = ''
self.load_weights(weights_path)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.pooling(x, 2, 2)
x = F.relu(self.conv2(x))
x = F.pooling(x, 2, 2)
x = F.relu(self.conv3(x))
x = F.pooling(x, 2, 2)
x = F.relu(self.conv4(x))
x = F.pooling(x, 2, 2)
x = F.reshape(x, (x.shape[0], -1))
x = F.dropout(F.relu(self.fc5(x)))
x = self.fc6(x)
return x
@staticmethod
def preprocess(image, size=(224, 224), dtype=np.float32):
image = image.convert('RGB')
if size:
image = image.resize(size)
image = np.asarray(image, dtype=dtype)
image = image[:, :, ::-1]
image -= np.array([103.939, 116.779, 123.68], dtype=dtype)
image = image.transpose((2, 0, 1))
return image
class IveDataset(Dataset):
def __init__(self, train=True, transform=TestModel.preprocess, target_transform=None):
super().__init__(train, transform, target_transform)
def prepare(self):
if self.train:
self.data = train_path
self.label = train_label
else:
self.data = test_path
self.label = test_label
def __getitem__(self, index):
img = Image.open(self.data[index])
img = self.transform(img)
return img, int(self.label[index])
def __len__(self):
return len(self.label)
batchsize = 8
train_set = IveDataset(train=True)
train_loader = DataLoader(train_set, batchsize, shuffle=True)
test_set = IveDataset(train=False)
test_loader = DataLoader(test_set, batchsize, shuffle=False)
train_loader.to_gpu()
test_loader.to_gpu()
max_epoch = 10
for epoch in range(max_epoch):
sum_loss, sum_acc = 0, 0
for x, t in train_loader:
y = model(x)
loss = F.softmax_cross_entropy(y, t)
acc = F.accuracy(y, t)
model.cleargrads()
loss.backward()
optimizer.update()
sum_loss += float(loss.data) * len(t)
sum_acc += float(acc.data) * len(t)
print('epoch: {}'.format(epoch+1))
print('train loss: {:.4f}, accuracy: {:.4f}'.format(sum_loss / len(train_set), sum_acc / len(train_set)))
sum_loss, sum_acc = 0, 0
with test_mode():
for x, t in test_loader:
y = model(x)
loss = F.softmax_cross_entropy(y, t)
acc = F.accuracy(y, t)
sum_loss += float(loss.data) * len(t)
sum_acc += float(acc.data) * len(t)
print('test loss: {:.4f}, accuracy: {:.4f}'.format(sum_loss / len(test_set), sum_acc / len(test_set)))
結果
