備忘録を兼ねて、ChainerでNN、CNN、RNNのモデルを作成し、MNISTのディープラーニングのスクリプトを作成します。
事前準備
Google ColaboratoryへのChainerの導入
GPUを使用できるよう、下記のコマンドで必要なモジュールをインストールします。
!apt-get install -y -qq libcusparse8.0 libnvrtc8.0 libnvtoolsext1
!ln -snf /usr/lib/x86_64-linux-gnu/libnvrtc-builtins.so.8.0 /usr/lib/x86_64-linux-gnu/libnvrtc-builtins.so
!pip install cupy-cuda80==4.0.0b4
!pip install chainer==4.0.0b4
必要なモジュールの読み込み
import numpy as np
import chainer
from chainer import cuda, Function, gradient_check, report, training, utils, Variable
from chainer import datasets, iterators, optimizers, serializers
from chainer import Link, Chain, ChainList
import chainer.functions as F
import chainer.links as L
from chainer.datasets import tuple_dataset
from chainer import training
from chainer.training import extensions
MNISTの読み込み
train, test = datasets.get_mnist(ndim=3)
データを確認します。
%matplotlib inline
import matplotlib.pyplot as plt
plt.imshow(np.array(train[0][0][0], dtype=np.float32), cmap='gray')
print('この画像データのラベルは{:0d}です。'.format(train[0][1]))
この画像データのラベルは5です。
ニューラルネットワークでの学習
モデルの定義
class NN(Chain):
def __init__(self, n_in, n_units, n_out):
super(NN, self).__init__()
with self.init_scope():
self.l1 = L.Linear(n_in, n_units)
self.l2 = L.Linear(n_units, n_units)
self.l3 = L.Linear(n_units, n_out)
def __call__(self, x):
h1 = F.relu(self.l1(x))
h2 = F.relu(self.l2(h1))
return self.l3(h2)
定数の定義
gpu_device = 0
epoch = 30
batch_size = 512
frequency = -1
n_in = 784
n_units = 100
n_out = 10
学習
model = L.Classifier(NN(n_in, n_units, n_out))
chainer.cuda.get_device_from_id(0)
model.to_gpu()
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
train_iter = chainer.iterators.SerialIterator(train, batch_size)
test_iter = chainer.iterators.SerialIterator(test, batch_size, repeat=False, shuffle=False)
updater = training.StandardUpdater(train_iter, optimizer, device=gpu_device)
trainer = training.Trainer(updater, (epoch, 'epoch'))
trainer.extend(extensions.Evaluator(test_iter, model,device=gpu_device))
trainer.extend(extensions.dump_graph('main/loss'))
frequency = epoch if frequency == -1 else max(1, frequency)
trainer.extend(extensions.snapshot(), trigger=(frequency, 'epoch'))
trainer.extend(extensions.LogReport())
trainer.extend(
extensions.PlotReport(['main/loss', 'validation/main/loss'],
'epoch', file_name='loss.png'))
trainer.extend(
extensions.PlotReport(['main/accuracy', 'validation/main/accuracy'],
'epoch', file_name='accuracy.png'))
trainer.extend(extensions.PrintReport(
['epoch', 'main/loss', 'validation/main/loss',
'main/accuracy', 'validation/main/accuracy', 'elapsed_time']))
trainer.run()
実行結果
epoch main/loss validation/main/loss main/accuracy validation/main/accuracy elapsed_time
1 0.609557 0.271625 0.83986 0.921639 2.24247
2 0.223956 0.190223 0.936165 0.942412 3.91926
3 0.169551 0.155889 0.95229 0.952826 5.49803
4 0.138876 0.130223 0.960069 0.960507 7.12243
5 0.114769 0.117665 0.966947 0.964746 8.73528
6 0.096983 0.113826 0.971994 0.965614 10.3822
7 0.0863535 0.0950723 0.97476 0.972306 11.9649
8 0.0738427 0.0943972 0.978816 0.971743 13.6363
9 0.0651506 0.0920507 0.98107 0.971375 15.2552
10 0.0575137 0.0831638 0.983407 0.97492 16.8571
11 0.0515834 0.0856705 0.984938 0.974575 18.4335
12 0.0451553 0.0791185 0.986946 0.975506 20.0122
13 0.0401197 0.0787027 0.988732 0.976201 21.5908
14 0.0361692 0.0773779 0.989934 0.976982 23.1778
15 0.0320661 0.0788221 0.991319 0.975615 24.7676
16 0.0292983 0.0776353 0.99187 0.976459 26.4491
17 0.0255289 0.0842461 0.99328 0.976017 28.0744
18 0.0226313 0.0798943 0.994308 0.976005 29.6888
19 0.0198879 0.0758416 0.995276 0.977935 31.3053
20 0.0175719 0.079733 0.99581 0.977372 32.9155
21 0.0158291 0.0838353 0.996244 0.976384 34.5392
22 0.015166 0.0863985 0.996524 0.975776 36.1806
23 0.013793 0.0871418 0.996578 0.975494 37.7816
24 0.0118357 0.0869616 0.997362 0.975982 39.3604
25 0.00977673 0.0892306 0.998013 0.976373 40.9641
26 0.00908512 0.0872242 0.998197 0.976298 42.679
27 0.0102062 0.0899191 0.997732 0.97708 44.2737
28 0.00712254 0.087929 0.998781 0.978315 45.8738
29 0.00570363 0.0855932 0.999182 0.977263 47.4935
30 0.00537472 0.0911612 0.999299 0.977752 49.1054
学習曲線
from IPython.display import Image, display_png
display_png(Image('result/loss.png'))
display_png(Image('result/accuracy.png'))
予測
import cupy as cp
plt.imshow(np.array(test[0][0][0], dtype=np.float32), cmap='gray')
print('この画像データのラベルは{:0d}です。'.format(test[0][1]))
prediction = model.predictor(cp.array(test[0][0][0]).reshape(1, 784))
probability = chainer.cuda.to_cpu(F.softmax(prediction).data[0])
np.set_printoptions(precision=20, floatmode='fixed', suppress=True)
print(probability)
この画像データのラベルは7です。
[0.00000000372259179038 0.00000000087596779830 0.00000003106548618348
0.00004456961323739961 0.00000000000002767143 0.00000000147911294324
0.00000000000000000014 0.99994862079620361328 0.00000134408105623152
0.00000531150408278336]
畳み込みニューラルネットワークでの学習
モデルの定義
class CNN(Chain):
def __init__(self):
super(CNN, self).__init__()
with self.init_scope():
self.cn1 = L.Convolution2D(1, 20, 5)
self.cn2 = L.Convolution2D(20, 50, 5)
self.fc1 = L.Linear(800, 500)
self.fc2 = L.Linear(500, 10)
def __call__(self, x):
h1 = F.max_pooling_2d(F.relu(self.cn1(x)), 2)
h2 = F.max_pooling_2d(F.relu(self.cn2(h1)), 2)
h3 = F.dropout(F.relu(self.fc1(h2)))
return self.fc2(h3)
定数の定義
gpu_device = 0
epoch = 30
batch_size = 512
frequency = -1
学習
model = L.Classifier(CNN())
chainer.cuda.get_device_from_id(0)
model.to_gpu()
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
train_iter = chainer.iterators.SerialIterator(train, batch_size)
test_iter = chainer.iterators.SerialIterator(test, batch_size, repeat=False, shuffle=False)
updater = training.StandardUpdater(train_iter, optimizer, device=gpu_device)
trainer = training.Trainer(updater, (epoch, 'epoch'))
trainer.extend(extensions.Evaluator(test_iter, model,device=gpu_device))
trainer.extend(extensions.dump_graph('main/loss'))
frequency = epoch if frequency == -1 else max(1, frequency)
trainer.extend(extensions.snapshot(), trigger=(frequency, 'epoch'))
trainer.extend(extensions.LogReport())
trainer.extend(
extensions.PlotReport(['main/loss', 'validation/main/loss'],
'epoch', file_name='loss.png'))
trainer.extend(
extensions.PlotReport(['main/accuracy', 'validation/main/accuracy'],
'epoch', file_name='accuracy.png'))
trainer.extend(extensions.PrintReport(
['epoch', 'main/loss', 'validation/main/loss',
'main/accuracy', 'validation/main/accuracy', 'elapsed_time']))
trainer.run()
epoch main/loss validation/main/loss main/accuracy validation/main/accuracy elapsed_time
1 0.354028 0.0805555 0.892197 0.974782 2.43793
2 0.0861291 0.0457051 0.97476 0.984651 5.02699
3 0.0603717 0.0358304 0.98152 0.98785 7.64132
4 0.0447528 0.029547 0.986278 0.990659 10.286
5 0.0393856 0.0300916 0.98748 0.990734 12.9046
6 0.0335107 0.0276662 0.98939 0.991355 15.5664
7 0.028251 0.0264475 0.991169 0.990682 18.4082
8 0.024206 0.0245272 0.992405 0.991234 21.0302
9 0.0213547 0.0207464 0.993339 0.992699 23.6642
10 0.0193615 0.0218372 0.993924 0.99271 26.3024
11 0.0163818 0.0214707 0.995051 0.992613 28.9265
12 0.0147773 0.0218644 0.995292 0.992894 31.5535
13 0.0131757 0.0236539 0.99591 0.992124 34.1893
14 0.0123087 0.0218821 0.99606 0.99259 36.9303
15 0.0108798 0.0218928 0.996595 0.992515 39.5811
16 0.00979711 0.0227453 0.996862 0.992722 42.2415
17 0.0107679 0.0222965 0.996557 0.992992 44.9246
18 0.00846383 0.0256104 0.997246 0.991625 47.5767
19 0.0078571 0.0252699 0.997429 0.991636 50.2381
20 0.00736991 0.0229966 0.99763 0.992905 52.8978
21 0.00675241 0.0229066 0.997963 0.992894 55.5817
22 0.00664717 0.0212612 0.997997 0.993578 58.3709
23 0.00643474 0.0251418 0.997963 0.99205 61.0157
24 0.00604594 0.0260536 0.997947 0.992015 63.663
25 0.00545904 0.0283221 0.998214 0.993003 66.3088
26 0.00546739 0.0324786 0.998047 0.991136 68.9498
27 0.00519143 0.0309123 0.998245 0.992124 71.6131
28 0.0049064 0.0308791 0.998264 0.991257 74.2487
29 0.00435375 0.0302761 0.998448 0.99194 76.9035
30 0.00312136 0.0303936 0.999115 0.992503 79.5643
学習曲線
from IPython.display import Image, display_png
display_png(Image('result/loss.png'))
display_png(Image('result/accuracy.png'))
予測
import cupy as cp
plt.imshow(np.array(test[0][0][0], dtype=np.float32), cmap='gray')
print('この画像データのラベルは{:0d}です。'.format(test[0][1]))
prediction = model.predictor(cp.array(test[0][0][0]).reshape(1, 1, 28, 28))
probability = chainer.cuda.to_cpu(F.softmax(prediction).data[0])
np.set_printoptions(precision=20, floatmode='fixed', suppress=True)
print(probability)
この画像データのラベルは7です。
[0.00000000000034654969 0.00000000001631320079 0.00000000017968711241
0.00000000000952125098 0.00000000180500603353 0.00000000000172482167
0.00000000000000372396 1.00000000000000000000 0.00000000000094331075
0.00000000918344689183]
リカレントニューラルネットワークでの学習
定数の定義
gpu_device = 0
epoch = 30
batch_size = 500
frequency = -1
モデルの定義
class RNN(Chain):
def __init__(self):
super(RNN, self).__init__()
with self.init_scope():
self.w1 = L.Linear(784, 100)
self.h1 = L.Linear(100, 100)
self.o = L.Linear(100, 10)
def reset_state(self):
self.last_z = chainer.Variable(cp.zeros((batch_size, 100), dtype=np.float32))
def __call__(self, x):
z = F.relu(self.w1(x) + self.h1(self.last_z))
self.last_z = z
y = F.relu(self.o(z))
return y
学習
rnn = RNN()
rnn.reset_state()
model = L.Classifier(rnn)
chainer.cuda.get_device_from_id(0)
model.to_gpu()
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.GradientClipping(10.0))
train_iter = chainer.iterators.SerialIterator(train, batch_size)
test_iter = chainer.iterators.SerialIterator(test, batch_size, repeat=False, shuffle=False)
updater = training.StandardUpdater(train_iter, optimizer, device=gpu_device)
trainer = training.Trainer(updater, (epoch, 'epoch'))
trainer.extend(extensions.Evaluator(test_iter, model,device=gpu_device))
trainer.extend(extensions.dump_graph('main/loss'))
frequency = epoch if frequency == -1 else max(1, frequency)
trainer.extend(extensions.snapshot(), trigger=(frequency, 'epoch'))
trainer.extend(extensions.LogReport())
trainer.extend(
extensions.PlotReport(['main/loss', 'validation/main/loss'],
'epoch', file_name='loss.png'))
trainer.extend(
extensions.PlotReport(['main/accuracy', 'validation/main/accuracy'],
'epoch', file_name='accuracy.png'))
trainer.extend(extensions.PrintReport(
['epoch', 'main/loss', 'validation/main/loss',
'main/accuracy', 'validation/main/accuracy', 'elapsed_time']))
trainer.run()
epoch main/loss validation/main/loss main/accuracy validation/main/accuracy elapsed_time
1 1.12647 0.67986 0.676317 0.7929 14.2889
2 0.637657 0.561256 0.7971 0.8164 28.7837
3 0.554199 0.508853 0.816633 0.8303 43.3512
4 0.514805 0.482239 0.826266 0.8361 57.8673
5 0.482919 0.462028 0.834117 0.8405 72.5115
6 0.459542 0.436305 0.84055 0.8471 87.0379
7 0.439237 0.422526 0.845017 0.8518 101.544
8 0.422248 0.407439 0.850083 0.8552 116.104
9 0.408515 0.397422 0.853317 0.8583 130.608
10 0.395496 0.386185 0.856666 0.8606 145.18
11 0.386715 0.383238 0.859083 0.8604 159.626
12 0.376893 0.380003 0.861716 0.8609 174.196
13 0.369096 0.366671 0.863667 0.8654 188.681
14 0.36017 0.360889 0.865816 0.8671 203.158
15 0.35465 0.358045 0.867233 0.8669 217.71
16 0.348177 0.354098 0.8692 0.8683 232.215
17 0.341731 0.346596 0.870567 0.871 246.759
18 0.335474 0.343758 0.872 0.8717 261.409
19 0.330673 0.341848 0.873383 0.8716 275.912
20 0.325912 0.337807 0.874617 0.8728 290.407
21 0.316205 0.159358 0.879183 0.9563 305.075
22 0.094005 0.109159 0.97395 0.9694 319.652
23 0.0830744 0.107424 0.976567 0.969 334.165
24 0.0779403 0.100136 0.977933 0.9712 348.725
25 0.0724973 0.0975639 0.97945 0.9719 363.134
26 0.0687905 0.0933854 0.980384 0.972 377.548
27 0.0643198 0.0951944 0.98205 0.9713 392.084
28 0.060734 0.0910737 0.98275 0.9726 406.522
29 0.0582916 0.089033 0.9841 0.9735 421.009
30 0.0541574 0.0874746 0.984867 0.9745 435.481
学習曲線
from IPython.display import Image, display_png
display_png(Image('result/loss.png'))
display_png(Image('result/accuracy.png'))
参考
リカレントニューラルネットワークのChainerでの実装について、下記ページを参考にいたしました。
https://qiita.com/knok/items/3077230a9dc6b7979173