1
2

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

More than 5 years have passed since last update.

「not only 数字 but also 平仮名 ~MNIST だけでなく NDL もまた~」Chainerで手書きを認識

Posted at

NDL

NDLラボへようこそ。
ここは、国立国会図書館の実験的なサービスを提供するサイトです。

サンプルプログラムの実行手順の例が公開されている
$ wget http://lab.ndl.go.jp/dataset/hiragana73.tar.gz
$ tar xfz hiragana73.tar.gz
$ wget http://lab.ndl.go.jp/dataset/example/train_ndlkana.tar.gz
$ tar xfz train_ndlkana.tar.gz
$ python example/ndlkana/train.py -d hiragana73

先に成果を公開

やってみる

環境

そのまま実行するとエラー
ValueError                                Traceback (most recent call last)
/data/Documents/ndl/example/ndlkana/train.py in <module>()
    123         for i in six.moves.range(0, n_test, batchsize):
    124 		x = chainer.Variable(xp.asarray(x_test[i:i + batchsize]),
--> 125 							 volatile='on')
    126 		t = chainer.Variable(xp.asarray(y_test[i:i + batchsize]),
    127 							 volatile='on')
  • chainer.Vaiableのvolatileが古い

Modified from examples/mnist programs of Chainer 1.9.0.

  • Variableなくすとかやり方がありそうです、簡単な対処はこんなものでしょうか
  • 編集後のpy
example/ndlkana/train.py.diff
-		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)
+		x = chainer.Variable(xp.asarray(x_test[i:i + batchsize]))
+		t = chainer.Variable(xp.asarray(y_test[i:i + batchsize]))
+		with chainer.no_backprop_mode():
+			loss = model(x, t)
学習の様子
Namespace(batchsize=100, datadir='hiragana73', epoch=20, fspec='*.png', initmodel='', resume='', testratio=0.14285714285714285)
load NDLKANA dataset
n_train=68597 n_test=11403
n_class=73
epoch 1
graph generated
train mean loss=0.57483, accuracy=0.86960, throughput=2499.3 images/sec
test  mean loss=0.13620, accuracy=0.96659
~略~
epoch 20
train mean loss=0.01600, accuracy=0.99509, throughput=2380.1 images/sec
test  mean loss=0.07081, accuracy=0.98606
save the model
save the optimizer
  • 学習成果デモのコードは公開されてないので書いてみる
  • jupyter上で実施したのでdisplay(Image...のところは必要に応じて読み替えてください
demo.py
from pathlib import Path
import sys
sys.path.append(str(Path('./example/ndlkana').absolute()))
import net
import data

import numpy as np
import random
from PIL import Image

import chainer
import chainer.links as L
from chainer import serializers

# Prepare dataset
print('load NDLKANA dataset')
ndlkana = data.load_ndlkana_data('hiragana73', '*.png', 1.0/7.0)
ndlkana['data'] = ndlkana['data'].astype(np.float32)
ndlkana['data'] /= 255
n_test = ndlkana['testsize']
n_train = ndlkana['data'].shape[0] - n_test
print("n_train={} n_test={}".format(n_train, n_test))

_, x_test = np.split(ndlkana['data'], [n_train])

# Prepare CNN model, defined in net.py
model = L.Classifier(net.NdlkanaCNN())
serializers.load_npz('hiragana73.model', model)

# 推測の対象画像をランダム選択
i = random.randrange(n_test)
img = x_test[i:i + 1]
print('--画像')
display(Image.fromarray(np.uint8(img[0][0]*255)))

# 推測
x = chainer.Variable(np.asarray(img))
with chainer.using_config('train', False):
    y = model.predictor(x)
hexcode = ndlkana['label'][y.data[0].argmax()]
print('--候補一位の文字')
print(chr(int(hexcode.replace('U', '0x'), 16)))

# print('--候補ぜんぶ(ラベルのインデックス)')
# print(np.argsort(y.data[0])[-1::-1])
# print('--確率ぜんぶ')
# print(np.sort(y.data[0])[-1::-1])
出力
--画像
※ここは画像がはいります
--候補一位の文字
せ
--候補ぜんぶ(ラベルのインデックス)
[21 22 51 66 40 25 33 50 59 37 36  9 31  4 42 63 28 19 71 39 11 18 61 34 72
 64  0 16 35 60 14 48 32 38  7 26 58 57  6 53 70  1 41 44 49 17 13  3 15 56
 52 27  5  8 29 10 23 24 43 45 68 47 65 67 62 55 30  2 54 20 46 69 12]
--確率ぜんぶ
[  7.39292622   0.80466545   0.5365994   -2.02391601  -2.12653828
  -2.44498348  -2.61160302  -3.17728567  -4.87509871  -5.18633127
  -5.77648926  -5.90091658  -7.84125566  -7.89595032  -8.23139668
  -8.379076    -8.40882969  -8.57416248  -8.83006859  -9.06515026
  -9.61909008  -9.74232578 -10.18440247 -10.19537926 -10.39629555
 -10.53983307 -10.58770752 -10.98420334 -11.07983494 -11.25813293
 -11.76102638 -11.86397743 -11.97500134 -11.98637295 -12.34544277
 -12.57084084 -12.58101082 -12.59549904 -12.79733276 -12.8950882
 -13.12607479 -13.21101952 -13.55077934 -13.6707716  -13.68242836
 -14.06533813 -14.11181068 -14.27572346 -14.33026981 -14.49097729
 -14.53163624 -14.62358761 -14.70830154 -14.91421604 -15.22727203
 -15.28550529 -15.61158752 -15.94737148 -16.43763351 -16.58153725
 -16.94737053 -17.14120674 -17.96526337 -18.85281563 -19.24281311
 -19.41680145 -19.64408112 -20.11096191 -20.14222908 -20.1970253
 -22.58972931 -23.84316826 -24.02269363]

3462.png

その他、国立国会図書館が公開してるデータ、いつか使えそうなのでメモ

次の一歩

  • 複数の文字、というか文字列というか、文書をOCRさせるには?
  • 文字でなく単語を学習する必要が合る?
  • ノートやホワイトボードのように自由な場所に書く場合はbboxが必要?

おしまい

1
2
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
1
2

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?