Cursive Hiragana MNIST using NN

5 min readOct 19, 2021

Hiragana is one component of a Japanese writing system, along with Kanji, Katakana, and Romaji.

Cursive Hiragana or Kuzushiji means cursive handwritten hiragana. There are millions of pages of handwritten hiragana (Kuzushiji) available now and writing in this style became difficult to be interpreted by common Japanese people. The variety of shapes for each symbol is quite large, due to the free-form of these characters. Therefore the automatic interpretation of such digitized texts might be difficult.

Recently, a Japanese-British Data Science team created a database of Kuzushiji images, available now on Kaggle. This is suggested as an alternative for the MNIST dataset and is called Kuzushiji-MNIST (Japanese Literature Alternative Dataset for Deep Learning Tasks). The dataset can be used to train models to recognize cursive hiragana characters in classical texts.

The present dataset is derived from the KMNIST dataset.

The training set contains 50,000 grey level images with dimensions 28 x 28.
The test set contains 20,000 images.

There are 10 different classes of images. For the training set, image labels are given in a separate file. Each class corresponds to a major type of hiragana, as shown below:
index (char): 0 (お), 1 (き), 2 (す), 3 (つ), 4 (な), 5 (は), 6 (ま), 7 (や), 8 (れ), 9 (を)

Input Libraries

Input Data

['kmnist-train-labels.npz', 'kmnist-train-imgs.npz', 'kmnist-test-labels.npz', 'kmnist_classmap.csv', 'kmnist-test-imgs.npz']

KMNIST train shape: (60000, 28, 28)
KMNIST test shape: (10000, 28, 28)
KMNIST train shape: (60000,)
KMNIST character map shape: (10, 3)

Percent for each category: [10. 10. 10. 10. 10. 10. 10. 10. 10. 10.]

Data Visualization

Data Preprocess

Train Test Split

KMNIST train -  rows: 48000  columns: (28, 28, 1)
KMNIST valid -  rows: 12000  columns: (28, 28, 1)
KMNIST test -  rows: 10000  columns: (28, 28, 1)

4(な):   4845 or 10.09375%
9(を):   4833 or 10.06875%
0(お):   4833 or 10.06875%
6(ま):   4809 or 10.01875%
8(れ):   4806 or 10.012500000000001%
2(す):   4800 or 10.0%
1(き):   4795 or 9.989583333333334%
3(つ):   4791 or 9.98125%
5(は):   4781 or 9.960416666666665%
7(や):   4707 or 9.80625%

7(や):   1293 or 10.775%
5(は):   1219 or 10.158333333333333%
3(つ):   1209 or 10.075000000000001%
1(き):   1205 or 10.041666666666666%
2(す):   1200 or 10.0%
8(れ):   1194 or 9.950000000000001%
6(ま):   1191 or 9.925%
9(を):   1167 or 9.725%
0(お):   1167 or 9.725%
4(な):   1155 or 9.625%

Model Architecture

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 28, 28, 16)        160       
_________________________________________________________________
batch_normalization (BatchNo (None, 28, 28, 16)        64        
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 26, 26, 16)        2320      
_________________________________________________________________
batch_normalization_1 (Batch (None, 26, 26, 16)        64        
_________________________________________________________________
flatten (Flatten)            (None, 10816)             0         
_________________________________________________________________
dense (Dense)                (None, 10)                108170    
=================================================================
Total params: 110,778
Trainable params: 110,714
Non-trainable params: 64

Training the model

Epoch 1/10
  1/375 [..............................] - ETA: 0s - loss: 3.0626 - accuracy: 0.0859WARNING:tensorflow:Callbacks method `on_train_batch_end` is slow compared to the batch time (batch time: 0.0019s vs `on_train_batch_end` time: 0.0033s). Check your callbacks.
375/375 [==============================] - 2s 4ms/step - loss: 0.3864 - accuracy: 0.8914 - val_loss: 0.3888 - val_accuracy: 0.8773
Epoch 2/10
375/375 [==============================] - 1s 4ms/step - loss: 0.1316 - accuracy: 0.9612 - val_loss: 0.2332 - val_accuracy: 0.9372
Epoch 3/10
375/375 [==============================] - 1s 4ms/step - loss: 0.0678 - accuracy: 0.9795 - val_loss: 0.2286 - val_accuracy: 0.9434
Epoch 4/10
375/375 [==============================] - 1s 4ms/step - loss: 0.0334 - accuracy: 0.9905 - val_loss: 0.2308 - val_accuracy: 0.9443
Epoch 5/10
375/375 [==============================] - 1s 4ms/step - loss: 0.0199 - accuracy: 0.9945 - val_loss: 0.2426 - val_accuracy: 0.9462
Epoch 6/10
375/375 [==============================] - 1s 4ms/step - loss: 0.0114 - accuracy: 0.9977 - val_loss: 0.2503 - val_accuracy: 0.9438
Epoch 7/10
375/375 [==============================] - 1s 4ms/step - loss: 0.0097 - accuracy: 0.9977 - val_loss: 0.2867 - val_accuracy: 0.9438
Epoch 8/10
375/375 [==============================] - 1s 4ms/step - loss: 0.0215 - accuracy: 0.9931 - val_loss: 0.3793 - val_accuracy: 0.9283
Epoch 9/10
375/375 [==============================] - 1s 4ms/step - loss: 0.0587 - accuracy: 0.9814 - val_loss: 0.3890 - val_accuracy: 0.9402
Epoch 10/10
375/375 [==============================] - 1s 4ms/step - loss: 0.0342 - accuracy: 0.9897 - val_loss: 0.3755 - val_accuracy: 0.9353

Accuracy and Loss Graphs

Predictions

Correct predicted classes: 11224
Incorrect predicted classes: 776

                 precision    recall  f1-score   support

Class 0 (お):       0.96      0.96      0.96      1167
Class 1 (き):       0.97      0.91      0.94      1205
Class 2 (す):       0.95      0.85      0.90      1200
Class 3 (つ):       0.92      0.96      0.94      1209
Class 4 (な):       0.92      0.93      0.92      1155
Class 5 (は):       0.95      0.90      0.93      1219
Class 6 (ま):       0.83      0.96      0.89      1191
Class 7 (や):       0.97      0.96      0.97      1293
Class 8 (れ):       0.93      0.96      0.95      1194
Class 9 (を):       0.98      0.95      0.96      1167

    accuracy                           0.94     12000
   macro avg       0.94      0.94      0.94     12000
weighted avg       0.94      0.94      0.94     12000

DeepCC

[INFO]
Reading [keras model] 'model.h5'
[SUCCESS]
Saved 'model_deepC/model.onnx'
[INFO]
Reading [onnx model] 'model_deepC/model.onnx'
[INFO]
Model info:
  ir_vesion : 4
  doc       : 
[WARNING]
[ONNX]: graph-node conv2d's attribute auto_pad has no meaningful data.
[WARNING]
[ONNX]: terminal (input/output) conv2d_input's shape is less than 1. Changing it to 1.
[WARNING]
[ONNX]: terminal (input/output) dense's shape is less than 1. Changing it to 1.
WARN (GRAPH): found operator node with the same name (dense) as io node.
[INFO]
Running DNNC graph sanity check ...
[SUCCESS]
Passed sanity check.
[INFO]
Writing C++ file 'model_deepC/model.cpp'
[INFO]
deepSea model files are ready in 'model_deepC/' 
[RUNNING COMMAND]
g++ -std=c++11 -O3 -fno-rtti -fno-exceptions -I. -I/opt/tljh/user/lib/python3.7/site-packages/deepC-0.13-py3.7-linux-x86_64.egg/deepC/include -isystem /opt/tljh/user/lib/python3.7/site-packages/deepC-0.13-py3.7-linux-x86_64.egg/deepC/packages/eigen-eigen-323c052e1731 "model_deepC/model.cpp" -D_AITS_MAIN -o "model_deepC/model.exe"
[RUNNING COMMAND]
size "model_deepC/model.exe"
   text	   data	    bss	    dec	    hex	filename
 602405	   3792	    760	 606957	  942ed	model_deepC/model.exe
[SUCCESS]
Saved model as executable "model_deepC/model.exe"

Notebook Lin- Here

Credits- Siddharth Ganjoo