                               ___                                     ___           ___     
    _____                     /\__\                                   /\  \         /\__\    
   /::\  \       ___         /:/ _/_       ___           ___         /::\  \       /:/ _/_   
  /:/\:\  \     /\__\       /:/ /\  \     /\__\         /\__\       /:/\:\__\     /:/ /\__\  
 /:/  \:\__\   /:/__/      /:/ /::\  \   /:/__/        /:/  /      /:/ /:/  /    /:/ /:/ _/_ 
/:/__/ \:|__| /::\  \     /:/__\/\:\__\ /::\  \       /:/__/      /:/_/:/__/___ /:/_/:/ /\__\
\:\  \ /:/  / \/\:\  \__  \:\  \ /:/  / \/\:\  \__   /::\  \      \:\/:::::/  / \:\/:/ /:/  /
 \:\  /:/  /   ~~\:\/\__\  \:\  /:/  /   ~~\:\/\__\ /:/\:\  \      \::/~~/~~~~   \::/_/:/  / 
  \:\/:/  /       \::/  /   \:\/:/  /       \::/  / \/__\:\  \      \:\~~\        \:\/:/  /  
   \::/  /        /:/  /     \::/  /        /:/  /       \:\__\      \:\__\        \::/  /   
    \/__/         \/__/       \/__/         \/__/         \/__/       \/__/         \/__/

December 2016

Build prediction model with tflearn and tensorflow
- Multilayer convolutional neural network
- Serialize trained CNN model for serving
Classify digit examples from Digitre
- Example step-by-step preprocessing
- Classify preprocessed images



In [1]:

    
# Standard library
import datetime
import time

# Third party libraries
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline



In [2]:

    
# Digitre code
import digitre_preprocessing as prep
import digitre_model
import digitre_classifier



In [15]:

    
# Reload digitre code in the same session (during development)
import imp
imp.reload(prep)
imp.reload(digitre_model)
imp.reload(digitre_classifier)









    Out[15]:





<module 'digitre_classifier' from '/Users/luis/Git/digitre/digitre/digitre_classifier.py'>



In [16]:

    
# Latest update
str(datetime.datetime.now())









    Out[16]:





'2017-03-21 19:35:34.787175'

Build prediction model with tflearn and tensorflow

Multilayer convolutional neural network

Code based on this tflearn example, with CNN architecture modeled after TensorFlow's tutorial Deep MNIST for experts.



In [5]:

    
# Data loading and preprocessing
X, Y, testX, testY = digitre_model.load_data()
#X = X.reshape([-1, 28, 28, 1])
#testX = testX.reshape([-1, 28, 28, 1])









    



Extracting mnist/train-images-idx3-ubyte.gz






    



/Users/luis/anaconda3/lib/python3.5/gzip.py:274: VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future
  return self._buffer.read(size)
/Users/luis/anaconda3/lib/python3.5/site-packages/tflearn/datasets/mnist.py:52: VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future
  data = data.reshape(num_images, rows, cols, 1)






    



Extracting mnist/train-labels-idx1-ubyte.gz
Extracting mnist/t10k-images-idx3-ubyte.gz
Extracting mnist/t10k-labels-idx1-ubyte.gz



In [6]:

    
# Plot functions
def plot_digit(digit, show=True, file_name=None):
    plt.imshow(digit, cmap = 'Greys', interpolation = 'none')
    plt.tick_params(axis='both', which='both', bottom='off', top='off',
                    labelbottom='off', right='off', left='off', labelleft='off')
    if file_name is not None:
        plt.savefig(file_name)
    if show:
        plt.show()
    
def plot_digits(digits, rows, columns):
    for i, digit in enumerate(digits):
        plt.subplot(rows, columns, i+1)
        plot_digit(digit, show=False)
    plt.show()



In [7]:

    
# Plot a few training examples
X_eg = X[10:20,:,:,:]
X_eg = [digit.reshape(28, 28) for digit in X_eg]

plot_digits(X_eg, 2, 5)

### Build convolutional cnn model cnn = digitre_model.build() # Training t0 = time.time() digitre_model.fit(cnn, X, Y, testX, testY, n_epoch=20, run_id='cnn_mnist') print('-----') print('-----') print('Completed in') prep.print_elapsed_time(t0)



In [103]:

    
# Visualization
# Used "tensorboard_verbose=0", meaning Loss & Metric
# Run "$ tensorboard --logdir='/tmp/tflearn_logs'"

# Save the model model.save('cnn.tflearn') ### Output after 20 epochs # Training Step: 17200 | total loss: 0.45195 # | Adam | epoch: 020 | loss: 0.45195 - acc: 0.9702 | val_loss: 0.03177 - val_acc: 0.9892 -- iter: 55000/55000 # -- # ----- # ----- # Completed in # 1.2 hr.



In [104]:

    
### Fit model using all data (merge training and test data)
# Done from command line:
# $ python digitre_model.py -f 'cnn_alldata.tflearn' -a -e 20

# Training Step: 20320  | total loss: 0.642990.9401 | val_loss: 0.052
# | Adam | epoch: 020 | loss: 0.64299 - acc: 0.9401 | val_loss: 0.05263 - val_acc: 0.9866 -- iter: 65000/65000
# --
# -----
# Completed training in
# 3.5 hr.
# -----
# ... Saving trained model as " cnn_alldata.tflearn "

Classify digit examples from Digitre

Example step-by-step preprocessing

Take example base64-encoded handwritten digit images (generated from html canvas element) and preprocess step-by-step to a format ready for classification model. Compare with MNIST example.



In [17]:

    
with open('b64_2_preprocessing.txt', 'r') as f:
    eg_2 = f.read()

# Preview base64 encoded image
print(eg_2[:500])









    



data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAPoAAAD6CAYAAACI7Fo9AAAQ5klEQVR4Xu2dTcg2VQGGb6lNUJrUJjL8oywkrY0FRtZGhCBtV7nQIq2dCtrvolwElUK6K4y0hWFt1LYtLMuFbkyJwOhPSEQqtCikRRk3ztBhmJ/nfZ+Z55kz9zXw8X3f+84zc+77Ptdzzpw5c+YMseEADmzegTM2rxCBOIADAnQqAQ4EOADoASEjEQcAnTqAAwEOAHpAyEjEAUCnDuBAgAOAHhAyEnEA0KkDOBDgAKAHhIxEHAB06gAOBDgA6AEhIxEHAJ06gAMBDgB6QMhIxAFApw7gQIADgB4QMhJxANCpAzgQ4ACgB4SMRBwAdOoADgQ4AOgBISMRBwCdOoADAQ4AekDISMQBQKcO4ECAA4AeEHKQxPMkXS3pGkn+97OSrij0/7z5v/9+RdKvJD0s6Wdb9wjQt55whj6DfZWkz3bktm



In [18]:

    
eg_2 = prep.b64_str_to_np(eg_2)
eg_2.shape









    Out[18]:





(250, 250)



In [19]:

    
# Plot the example handwritten digit
plot_digit(eg_2, file_name='b64_2_preprocessing_1.png')



In [20]:

    
eg_2 = prep.crop_img(eg_2)



In [21]:

    
plot_digit(eg_2, file_name='b64_2_preprocessing_2.png')



In [22]:

    
eg_2 = prep.center_img(eg_2)



In [23]:

    
plot_digit(eg_2, file_name='b64_2_preprocessing_3.png')



In [24]:

    
eg_2 = prep.resize_img(eg_2)
eg_2.shape









    Out[24]:





(28, 28)



In [25]:

    
plot_digit(eg_2, file_name='b64_2_preprocessing_4.png')



In [26]:

    
eg_2 = prep.min_max_scaler(eg_2, final_range=(0, 1))



In [27]:

    
plot_digit(eg_2)



In [28]:

    
# Plot processed Digitre image together with MNIST example
plot_digits([eg_2, X_eg[6]], 1, 2)



In [29]:

    
# Save MNIST example too
plot_digit(X_eg[6], file_name='MNIST_2.png')



In [116]:

    
eg_2.max()









    Out[116]:





1.0



In [117]:

    
eg_2.shape









    Out[117]:





(28, 28)

Classify preprocessed images



In [58]:

    
# Instantiate Classifier (loads the tflearn pre-trained model)
model = digitre_classifier.Classifier(file_name='cnn.tflearn')



In [118]:

    
# Classify same example digit
with open('b64_2_preprocessing.txt', 'r') as f:
    eg_2 = f.read()

eg_2 = model.preprocess(eg_2)
pred = np.around(model.classify(eg_2)[0], 2)
pred









    Out[118]:





array([ 0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.])



In [119]:

    
from altair import Chart, Data, X, Y, Axis, Scale

# Plot prediction
def prob_distribution_plot(pred):
    prediction = pred.reshape([10])
    data = Data(values=[{'x': i, 'y': value} for i, value in enumerate(pred)])
    plot = Chart(data).mark_bar(color='#f6755e').encode(
        x=X('x:O', axis=Axis(title='Digit', labelAngle=0.5,
                             tickLabelFontSize=15, titleFontSize=15)),
        y=Y('y:Q', axis=Axis(format='%', title='Probability',
                             tickLabelFontSize=15, titleFontSize=15),
            scale=Scale(domain=(0, 1))))
    
    return plot

prob_distribution_plot(pred)



In [120]:

    
from altair import Chart, Data, X, Y, Axis

# Plot prediction
def prob_distribution_plot(pred):
    prediction = pred.reshape([10])
    data = Data(values=[{'x': i, 'y': value} for i, value in enumerate(prediction)])
    plot = Chart(data).mark_bar(color='#f6755e').encode(
        x=X('x:O', axis=Axis(title='Digit', labelAngle=0.5, tickLabelFontSize=15, titleFontSize=15)),
        y=Y('y:Q', axis=Axis(format='%', title='Probability', tickLabelFontSize=15, titleFontSize=15)))
    
    return plot.to_json(indent=2)

prob_distribution_plot(pred)