In [1]:

    

import gzip
import numpy as np
import pickle
import matplotlib.pyplot as plt
import tensorflow as tf


    

In [2]:

    

f=gzip.open('/home/jesse/Code/neural-networks-and-deep-learning/data/mnist.pkl.gz')


    

In [3]:

    

training_data, validation_data, test_data = pickle.load(f,encoding='latin1')
f.close()


    

In [4]:

    

training_data[0].shape


    

    
Out[4]:





(50000, 784)

In [5]:

    

exampl = training_data[0][0].reshape(28,28)
plt.subplot(2,3,1)
plt.imshow(exampl)

exampl = training_data[0][1].reshape(28,28)
plt.subplot(2,3,2)
plt.imshow(exampl)

exampl = training_data[0][2].reshape(28,28)
plt.subplot(2,3,3)
plt.imshow(exampl)

exampl = training_data[0][3].reshape(28,28)
plt.subplot(2,3,4)
plt.imshow(exampl)

exampl = training_data[0][4].reshape(28,28)
plt.subplot(2,3,5)
plt.imshow(exampl)

exampl = training_data[0][5].reshape(28,28)
plt.subplot(2,3,6)
plt.imshow(exampl)

plt.show()
print(training_data[1][2])


    

    













    




4

In [6]:

    

f=gzip.open('/home/jesse/Code/neural-nets/MNIST_data/train-images-idx3-ubyte.gz')


    

In [7]:

    

import keras
from keras.models import Sequential


    

    




Using TensorFlow backend.

In [39]:

    

model = Sequential()


    

In [8]:

    

from keras.layers import Dense, Activation


    

In [16]:

    

model = Sequential([
    Dense(256, input_shape=(784,),activation='relu'),
    Dense(256,activation='relu'),
    Dense(10,activation='softmax',trainable=),
])


    

In [17]:

    

model.summary()


    

    




_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_7 (Dense)              (None, 256)               200960    
_________________________________________________________________
dense_8 (Dense)              (None, 256)               65792     
_________________________________________________________________
dense_9 (Dense)              (None, 10)                2570      
=================================================================
Total params: 269,322
Trainable params: 266,752
Non-trainable params: 2,570
_________________________________________________________________

In [15]:

    

65792-256*256


    

    
Out[15]:





256

In [54]:

    

model.compile(optimizer = 'Adam',
             loss = 'categorical_crossentropy',
             metrics = ['accuracy'])


    

In [50]:

    

# Convert labels to categorical one-hot encoding
one_hot_labels = keras.utils.to_categorical(training_data[1], num_classes=10)


    

In [55]:

    

# Train the model, iterating on the data in batches of 32 samples
model.fit(training_data[0],one_hot_labels,epochs=10,batch_size=64)


    

    




Epoch 1/10
50000/50000 [==============================] - 90s - loss: 0.2435 - acc: 0.9269        
Epoch 2/10
50000/50000 [==============================] - 3s - loss: 0.0943 - acc: 0.9708     
Epoch 3/10
50000/50000 [==============================] - 3s - loss: 0.0633 - acc: 0.9802     
Epoch 4/10
50000/50000 [==============================] - 3s - loss: 0.0447 - acc: 0.9858     
Epoch 5/10
50000/50000 [==============================] - 3s - loss: 0.0368 - acc: 0.9880     
Epoch 6/10
50000/50000 [==============================] - 3s - loss: 0.0264 - acc: 0.9911     
Epoch 7/10
50000/50000 [==============================] - 3s - loss: 0.0232 - acc: 0.9927     
Epoch 8/10
50000/50000 [==============================] - 3s - loss: 0.0208 - acc: 0.9928     
Epoch 9/10
50000/50000 [==============================] - 3s - loss: 0.0204 - acc: 0.9933         
Epoch 10/10
50000/50000 [==============================] - 3s - loss: 0.0150 - acc: 0.9953     






    
Out[55]:





<keras.callbacks.History at 0x7f895a3f1a58>

In [56]:

    

y_test = keras.utils.to_categorical(test_data[1],num_classes=10)


    

In [57]:

    

score = model.evaluate(test_data[0],y_test)


    

    




 9536/10000 [===========================>..] - ETA: 0s 

In [58]:

    

print('Test loss: ',score[0])


    

    




Test loss:  0.139833982974

In [59]:

    

print('Test accuracy:',score[1])


    

    




Test accuracy: 0.9671

In [175]:

    

def make_subset(data: tuple,numbers: list):
    """
    Takes a list of numbers and returns the part of the MNIST dataset corresponding to these numbers. 
    """
    if isinstance(numbers,int):
        numbers = [numbers]
    idx = [np.squeeze(np.argwhere(data[1]==i)) for i in numbers]
    idx = np.sort(np.concatenate( idx, axis=0 ))
    return data[0][idx], data[1][idx]


    

In [210]:

    

model = Sequential([
    Dense(256, input_shape=(784,)),
    Activation('relu'),
    Dense(256),
    Activation('relu'),
    Dense(10),
    Activation('softmax')
])


    

In [160]:

    




    

In [186]:

    




    

In [226]:

    

task = 0
score_task1=[0,0]
score_task2=[0,0]
for i in range(0,4,2):
        
    train_x, train_y = make_subset(training_data,[i,i+1])

    one_hot_labels = keras.utils.to_categorical(train_y, num_classes=10)

    model.compile(optimizer = 'Adam',
             loss = 'categorical_crossentropy',
             metrics = ['accuracy'])
    model.fit(train_x,one_hot_labels,epochs=10,batch_size=64)
    
    # first, test on task 1 
    test_x, test_y = make_subset(test_data,[0,1])
    test_y = keras.utils.to_categorical(test_y,num_classes=10)
    score_task1[task] = model.evaluate(test_x,test_y)
    
    # then, test on task 2
    test_x, test_y = make_subset(test_data,[2,3])
    test_y = keras.utils.to_categorical(test_y,num_classes=10)
    score_task2[task] = model.evaluate(test_x,test_y)

    task += 1


    

    




Epoch 1/10
10610/10610 [==============================] - 1s - loss: 0.0011 - acc: 0.9998          
Epoch 2/10
10610/10610 [==============================] - 0s - loss: 2.0129e-07 - acc: 1.0000     
Epoch 3/10
10610/10610 [==============================] - 0s - loss: 1.1947e-07 - acc: 1.0000     
Epoch 4/10
10610/10610 [==============================] - 0s - loss: 1.1932e-07 - acc: 1.0000     
Epoch 5/10
10610/10610 [==============================] - 0s - loss: 1.1931e-07 - acc: 1.0000     
Epoch 6/10
10610/10610 [==============================] - 0s - loss: 1.1930e-07 - acc: 1.0000     
Epoch 7/10
10610/10610 [==============================] - 0s - loss: 1.1929e-07 - acc: 1.0000     
Epoch 8/10
10610/10610 [==============================] - 0s - loss: 1.1928e-07 - acc: 1.0000     
Epoch 9/10
10610/10610 [==============================] - 0s - loss: 1.1928e-07 - acc: 1.0000     
Epoch 10/10
10610/10610 [==============================] - 0s - loss: 1.1927e-07 - acc: 1.0000     
1856/2042 [==========================>...] - ETA: 0sEpoch 1/10
10069/10069 [==============================] - 0s - loss: 1.2005 - acc: 0.8997           
Epoch 2/10
10069/10069 [==============================] - 0s - loss: 0.0066 - acc: 0.9981     
Epoch 3/10
10069/10069 [==============================] - 0s - loss: 0.0015 - acc: 0.9997         
Epoch 4/10
10069/10069 [==============================] - 0s - loss: 5.3637e-04 - acc: 1.0000     
Epoch 5/10
10069/10069 [==============================] - 0s - loss: 1.8916e-04 - acc: 1.0000     
Epoch 6/10
10069/10069 [==============================] - 0s - loss: 1.0079e-04 - acc: 1.0000     
Epoch 7/10
10069/10069 [==============================] - 0s - loss: 5.2771e-05 - acc: 1.0000     
Epoch 8/10
10069/10069 [==============================] - 0s - loss: 3.7339e-05 - acc: 1.0000     
Epoch 9/10
10069/10069 [==============================] - 0s - loss: 2.6068e-05 - acc: 1.0000     
Epoch 10/10
10069/10069 [==============================] - 0s - loss: 1.8862e-05 - acc: 1.0000     
2042/2042 [==============================] - 0s     

In [227]:

    

score_task1


    

    
Out[227]:





[[0.0076209683508598001, 0.99952718676122931],
 [11.456930575359516, 0.011820330969267139]]

In [228]:

    

score_task2


    

    
Out[228]:





[[15.699201815508955, 0.0], [0.013092741012352358, 0.99461312479650588]]

In [240]:

    

test_x1, test_y1 = make_subset(test_data,[0,1])
test_x2, test_y2 = make_subset(test_data,[2,3])
test_y1 = keras.utils.to_categorical(test_y1,num_classes=10)
test_y2 = keras.utils.to_categorical(test_y2,num_classes=10)


    

In [251]:

    

plt.subplot(1,2,1)
plt.imshow(test_x1[3].reshape((28,28)))
plt.subplot(1,2,2)
plt.imshow(test_x1[5].reshape((28,28)))
plt.show()


    

In [ ]:

    




    

In [ ]:

    




    

In [ ]:

    




    

In [ ]: