In [1]:

    

import keras.backend as K
from data_manager import ClutteredMNIST
from visualizer import plot_mnist_sample
from visualizer import print_evaluation
from visualizer import plot_mnist_grid
from models import STN


    

    




Using TensorFlow backend.

In [2]:

    

dataset_path = "../datasets/mnist_cluttered_60x60_6distortions.npz"
batch_size = 256
num_epochs = 30

data_manager = ClutteredMNIST(dataset_path)
train_data, val_data, test_data = data_manager.load()
x_train, y_train = train_data
plot_mnist_sample(x_train[7])


    

In [3]:

    

model = STN()
model.compile(loss='categorical_crossentropy', optimizer='adam')
model.summary()


    

    




__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            (None, 60, 60, 1)    0                                            
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D)  (None, 30, 30, 1)    0           input_1[0][0]                    
__________________________________________________________________________________________________
conv2d_1 (Conv2D)               (None, 26, 26, 20)   520         max_pooling2d_1[0][0]            
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D)  (None, 13, 13, 20)   0           conv2d_1[0][0]                   
__________________________________________________________________________________________________
conv2d_2 (Conv2D)               (None, 9, 9, 20)     10020       max_pooling2d_2[0][0]            
__________________________________________________________________________________________________
flatten_1 (Flatten)             (None, 1620)         0           conv2d_2[0][0]                   
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, 50)           81050       flatten_1[0][0]                  
__________________________________________________________________________________________________
activation_1 (Activation)       (None, 50)           0           dense_1[0][0]                    
__________________________________________________________________________________________________
dense_2 (Dense)                 (None, 6)            306         activation_1[0][0]               
__________________________________________________________________________________________________
bilinear_interpolation_1 (Bilin (None, 30, 30, 1)    0           input_1[0][0]                    
                                                                 dense_2[0][0]                    
__________________________________________________________________________________________________
conv2d_3 (Conv2D)               (None, 30, 30, 32)   320         bilinear_interpolation_1[0][0]   
__________________________________________________________________________________________________
activation_2 (Activation)       (None, 30, 30, 32)   0           conv2d_3[0][0]                   
__________________________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D)  (None, 15, 15, 32)   0           activation_2[0][0]               
__________________________________________________________________________________________________
conv2d_4 (Conv2D)               (None, 13, 13, 32)   9248        max_pooling2d_3[0][0]            
__________________________________________________________________________________________________
activation_3 (Activation)       (None, 13, 13, 32)   0           conv2d_4[0][0]                   
__________________________________________________________________________________________________
max_pooling2d_4 (MaxPooling2D)  (None, 6, 6, 32)     0           activation_3[0][0]               
__________________________________________________________________________________________________
flatten_2 (Flatten)             (None, 1152)         0           max_pooling2d_4[0][0]            
__________________________________________________________________________________________________
dense_3 (Dense)                 (None, 256)          295168      flatten_2[0][0]                  
__________________________________________________________________________________________________
activation_4 (Activation)       (None, 256)          0           dense_3[0][0]                    
__________________________________________________________________________________________________
dense_4 (Dense)                 (None, 10)           2570        activation_4[0][0]               
__________________________________________________________________________________________________
activation_5 (Activation)       (None, 10)           0           dense_4[0][0]                    
==================================================================================================
Total params: 399,202
Trainable params: 399,202
Non-trainable params: 0
__________________________________________________________________________________________________

In [6]:

    

input_image = model.input
output_STN = model.get_layer('bilinear_interpolation_1').output
STN_function = K.function([input_image], [output_STN])

for epoch_arg in range(num_epochs):
    for batch_arg in range(150):
        arg_0 = batch_arg * batch_size
        arg_1 = (batch_arg + 1) * batch_size
        x_batch, y_batch = x_train[arg_0:arg_1], y_train[arg_0:arg_1]
        loss = model.train_on_batch(x_batch, y_batch)
    if epoch_arg % 10 == 0:
        val_score = model.evaluate(*val_data, verbose=1)
        test_score = model.evaluate(*test_data, verbose=1)
        print_evaluation(epoch_arg, val_score, test_score)
        plot_mnist_grid(x_batch, STN_function)
        print('-' * 40)


    

    




10000/10000 [==============================] - 0s 49us/step
10000/10000 [==============================] - 0s 49us/step
Epoch: 0 | Val: 0.17136050443053247 | Test: 0.17655548879206182






    













    




----------------------------------------
10000/10000 [==============================] - 1s 55us/step
10000/10000 [==============================] - 1s 54us/step
Epoch: 10 | Val: 0.17013455713624134 | Test: 0.17813963381517678






    













    




----------------------------------------
10000/10000 [==============================] - 0s 49us/step
10000/10000 [==============================] - 1s 55us/step
Epoch: 20 | Val: 0.19145010298341514 | Test: 0.1978291994124651






    













    




----------------------------------------

In [7]:

    

plot_mnist_grid(x_train)
plot_mnist_grid(x_train, STN_function)