In [1]:

    

import numpy as np
from keras.utils import np_utils
from scipy.misc import imresize
from sklearn.preprocessing import OneHotEncoder

seed = 1337
np.random.seed(seed)  # for reproducibility


    

    




Using TensorFlow backend.

In [2]:

    

train = np.load('train.npy')
test = np.load('test.npy')


    

In [3]:

    

x_train, y_train = np.invert(train[:, 0]), train[:, 1].reshape(train.shape[0], 1)


    

In [4]:

    

enc = OneHotEncoder()
Y_train = enc.fit_transform(y_train).toarray()


    

In [5]:

    

size = 64
img_rows = size
img_cols = size


    

In [6]:

    

for i in range(len(x_train)):
    x_train[i] = imresize(x_train[i], (img_rows, img_cols), interp='bicubic')
x_train = np.dstack(x_train)
x_train = np.transpose(x_train, axes=(2, 0, 1))
x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
X_train = x_train.astype('float32')
X_train -= np.mean(X_train)
X_train /= np.std(X_train)


    

In [7]:

    

for i in range(len(test)):
    test[i] = imresize(np.invert(test[i]), (img_rows, img_cols), interp='bicubic')
test = np.dstack(test)
test = np.transpose(test, axes=(2, 0, 1))
test = test.reshape(test.shape[0], img_rows, img_cols, 1).astype("float32")
test -= np.mean(test)
test /= np.std(test)


    

In [9]:

    

input_shape = (img_rows, img_cols, 1)
nb_classes = 500
batch_size = 128
nb_epoch = 30

from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten, BatchNormalization, PReLU
from keras.layers import Convolution2D, MaxPooling2D
from keras.optimizers import *

model = Sequential()

model.add(Dropout(0.1, input_shape=input_shape))

model.add(Convolution2D(64, 5, 5, border_mode='same'))
model.add(PReLU())
model.add(BatchNormalization(axis=3))

model.add(Convolution2D(64, 5, 5, border_mode='same'))
model.add(PReLU())
model.add(BatchNormalization(axis=3))

model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
model.add(Dropout(0.25))

model.add(Convolution2D(128, 5, 5, border_mode='same'))
model.add(PReLU())
model.add(BatchNormalization(axis=3))

model.add(Convolution2D(128, 5, 5, border_mode='same'))
model.add(PReLU())
model.add(BatchNormalization(axis=3))

model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
model.add(Dropout(0.25))

model.add(Convolution2D(256, 5, 5, border_mode='same'))
model.add(PReLU())
model.add(BatchNormalization(axis=3))

model.add(MaxPooling2D(pool_size=(3, 3), strides=(2, 2)))

model.add(Flatten())

model.add(Dropout(0.5))
model.add(Dense(2000))
model.add(PReLU())
model.add(Dropout(0.5))
model.add(Dense(2000))
model.add(PReLU())
model.add(Dropout(0.5))
model.add(Dense(nb_classes))
model.add(Activation('softmax'))

model.summary()
opt = Adadelta()
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])


    

    




____________________________________________________________________________________________________
Layer (type)                     Output Shape          Param #     Connected to                     
====================================================================================================
dropout_7 (Dropout)              (None, 64, 64, 1)     0           dropout_input_2[0][0]            
____________________________________________________________________________________________________
convolution2d_6 (Convolution2D)  (None, 64, 64, 64)    1664        dropout_7[0][0]                  
____________________________________________________________________________________________________
prelu_8 (PReLU)                  (None, 64, 64, 64)    262144      convolution2d_6[0][0]            
____________________________________________________________________________________________________
batchnormalization_6 (BatchNorma (None, 64, 64, 64)    256         prelu_8[0][0]                    
____________________________________________________________________________________________________
convolution2d_7 (Convolution2D)  (None, 64, 64, 64)    102464      batchnormalization_6[0][0]       
____________________________________________________________________________________________________
prelu_9 (PReLU)                  (None, 64, 64, 64)    262144      convolution2d_7[0][0]            
____________________________________________________________________________________________________
batchnormalization_7 (BatchNorma (None, 64, 64, 64)    256         prelu_9[0][0]                    
____________________________________________________________________________________________________
maxpooling2d_4 (MaxPooling2D)    (None, 32, 32, 64)    0           batchnormalization_7[0][0]       
____________________________________________________________________________________________________
dropout_8 (Dropout)              (None, 32, 32, 64)    0           maxpooling2d_4[0][0]             
____________________________________________________________________________________________________
convolution2d_8 (Convolution2D)  (None, 32, 32, 128)   204928      dropout_8[0][0]                  
____________________________________________________________________________________________________
prelu_10 (PReLU)                 (None, 32, 32, 128)   131072      convolution2d_8[0][0]            
____________________________________________________________________________________________________
batchnormalization_8 (BatchNorma (None, 32, 32, 128)   512         prelu_10[0][0]                   
____________________________________________________________________________________________________
convolution2d_9 (Convolution2D)  (None, 32, 32, 128)   409728      batchnormalization_8[0][0]       
____________________________________________________________________________________________________
prelu_11 (PReLU)                 (None, 32, 32, 128)   131072      convolution2d_9[0][0]            
____________________________________________________________________________________________________
batchnormalization_9 (BatchNorma (None, 32, 32, 128)   512         prelu_11[0][0]                   
____________________________________________________________________________________________________
maxpooling2d_5 (MaxPooling2D)    (None, 16, 16, 128)   0           batchnormalization_9[0][0]       
____________________________________________________________________________________________________
dropout_9 (Dropout)              (None, 16, 16, 128)   0           maxpooling2d_5[0][0]             
____________________________________________________________________________________________________
convolution2d_10 (Convolution2D) (None, 16, 16, 256)   819456      dropout_9[0][0]                  
____________________________________________________________________________________________________
prelu_12 (PReLU)                 (None, 16, 16, 256)   65536       convolution2d_10[0][0]           
____________________________________________________________________________________________________
batchnormalization_10 (BatchNorm (None, 16, 16, 256)   1024        prelu_12[0][0]                   
____________________________________________________________________________________________________
maxpooling2d_6 (MaxPooling2D)    (None, 7, 7, 256)     0           batchnormalization_10[0][0]      
____________________________________________________________________________________________________
flatten_2 (Flatten)              (None, 12544)         0           maxpooling2d_6[0][0]             
____________________________________________________________________________________________________
dropout_10 (Dropout)             (None, 12544)         0           flatten_2[0][0]                  
____________________________________________________________________________________________________
dense_4 (Dense)                  (None, 2000)          25090000    dropout_10[0][0]                 
____________________________________________________________________________________________________
prelu_13 (PReLU)                 (None, 2000)          2000        dense_4[0][0]                    
____________________________________________________________________________________________________
dropout_11 (Dropout)             (None, 2000)          0           prelu_13[0][0]                   
____________________________________________________________________________________________________
dense_5 (Dense)                  (None, 2000)          4002000     dropout_11[0][0]                 
____________________________________________________________________________________________________
prelu_14 (PReLU)                 (None, 2000)          2000        dense_5[0][0]                    
____________________________________________________________________________________________________
dropout_12 (Dropout)             (None, 2000)          0           prelu_14[0][0]                   
____________________________________________________________________________________________________
dense_6 (Dense)                  (None, 500)           1000500     dropout_12[0][0]                 
____________________________________________________________________________________________________
activation_2 (Activation)        (None, 500)           0           dense_6[0][0]                    
====================================================================================================
Total params: 32,489,268
Trainable params: 32,487,988
Non-trainable params: 1,280
____________________________________________________________________________________________________

In [ ]:

    

model.fit(X_train, Y_train, verbose=0, batch_size=batch_size, nb_epoch=nb_epoch)


    

In [ ]:

    

from keras.preprocessing.image import ImageDataGenerator

datagen = ImageDataGenerator(rotation_range=10, width_shift_range=0.1, height_shift_range=0.1, zoom_range=0.1,
                             fill_mode='constant')

model.fit_generator(datagen.flow(X_train, Y_train, batch_size=batch_size), samples_per_epoch=len(X_train), verbose=0,
                    nb_epoch=nb_epoch)


    

In [ ]:

    

model.fit(X_train, Y_train, verbose=0, batch_size=96, nb_epoch=10)


    

In [ ]:

    

model.evaluate(X_train, Y_train, verbose=0)[1]


    

In [ ]:

    

pred_cl = model.predict_classes(test, verbose=0)


    

In [ ]:

    

write_frame = ["{0},{1}".format(i + 1, enc.active_features_[pred_cl[i]]) for i in range(len(pred_cl))]
write_frame = ["Id,Category"] + write_frame
with open("predictions.csv", "w+") as f:
    f.write("\n".join(write_frame))


    

In [ ]: