In [1]:
from __future__ import print_function
#Basic libraries
import numpy as np
import tensorflow as tf
print('Tensorflow version: ', tf.__version__)
import time
#Show images
import matplotlib.pyplot as plt
%matplotlib inline
# plt configuration
plt.rcParams['figure.figsize'] = (10, 10) # size of images
plt.rcParams['image.interpolation'] = 'nearest' # show exact image
plt.rcParams['image.cmap'] = 'gray' # use grayscale
# Select GPU
import os
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"]="0"
In [2]:
# Import mnist dataset and rescale between [0,1]
from tensorflow.contrib.keras import datasets
(X_train, y_train), (X_test, y_test) = datasets.mnist.load_data(path='mnist.npz')
X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255.
X_test /= 255.
print('X train shape: ',X_train.shape)
print('y train shape: ',y_train.shape)
print('X test shape: ', X_test.shape)
print('y test shape: ', y_test.shape)
In [3]:
# Reshape (the input format is vector)
X_train = np.reshape(X_train, (len(X_train),28,28,1))
X_test = np.reshape(X_test, (len(X_test) ,28,28,1))
print('X train shape: ',X_train.shape)
print('X test shape: ', X_test.shape)
In [4]:
#Select a random subsample of the train data for each class
def create_sample(x, y, sample_size=100):
class_list = np.unique(y_train)
sample_ix = []
for i in class_list:
class_ix = np.where(y==i)[0]
class_ix_sample = np.random.choice(class_ix, sample_size)
sample_ix += list(class_ix_sample)
#Shuffle sample
sample_ix = np.random.permutation(sample_ix)
return x[sample_ix], y[sample_ix]
#Create new sample datasets
X_train_sample, y_train_sample = create_sample(X_train, y_train, sample_size=100)
print('X train sample shape: ',X_train_sample.shape)
print('y train sample shape: ',y_train_sample.shape)
In [5]:
#Examine the data
fig = plt.figure()
for n, i in enumerate(range(32)):
a = fig.add_subplot(4,8,n+1)
a.set_title(str(y_train_sample[i]))
fig.tight_layout()
plt.imshow(np.reshape(X_train_sample[i],(28,28)))
In [6]:
#Define the architecture
from tensorflow.contrib.keras import models, layers, optimizers, callbacks, preprocessing
print('Creating Lenet architecture.')
#Inputs
images = layers.Input(batch_shape=(None, 28, 28, 1), dtype='float32', name='Images')
#First convolutional layer
conv1 = layers.Conv2D(20, (5,5))(images)
pool1 = layers.MaxPooling2D(pool_size=(2, 2))(conv1)
#Second convolutional layer
conv2 = layers.Conv2D(20, (5,5))(pool1)
pool2 = layers.MaxPooling2D(pool_size=(2, 2))(conv2)
#Flatten convolution output
flat = layers.Flatten(name='Flat_image')(pool2)
# First dense layer
dense1 = layers.Dense(500, activation='relu', name='Dense_1')(flat)
# Second dense layer
output = layers.Dense(10, activation='softmax', name='Dense_output')(dense1)
In [7]:
# Configure input and output data for the Model
model_conv = models.Model(inputs=images, outputs=output)
model_conv.summary()
# Select optimizer and compile model
sgd_optimizer = optimizers.SGD(lr=0.01, momentum=0.99, decay=0.005, nesterov=True)
model_conv.compile(loss='sparse_categorical_crossentropy',
optimizer=sgd_optimizer, metrics=['accuracy'])
In [8]:
#Fit the model
batch_size = 64
epochs = 30
# Tensorflow callback
tb_callback_ln = callbacks.TensorBoard(log_dir='/tmp/tensorboard/mnist/conv_aug1')
history_conv = model_conv.fit(X_train_sample, y_train_sample, batch_size=batch_size, epochs=epochs,
validation_data=(X_test, y_test), callbacks=[tb_callback_ln])
In [9]:
plt.plot(history_conv.history['acc'], label='acc')
plt.plot(history_conv.history['val_acc'], label='val acc')
plt.legend(loc='lower right')
plt.show()
In [ ]:
In [10]:
print('Second model. The same Lenet architecture')
#Inputs
images = layers.Input(batch_shape=(None, 28, 28, 1), dtype='float32', name='Images')
#First convolutional layer
conv1 = layers.Conv2D(20, (5,5))(images)
pool1 = layers.MaxPooling2D(pool_size=(2, 2))(conv1)
#Second convolutional layer
conv2 = layers.Conv2D(20, (5,5))(pool1)
pool2 = layers.MaxPooling2D(pool_size=(2, 2))(conv2)
#Flatten convolution output
flat = layers.Flatten(name='Flat_image')(pool2)
# First dense layer
dense1 = layers.Dense(500, activation='relu', name='Dense_1')(flat)
# Second dense layer
output = layers.Dense(10, activation='softmax', name='Dense_output')(dense1)
In [11]:
# Define inputs - outputs
model_aug = models.Model(inputs=images, outputs=output)
model_aug.summary()
# Select optimizer and compile model
sgd_optimizer = optimizers.SGD(lr=0.01, momentum=0.99, decay=0.005, nesterov=True)
model_aug.compile(loss='sparse_categorical_crossentropy',
optimizer=sgd_optimizer, metrics=['accuracy'])
In [12]:
# Augmentation for training
train_datagen = preprocessing.image.ImageDataGenerator(
#rotation_range=0.25,
width_shift_range=0.2,
height_shift_range=0.12,
shear_range=0.3,
zoom_range=[0.9,1.3])
In [13]:
#Visualize the ugmented images
fig = plt.figure()
for n, i in enumerate(range(32)):
a = fig.add_subplot(4,8,n+1)
a.set_title(str(y_train_sample[i]))
fig.tight_layout()
plt.imshow(np.reshape(train_datagen.random_transform(X_train_sample[i]),(28,28)))
In [14]:
batch_size = 64
epochs = 30
tb_callback_ln = callbacks.TensorBoard(log_dir='/tmp/tensorboard/mnist/conv_aug2')
history_aug = model_aug.fit_generator(train_datagen.flow(X_train_sample, y_train_sample, batch_size=batch_size),
steps_per_epoch=len(X_train) // batch_size,
epochs=epochs,
validation_data=(X_test, y_test),
callbacks=[tb_callback_ln])
In [15]:
# Plot history
plt.plot(history_aug.history['acc'], label='acc')
plt.plot(history_aug.history['val_acc'], label='val acc')
plt.legend(loc='lower right')
plt.show()
In [ ]:
In [16]:
print('Model 3. The same Lenet architecture')
#Inputs
images = layers.Input(batch_shape=(None, 28, 28, 1), dtype='float32', name='Images')
#First convolutional layer
conv1 = layers.Conv2D(20, (5,5))(images)
pool1 = layers.MaxPooling2D(pool_size=(2, 2))(conv1)
#Second convolutional layer
conv2 = layers.Conv2D(20, (5,5))(pool1)
pool2 = layers.MaxPooling2D(pool_size=(2, 2))(conv2)
#Flatten convolution output
flat = layers.Flatten(name='Flat_image')(pool2)
# First dense layer
dense1 = layers.Dense(500, activation='relu', name='Dense_1')(flat)
# Second dense layer
output = layers.Dense(10, activation='softmax', name='Dense_output')(dense1)
In [17]:
# Model Architecture defined
model_aug2 = models.Model(inputs=images, outputs=output)
model_aug2.summary()
# Select optimizer and compile model
sgd_optimizer = optimizers.SGD(lr=0.01, momentum=0.99, decay=0.005, nesterov=True)
model_aug2.compile(loss='sparse_categorical_crossentropy',
optimizer=sgd_optimizer, metrics=['accuracy'])
In [18]:
# Augmentation for training
train_datagen = preprocessing.image.ImageDataGenerator(
rotation_range=0.25,
width_shift_range=0.2,
height_shift_range=0.12,
shear_range=0.3,
zoom_range=[0.9,1.3])
In [19]:
def img_augmented(img):
import random
import cv2
#Erode - dilate
if random.random() < 0.7:
img = cv2.erode(img, np.ones(2, np.uint8), iterations=1)
elif random.random() < 0.7:
img = cv2.dilate(img, np.ones(2, np.uint8), iterations=1)
return img
def data_generator(x, y, batch_size=32):
for x_batch, y_batch in train_datagen.flow(x, y, batch_size=batch_size):
x_batch_transform=[]
for img in x_batch:
x_batch_transform += [np.reshape(img_augmented(img),(28,28,1))]
yield np.array(x_batch_transform), y_batch
In [20]:
#Visualize data augmentation
dg = data_generator(X_train_sample, y_train_sample, batch_size=32)
x_sample, y_sample = next(dg)
#Examine the data
fig = plt.figure()
for n, i in enumerate(range(32)):
a = fig.add_subplot(4,8,n+1)
a.set_title(str(y_sample[i]))
fig.tight_layout()
plt.imshow(np.reshape(x_sample[i],(28,28)))
In [21]:
batch_size = 64
epochs = 30
tb_callback_ln = callbacks.TensorBoard(log_dir='/tmp/tensorboard/mnist/conv_aug2')
history_aug2 = model_aug2.fit_generator(data_generator(X_train_sample, y_train_sample, batch_size=batch_size),
steps_per_epoch=len(X_train) // batch_size,
epochs=epochs,
validation_data=(X_test, y_test),
callbacks=[tb_callback_ln])
In [22]:
# Compare test
plt.plot(history_aug.history['val_acc'] , label='val acc aug1')
plt.plot(history_aug2.history['val_acc'], label='val acc aug2')
plt.legend(loc='lower right')
plt.show()
In [ ]: