TensorFlow学习日记22

1. MNIST MLP

解析：

'''
Trains a simple deep NN on the MNIST dataset.
'''

from __future__ import print_function

import keras
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout
from keras.optimizers import RMSprop

batch_size = 128
num_classes = 10
epochs = 20

# the data, shuffled and split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()

x_train = x_train.reshape(60000, 784)
x_test = x_test.reshape(10000, 784)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')

# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)

model = Sequential()
model.add(Dense(512, activation='relu', input_shape=(784,)))
model.add(Dropout(0.2))
model.add(Dense(512, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(num_classes, activation='softmax'))

model.summary()

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

history = model.fit(x_train, y_train,
                    batch_size=batch_size,
                    epochs=epochs,
                    verbose=1,
                    validation_data=(x_test, y_test))
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])

2. MNIST CNN
解析：

'''
Trains a simple convnet on the MNIST dataset.
'''

from __future__ import print_function
import keras
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D
from keras import backend as K

batch_size = 128
num_classes = 10
epochs = 12

# input image dimensions
img_rows, img_cols = 28, 28

# the data, shuffled and split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()

if K.image_data_format() == 'channels_first':
    x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
    x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
    input_shape = (1, img_rows, img_cols)
else:
    x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
    x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
    input_shape = (img_rows, img_cols, 1)

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(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')

# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)

model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
                 activation='relu',
                 input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))

model.compile(loss=keras.losses.categorical_crossentropy,
              optimizer=keras.optimizers.Adadelta(),
              metrics=['accuracy'])

model.fit(x_train, y_train,
          batch_size=batch_size,
          epochs=epochs,
          verbose=1,
          validation_data=(x_test, y_test))
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])

3. CIFAR10 CNN

解析：

'''
Train a simple deep CNN on the CIFAR10 small images dataset.
'''

from __future__ import print_function
import keras
from keras.datasets import cifar10
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten
from keras.layers import Conv2D, MaxPooling2D
import numpy as np
import os

batch_size = 32
num_classes = 10
epochs = 200
data_augmentation = True
num_predictions = 20
save_dir = os.path.join(os.getcwd(), 'saved_models')
model_name = 'keras_cifar10_trained_model.h5'

# The data, shuffled and split between train and test sets:
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')

# Convert class vectors to binary class matrices.
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)

model = Sequential()
model.add(Conv2D(32, (3, 3), padding='same',
                 input_shape=x_train.shape[1:]))
model.add(Activation('relu'))
model.add(Conv2D(32, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))

model.add(Conv2D(64, (3, 3), padding='same'))
model.add(Activation('relu'))
model.add(Conv2D(64, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))

model.add(Flatten())
model.add(Dense(512))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes))
model.add(Activation('softmax'))

# initiate RMSprop optimizer
opt = keras.optimizers.rmsprop(lr=0.0001, decay=1e-6)

# Let's train the model using RMSprop
model.compile(loss='categorical_crossentropy',
              optimizer=opt,
              metrics=['accuracy'])

x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255

if not data_augmentation:
    print('Not using data augmentation.')
    model.fit(x_train, y_train,
              batch_size=batch_size,
              epochs=epochs,
              validation_data=(x_test, y_test),
              shuffle=True)
else:
    print('Using real-time data augmentation.')
    # This will do preprocessing and realtime data augmentation:
    datagen = ImageDataGenerator(
        featurewise_center=False,  # set input mean to 0 over the dataset
        samplewise_center=False,  # set each sample mean to 0
        featurewise_std_normalization=False,  # divide inputs by std of the dataset
        samplewise_std_normalization=False,  # divide each input by its std
        zca_whitening=False,  # apply ZCA whitening
        rotation_range=0,  # randomly rotate images in the range (degrees, 0 to 180)
        width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)
        height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)
        horizontal_flip=True,  # randomly flip images
        vertical_flip=False)  # randomly flip images

    # Compute quantities required for feature-wise normalization
    # (std, mean, and principal components if ZCA whitening is applied).
    datagen.fit(x_train)

    # Fit the model on the batches generated by datagen.flow().
    model.fit_generator(datagen.flow(x_train, y_train,
                                     batch_size=batch_size),
                        steps_per_epoch=int(np.ceil(x_train.shape[0] / float(batch_size))),
                        epochs=epochs,
                        validation_data=(x_test, y_test),
                        workers=4)

# Save model and weights
if not os.path.isdir(save_dir):
    os.makedirs(save_dir)
model_path = os.path.join(save_dir, model_name)
model.save(model_path)
print('Saved trained model at %s ' % model_path)

# Score trained model.
scores = model.evaluate(x_test, y_test, verbose=1)
print('Test loss:', scores[0])
print('Test accuracy:', scores[1])

4. CIFAR10 ResNet

解析：

"""
Trains a ResNet on the CIFAR10 dataset.
"""

from __future__ import print_function
import keras
from keras.layers import Dense, Conv2D, BatchNormalization, Activation
from keras.layers import MaxPooling2D, AveragePooling2D, Input, Flatten
from keras.optimizers import Adam
from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau
from keras.preprocessing.image import ImageDataGenerator
from keras.regularizers import l2
from keras import backend as K
from keras.models import Model
from keras.datasets import cifar10
import numpy as np
import os

# Training params.
batch_size = 32
epochs = 100
data_augmentation = True

# Network architecture params.
num_classes = 10
num_filters = 64
num_blocks = 4
num_sub_blocks = 2
use_max_pool = False

# Load the CIFAR10 data.
(x_train, y_train), (x_test, y_test) = cifar10.load_data()

# Input image dimensions.
# We assume data format "channels_last".
img_rows = x_train.shape[1]
img_cols = x_train.shape[2]
channels = x_train.shape[3]

if K.image_data_format() == 'channels_first':
    img_rows = x_train.shape[2]
    img_cols = x_train.shape[3]
    channels = x_train.shape[1]
    x_train = x_train.reshape(x_train.shape[0], channels, img_rows, img_cols)
    x_test = x_test.reshape(x_test.shape[0], channels, img_rows, img_cols)
    input_shape = (channels, img_rows, img_cols)
else:
    img_rows = x_train.shape[1]
    img_cols = x_train.shape[2]
    channels = x_train.shape[3]
    x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, channels)
    x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, channels)
    input_shape = (img_rows, img_cols, channels)

# Normalize data.
x_train = x_train.astype('float32') / 255
x_test = x_test.astype('float32') / 255
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
print('y_train shape:', y_train.shape)

# Convert class vectors to binary class matrices.
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)

# Start model definition.
inputs = Input(shape=input_shape)
x = Conv2D(num_filters,
           kernel_size=7,
           padding='same',
           strides=2,
           kernel_initializer='he_normal',
           kernel_regularizer=l2(1e-4))(inputs)
x = BatchNormalization()(x)
x = Activation('relu')(x)

# Orig paper uses max pool after 1st conv.
# Reaches up 87% acc if use_max_pool = True.
# Cifar10 images are already too small at 32x32 to be maxpooled. So, we skip.
if use_max_pool:
    x = MaxPooling2D(pool_size=3, strides=2, padding='same')(x)
    num_blocks = 3

# Instantiate convolutional base (stack of blocks).
for i in range(num_blocks):
    for j in range(num_sub_blocks):
        strides = 1
        is_first_layer_but_not_first_block = j == 0 and i > 0
        if is_first_layer_but_not_first_block:
            strides = 2
        y = Conv2D(num_filters,
                   kernel_size=3,
                   padding='same',
                   strides=strides,
                   kernel_initializer='he_normal',
                   kernel_regularizer=l2(1e-4))(x)
        y = BatchNormalization()(y)
        y = Activation('relu')(y)
        y = Conv2D(num_filters,
                   kernel_size=3,
                   padding='same',
                   kernel_initializer='he_normal',
                   kernel_regularizer=l2(1e-4))(y)
        y = BatchNormalization()(y)
        if is_first_layer_but_not_first_block:
            x = Conv2D(num_filters,
                       kernel_size=1,
                       padding='same',
                       strides=2,
                       kernel_initializer='he_normal',
                       kernel_regularizer=l2(1e-4))(x)
        x = keras.layers.add([x, y])
        x = Activation('relu')(x)

    num_filters = 2 * num_filters

# Add classifier on top.
x = AveragePooling2D()(x)
y = Flatten()(x)
outputs = Dense(num_classes,
                activation='softmax',
                kernel_initializer='he_normal')(y)

# Instantiate and compile model.
model = Model(inputs=inputs, outputs=outputs)
model.compile(loss='categorical_crossentropy',
              optimizer=Adam(),
              metrics=['accuracy'])
model.summary()

# Prepare model model saving directory.
save_dir = os.path.join(os.getcwd(), 'saved_models')
model_name = 'cifar10_resnet_model.h5'
if not os.path.isdir(save_dir):
    os.makedirs(save_dir)
filepath = os.path.join(save_dir, model_name)

# Prepare callbacks for model saving and for learning rate decaying.
checkpoint = ModelCheckpoint(filepath=filepath,
                             verbose=1,
                             save_best_only=True)
lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1),
                               cooldown=0,
                               patience=5,
                               min_lr=0.5e-6)
callbacks = [checkpoint, lr_reducer]

# Run training, with or without data augmentation.
if not data_augmentation:
    print('Not using data augmentation.')
    model.fit(x_train, y_train,
              batch_size=batch_size,
              epochs=epochs,
              validation_data=(x_test, y_test),
              shuffle=True,
              callbacks=callbacks)
else:
    print('Using real-time data augmentation.')
    # This will do preprocessing and realtime data augmentation:
    datagen = ImageDataGenerator(
        featurewise_center=False,  # set input mean to 0 over the dataset
        samplewise_center=False,  # set each sample mean to 0
        featurewise_std_normalization=False,  # divide inputs by std of the dataset
        samplewise_std_normalization=False,  # divide each input by its std
        zca_whitening=False,  # apply ZCA whitening
        rotation_range=0,  # randomly rotate images in the range (degrees, 0 to 180)
        width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)
        height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)
        horizontal_flip=True,  # randomly flip images
        vertical_flip=False)  # randomly flip images

    # Compute quantities required for featurewise normalization
    # (std, mean, and principal components if ZCA whitening is applied).
    datagen.fit(x_train)

    # Fit the model on the batches generated by datagen.flow().
    model.fit_generator(datagen.flow(x_train, y_train, batch_size=batch_size),
                        steps_per_epoch=int(np.ceil(x_train.shape[0] / float(batch_size))),
                        validation_data=(x_test, y_test),
                        epochs=epochs, verbose=1, workers=4,
                        callbacks=callbacks)

# Score trained model.
scores = model.evaluate(x_test, y_test, verbose=1)
print('Test loss:', scores[0])
print('Test accuracy:', scores[1])

5. CONV LSTM

解析：

"""
This script demonstrates the use of a convolutional LSTM network.
This network is used to predict the next frame of an artificially
generated movie which contains moving squares.
"""
from keras.models import Sequential
from keras.layers.convolutional import Conv3D
from keras.layers.convolutional_recurrent import ConvLSTM2D
from keras.layers.normalization import BatchNormalization
import numpy as np
import pylab as plt

# We create a layer which take as input movies of shape
# (n_frames, width, height, channels) and returns a movie
# of identical shape.

seq = Sequential()
seq.add(ConvLSTM2D(filters=40, kernel_size=(3, 3),
                   input_shape=(None, 40, 40, 1),
                   padding='same', return_sequences=True))
seq.add(BatchNormalization())

seq.add(ConvLSTM2D(filters=40, kernel_size=(3, 3),
                   padding='same', return_sequences=True))
seq.add(BatchNormalization())

seq.add(ConvLSTM2D(filters=40, kernel_size=(3, 3),
                   padding='same', return_sequences=True))
seq.add(BatchNormalization())

seq.add(ConvLSTM2D(filters=40, kernel_size=(3, 3),
                   padding='same', return_sequences=True))
seq.add(BatchNormalization())

seq.add(Conv3D(filters=1, kernel_size=(3, 3, 3),
               activation='sigmoid',
               padding='same', data_format='channels_last'))
seq.compile(loss='binary_crossentropy', optimizer='adadelta')


# Artificial data generation:
# Generate movies with 3 to 7 moving squares inside.
# The squares are of shape 1x1 or 2x2 pixels,
# which move linearly over time.
# For convenience we first create movies with bigger width and height (80x80)
# and at the end we select a 40x40 window.

def generate_movies(n_samples=1200, n_frames=15):
    row = 80
    col = 80
    noisy_movies = np.zeros((n_samples, n_frames, row, col, 1), dtype=np.float)
    shifted_movies = np.zeros((n_samples, n_frames, row, col, 1),
                              dtype=np.float)

    for i in range(n_samples):
        # Add 3 to 7 moving squares
        n = np.random.randint(3, 8)

        for j in range(n):
            # Initial position
            xstart = np.random.randint(20, 60)
            ystart = np.random.randint(20, 60)
            # Direction of motion
            directionx = np.random.randint(0, 3) - 1
            directiony = np.random.randint(0, 3) - 1

            # Size of the square
            w = np.random.randint(2, 4)

            for t in range(n_frames):
                x_shift = xstart + directionx * t
                y_shift = ystart + directiony * t
                noisy_movies[i, t, x_shift - w: x_shift + w,
                y_shift - w: y_shift + w, 0] += 1

                # Make it more robust by adding noise.
                # The idea is that if during inference,
                # the value of the pixel is not exactly one,
                # we need to train the network to be robust and still
                # consider it as a pixel belonging to a square.
                if np.random.randint(0, 2):
                    noise_f = (-1) ** np.random.randint(0, 2)
                    noisy_movies[i, t,
                    x_shift - w - 1: x_shift + w + 1,
                    y_shift - w - 1: y_shift + w + 1,
                    0] += noise_f * 0.1

                # Shift the ground truth by 1
                x_shift = xstart + directionx * (t + 1)
                y_shift = ystart + directiony * (t + 1)
                shifted_movies[i, t, x_shift - w: x_shift + w,
                y_shift - w: y_shift + w, 0] += 1

    # Cut to a 40x40 window
    noisy_movies = noisy_movies[::, ::, 20:60, 20:60, ::]
    shifted_movies = shifted_movies[::, ::, 20:60, 20:60, ::]
    noisy_movies[noisy_movies >= 1] = 1
    shifted_movies[shifted_movies >= 1] = 1
    return noisy_movies, shifted_movies


# Train the network
noisy_movies, shifted_movies = generate_movies(n_samples=1200)
seq.fit(noisy_movies[:1000], shifted_movies[:1000], batch_size=10,
        epochs=300, validation_split=0.05)

# Testing the network on one movie
# feed it with the first 7 positions and then
# predict the new positions
which = 1004
track = noisy_movies[which][:7, ::, ::, ::]

for j in range(16):
    new_pos = seq.predict(track[np.newaxis, ::, ::, ::, ::])
    new = new_pos[::, -1, ::, ::, ::]
    track = np.concatenate((track, new), axis=0)

# And then compare the predictions
# to the ground truth
track2 = noisy_movies[which][::, ::, ::, ::]
for i in range(15):
    fig = plt.figure(figsize=(10, 5))

    ax = fig.add_subplot(121)

    if i >= 7:
        ax.text(1, 3, 'Predictions !', fontsize=20, color='w')
    else:
        ax.text(1, 3, 'Initial trajectory', fontsize=20)

    toplot = track[i, ::, ::, 0]

    plt.imshow(toplot)
    ax = fig.add_subplot(122)
    plt.text(1, 3, 'Ground truth', fontsize=20)

    toplot = track2[i, ::, ::, 0]
    if i >= 2:
        toplot = shifted_movies[which][i - 1, ::, ::, 0]

    plt.imshow(toplot)
    plt.savefig('%i_animate.png' % (i + 1))

6. Image OCR

解析：

# -*- coding: utf-8 -*-
'''
This example uses a convolutional stack followed by a recurrent stack
and a CTC logloss function to perform optical character recognition
of generated text images.

'''
import os
import itertools
import codecs
import re
import datetime
import cairocffi as cairo
import editdistance
import numpy as np
from scipy import ndimage
import pylab
from keras import backend as K
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras.layers import Input, Dense, Activation
from keras.layers import Reshape, Lambda
from keras.layers.merge import add, concatenate
from keras.models import Model
from keras.layers.recurrent import GRU
from keras.optimizers import SGD
from keras.utils.data_utils import get_file
from keras.preprocessing import image
import keras.callbacks

OUTPUT_DIR = 'image_ocr'

# character classes and matching regex filter
regex = r'^[a-z ]+$'
alphabet = u'abcdefghijklmnopqrstuvwxyz '

np.random.seed(55)


# this creates larger "blotches" of noise which look
# more realistic than just adding gaussian noise
# assumes greyscale with pixels ranging from 0 to 1

def speckle(img):
    severity = np.random.uniform(0, 0.6)
    blur = ndimage.gaussian_filter(np.random.randn(*img.shape) * severity, 1)
    img_speck = (img + blur)
    img_speck[img_speck > 1] = 1
    img_speck[img_speck <= 0] = 0
    return img_speck


# paints the string in a random location the bounding box
# also uses a random font, a slight random rotation,
# and a random amount of speckle noise

def paint_text(text, w, h, rotate=False, ud=False, multi_fonts=False):
    surface = cairo.ImageSurface(cairo.FORMAT_RGB24, w, h)
    with cairo.Context(surface) as context:
        context.set_source_rgb(1, 1, 1)  # White
        context.paint()
        # this font list works in CentOS 7
        if multi_fonts:
            fonts = ['Century Schoolbook', 'Courier', 'STIX', 'URW Chancery L', 'FreeMono']
            context.select_font_face(np.random.choice(fonts), cairo.FONT_SLANT_NORMAL,
                                     np.random.choice([cairo.FONT_WEIGHT_BOLD, cairo.FONT_WEIGHT_NORMAL]))
        else:
            context.select_font_face('Courier', cairo.FONT_SLANT_NORMAL, cairo.FONT_WEIGHT_BOLD)
        context.set_font_size(25)
        box = context.text_extents(text)
        border_w_h = (4, 4)
        if box[2] > (w - 2 * border_w_h[1]) or box[3] > (h - 2 * border_w_h[0]):
            raise IOError('Could not fit string into image. Max char count is too large for given image width.')

        # teach the RNN translational invariance by
        # fitting text box randomly on canvas, with some room to rotate
        max_shift_x = w - box[2] - border_w_h[0]
        max_shift_y = h - box[3] - border_w_h[1]
        top_left_x = np.random.randint(0, int(max_shift_x))
        if ud:
            top_left_y = np.random.randint(0, int(max_shift_y))
        else:
            top_left_y = h // 2
        context.move_to(top_left_x - int(box[0]), top_left_y - int(box[1]))
        context.set_source_rgb(0, 0, 0)
        context.show_text(text)

    buf = surface.get_data()
    a = np.frombuffer(buf, np.uint8)
    a.shape = (h, w, 4)
    a = a[:, :, 0]  # grab single channel
    a = a.astype(np.float32) / 255
    a = np.expand_dims(a, 0)
    if rotate:
        a = image.random_rotation(a, 3 * (w - top_left_x) / w + 1)
    a = speckle(a)

    return a


def shuffle_mats_or_lists(matrix_list, stop_ind=None):
    ret = []
    assert all([len(i) == len(matrix_list[0]) for i in matrix_list])
    len_val = len(matrix_list[0])
    if stop_ind is None:
        stop_ind = len_val
    assert stop_ind <= len_val

    a = list(range(stop_ind))
    np.random.shuffle(a)
    a += list(range(stop_ind, len_val))
    for mat in matrix_list:
        if isinstance(mat, np.ndarray):
            ret.append(mat[a])
        elif isinstance(mat, list):
            ret.append([mat[i] for i in a])
        else:
            raise TypeError('`shuffle_mats_or_lists` only supports '
                            'numpy.array and list objects.')
    return ret


# Translation of characters to unique integer values
def text_to_labels(text):
    ret = []
    for char in text:
        ret.append(alphabet.find(char))
    return ret


# Reverse translation of numerical classes back to characters
def labels_to_text(labels):
    ret = []
    for c in labels:
        if c == len(alphabet):  # CTC Blank
            ret.append("")
        else:
            ret.append(alphabet[c])
    return "".join(ret)


# only a-z and space..probably not to difficult
# to expand to uppercase and symbols

def is_valid_str(in_str):
    search = re.compile(regex, re.UNICODE).search
    return bool(search(in_str))


# Uses generator functions to supply train/test with
# data. Image renderings are text are created on the fly
# each time with random perturbations

class TextImageGenerator(keras.callbacks.Callback):
    def __init__(self, monogram_file, bigram_file, minibatch_size,
                 img_w, img_h, downsample_factor, val_split,
                 absolute_max_string_len=16):

        self.minibatch_size = minibatch_size
        self.img_w = img_w
        self.img_h = img_h
        self.monogram_file = monogram_file
        self.bigram_file = bigram_file
        self.downsample_factor = downsample_factor
        self.val_split = val_split
        self.blank_label = self.get_output_size() - 1
        self.absolute_max_string_len = absolute_max_string_len

    def get_output_size(self):
        return len(alphabet) + 1

    # num_words can be independent of the epoch size due to the use of generators
    # as max_string_len grows, num_words can grow
    def build_word_list(self, num_words, max_string_len=None, mono_fraction=0.5):
        assert max_string_len <= self.absolute_max_string_len
        assert num_words % self.minibatch_size == 0
        assert (self.val_split * num_words) % self.minibatch_size == 0
        self.num_words = num_words
        self.string_list = [''] * self.num_words
        tmp_string_list = []
        self.max_string_len = max_string_len
        self.Y_data = np.ones([self.num_words, self.absolute_max_string_len]) * -1
        self.X_text = []
        self.Y_len = [0] * self.num_words

        # monogram file is sorted by frequency in english speech
        with codecs.open(self.monogram_file, mode='rt', encoding='utf-8') as f:
            for line in f:
                if len(tmp_string_list) == int(self.num_words * mono_fraction):
                    break
                word = line.rstrip()
                if max_string_len == -1 or max_string_len is None or len(word) <= max_string_len:
                    tmp_string_list.append(word)

        # bigram file contains common word pairings in english speech
        with codecs.open(self.bigram_file, mode='rt', encoding='utf-8') as f:
            lines = f.readlines()
            for line in lines:
                if len(tmp_string_list) == self.num_words:
                    break
                columns = line.lower().split()
                word = columns[0] + ' ' + columns[1]
                if is_valid_str(word) and \
                        (max_string_len == -1 or max_string_len is None or len(word) <= max_string_len):
                    tmp_string_list.append(word)
        if len(tmp_string_list) != self.num_words:
            raise IOError('Could not pull enough words from supplied monogram and bigram files. ')
        # interlace to mix up the easy and hard words
        self.string_list[::2] = tmp_string_list[:self.num_words // 2]
        self.string_list[1::2] = tmp_string_list[self.num_words // 2:]

        for i, word in enumerate(self.string_list):
            self.Y_len[i] = len(word)
            self.Y_data[i, 0:len(word)] = text_to_labels(word)
            self.X_text.append(word)
        self.Y_len = np.expand_dims(np.array(self.Y_len), 1)

        self.cur_val_index = self.val_split
        self.cur_train_index = 0

    # each time an image is requested from train/val/test, a new random
    # painting of the text is performed
    def get_batch(self, index, size, train):
        # width and height are backwards from typical Keras convention
        # because width is the time dimension when it gets fed into the RNN
        if K.image_data_format() == 'channels_first':
            X_data = np.ones([size, 1, self.img_w, self.img_h])
        else:
            X_data = np.ones([size, self.img_w, self.img_h, 1])

        labels = np.ones([size, self.absolute_max_string_len])
        input_length = np.zeros([size, 1])
        label_length = np.zeros([size, 1])
        source_str = []
        for i in range(size):
            # Mix in some blank inputs.  This seems to be important for
            # achieving translational invariance
            if train and i > size - 4:
                if K.image_data_format() == 'channels_first':
                    X_data[i, 0, 0:self.img_w, :] = self.paint_func('')[0, :, :].T
                else:
                    X_data[i, 0:self.img_w, :, 0] = self.paint_func('', )[0, :, :].T
                labels[i, 0] = self.blank_label
                input_length[i] = self.img_w // self.downsample_factor - 2
                label_length[i] = 1
                source_str.append('')
            else:
                if K.image_data_format() == 'channels_first':
                    X_data[i, 0, 0:self.img_w, :] = self.paint_func(self.X_text[index + i])[0, :, :].T
                else:
                    X_data[i, 0:self.img_w, :, 0] = self.paint_func(self.X_text[index + i])[0, :, :].T
                labels[i, :] = self.Y_data[index + i]
                input_length[i] = self.img_w // self.downsample_factor - 2
                label_length[i] = self.Y_len[index + i]
                source_str.append(self.X_text[index + i])
        inputs = {'the_input': X_data,
                  'the_labels': labels,
                  'input_length': input_length,
                  'label_length': label_length,
                  'source_str': source_str  # used for visualization only
                  }
        outputs = {'ctc': np.zeros([size])}  # dummy data for dummy loss function
        return (inputs, outputs)

    def next_train(self):
        while 1:
            ret = self.get_batch(self.cur_train_index, self.minibatch_size, train=True)
            self.cur_train_index += self.minibatch_size
            if self.cur_train_index >= self.val_split:
                self.cur_train_index = self.cur_train_index % 32
                (self.X_text, self.Y_data, self.Y_len) = shuffle_mats_or_lists(
                    [self.X_text, self.Y_data, self.Y_len], self.val_split)
            yield ret

    def next_val(self):
        while 1:
            ret = self.get_batch(self.cur_val_index, self.minibatch_size, train=False)
            self.cur_val_index += self.minibatch_size
            if self.cur_val_index >= self.num_words:
                self.cur_val_index = self.val_split + self.cur_val_index % 32
            yield ret

    def on_train_begin(self, logs={}):
        self.build_word_list(16000, 4, 1)
        self.paint_func = lambda text: paint_text(text, self.img_w, self.img_h,
                                                  rotate=False, ud=False, multi_fonts=False)

    def on_epoch_begin(self, epoch, logs={}):
        # rebind the paint function to implement curriculum learning
        if 3 <= epoch < 6:
            self.paint_func = lambda text: paint_text(text, self.img_w, self.img_h,
                                                      rotate=False, ud=True, multi_fonts=False)
        elif 6 <= epoch < 9:
            self.paint_func = lambda text: paint_text(text, self.img_w, self.img_h,
                                                      rotate=False, ud=True, multi_fonts=True)
        elif epoch >= 9:
            self.paint_func = lambda text: paint_text(text, self.img_w, self.img_h,
                                                      rotate=True, ud=True, multi_fonts=True)
        if epoch >= 21 and self.max_string_len < 12:
            self.build_word_list(32000, 12, 0.5)


# the actual loss calc occurs here despite it not being
# an internal Keras loss function

def ctc_lambda_func(args):
    y_pred, labels, input_length, label_length = args
    # the 2 is critical here since the first couple outputs of the RNN
    # tend to be garbage:
    y_pred = y_pred[:, 2:, :]
    return K.ctc_batch_cost(labels, y_pred, input_length, label_length)


# For a real OCR application, this should be beam search with a dictionary
# and language model.  For this example, best path is sufficient.

def decode_batch(test_func, word_batch):
    out = test_func([word_batch])[0]
    ret = []
    for j in range(out.shape[0]):
        out_best = list(np.argmax(out[j, 2:], 1))
        out_best = [k for k, g in itertools.groupby(out_best)]
        outstr = labels_to_text(out_best)
        ret.append(outstr)
    return ret


class VizCallback(keras.callbacks.Callback):
    def __init__(self, run_name, test_func, text_img_gen, num_display_words=6):
        self.test_func = test_func
        self.output_dir = os.path.join(
            OUTPUT_DIR, run_name)
        self.text_img_gen = text_img_gen
        self.num_display_words = num_display_words
        if not os.path.exists(self.output_dir):
            os.makedirs(self.output_dir)

    def show_edit_distance(self, num):
        num_left = num
        mean_norm_ed = 0.0
        mean_ed = 0.0
        while num_left > 0:
            word_batch = next(self.text_img_gen)[0]
            num_proc = min(word_batch['the_input'].shape[0], num_left)
            decoded_res = decode_batch(self.test_func, word_batch['the_input'][0:num_proc])
            for j in range(num_proc):
                edit_dist = editdistance.eval(decoded_res[j], word_batch['source_str'][j])
                mean_ed += float(edit_dist)
                mean_norm_ed += float(edit_dist) / len(word_batch['source_str'][j])
            num_left -= num_proc
        mean_norm_ed = mean_norm_ed / num
        mean_ed = mean_ed / num
        print('\nOut of %d samples:  Mean edit distance: %.3f Mean normalized edit distance: %0.3f'
              % (num, mean_ed, mean_norm_ed))

    def on_epoch_end(self, epoch, logs={}):
        self.model.save_weights(os.path.join(self.output_dir, 'weights%02d.h5' % (epoch)))
        self.show_edit_distance(256)
        word_batch = next(self.text_img_gen)[0]
        res = decode_batch(self.test_func, word_batch['the_input'][0:self.num_display_words])
        if word_batch['the_input'][0].shape[0] < 256:
            cols = 2
        else:
            cols = 1
        for i in range(self.num_display_words):
            pylab.subplot(self.num_display_words // cols, cols, i + 1)
            if K.image_data_format() == 'channels_first':
                the_input = word_batch['the_input'][i, 0, :, :]
            else:
                the_input = word_batch['the_input'][i, :, :, 0]
            pylab.imshow(the_input.T, cmap='Greys_r')
            pylab.xlabel('Truth = \'%s\'\nDecoded = \'%s\'' % (word_batch['source_str'][i], res[i]))
        fig = pylab.gcf()
        fig.set_size_inches(10, 13)
        pylab.savefig(os.path.join(self.output_dir, 'e%02d.png' % (epoch)))
        pylab.close()


def train(run_name, start_epoch, stop_epoch, img_w):
    # Input Parameters
    img_h = 64
    words_per_epoch = 16000
    val_split = 0.2
    val_words = int(words_per_epoch * (val_split))

    # Network parameters
    conv_filters = 16
    kernel_size = (3, 3)
    pool_size = 2
    time_dense_size = 32
    rnn_size = 512
    minibatch_size = 32

    if K.image_data_format() == 'channels_first':
        input_shape = (1, img_w, img_h)
    else:
        input_shape = (img_w, img_h, 1)

    fdir = os.path.dirname(get_file('wordlists.tgz',
                                    origin='http://www.mythic-ai.com/datasets/wordlists.tgz', untar=True))

    img_gen = TextImageGenerator(monogram_file=os.path.join(fdir, 'wordlist_mono_clean.txt'),
                                 bigram_file=os.path.join(fdir, 'wordlist_bi_clean.txt'),
                                 minibatch_size=minibatch_size,
                                 img_w=img_w,
                                 img_h=img_h,
                                 downsample_factor=(pool_size ** 2),
                                 val_split=words_per_epoch - val_words
                                 )
    act = 'relu'
    input_data = Input(name='the_input', shape=input_shape, dtype='float32')
    inner = Conv2D(conv_filters, kernel_size, padding='same',
                   activation=act, kernel_initializer='he_normal',
                   name='conv1')(input_data)
    inner = MaxPooling2D(pool_size=(pool_size, pool_size), name='max1')(inner)
    inner = Conv2D(conv_filters, kernel_size, padding='same',
                   activation=act, kernel_initializer='he_normal',
                   name='conv2')(inner)
    inner = MaxPooling2D(pool_size=(pool_size, pool_size), name='max2')(inner)

    conv_to_rnn_dims = (img_w // (pool_size ** 2), (img_h // (pool_size ** 2)) * conv_filters)
    inner = Reshape(target_shape=conv_to_rnn_dims, name='reshape')(inner)

    # cuts down input size going into RNN:
    inner = Dense(time_dense_size, activation=act, name='dense1')(inner)

    # Two layers of bidirectional GRUs
    # GRU seems to work as well, if not better than LSTM:
    gru_1 = GRU(rnn_size, return_sequences=True, kernel_initializer='he_normal', name='gru1')(inner)
    gru_1b = GRU(rnn_size, return_sequences=True, go_backwards=True, kernel_initializer='he_normal', name='gru1_b')(
        inner)
    gru1_merged = add([gru_1, gru_1b])
    gru_2 = GRU(rnn_size, return_sequences=True, kernel_initializer='he_normal', name='gru2')(gru1_merged)
    gru_2b = GRU(rnn_size, return_sequences=True, go_backwards=True, kernel_initializer='he_normal', name='gru2_b')(
        gru1_merged)

    # transforms RNN output to character activations:
    inner = Dense(img_gen.get_output_size(), kernel_initializer='he_normal',
                  name='dense2')(concatenate([gru_2, gru_2b]))
    y_pred = Activation('softmax', name='softmax')(inner)
    Model(inputs=input_data, outputs=y_pred).summary()

    labels = Input(name='the_labels', shape=[img_gen.absolute_max_string_len], dtype='float32')
    input_length = Input(name='input_length', shape=[1], dtype='int64')
    label_length = Input(name='label_length', shape=[1], dtype='int64')
    # Keras doesn't currently support loss funcs with extra parameters
    # so CTC loss is implemented in a lambda layer
    loss_out = Lambda(ctc_lambda_func, output_shape=(1,), name='ctc')([y_pred, labels, input_length, label_length])

    # clipnorm seems to speeds up convergence
    sgd = SGD(lr=0.02, decay=1e-6, momentum=0.9, nesterov=True, clipnorm=5)

    model = Model(inputs=[input_data, labels, input_length, label_length], outputs=loss_out)

    # the loss calc occurs elsewhere, so use a dummy lambda func for the loss
    model.compile(loss={'ctc': lambda y_true, y_pred: y_pred}, optimizer=sgd)
    if start_epoch > 0:
        weight_file = os.path.join(OUTPUT_DIR, os.path.join(run_name, 'weights%02d.h5' % (start_epoch - 1)))
        model.load_weights(weight_file)
    # captures output of softmax so we can decode the output during visualization
    test_func = K.function([input_data], [y_pred])

    viz_cb = VizCallback(run_name, test_func, img_gen.next_val())

    model.fit_generator(generator=img_gen.next_train(),
                        steps_per_epoch=(words_per_epoch - val_words) // minibatch_size,
                        epochs=stop_epoch,
                        validation_data=img_gen.next_val(),
                        validation_steps=val_words // minibatch_size,
                        callbacks=[viz_cb, img_gen],
                        initial_epoch=start_epoch)


if __name__ == '__main__':
    run_name = datetime.datetime.now().strftime('%Y:%m:%d:%H:%M:%S')
    train(run_name, 0, 20, 128)
    # increase to wider images and start at epoch 20. The learned weights are reloaded
    train(run_name, 20, 25, 512)

参考文献：

[1] Keras Examples：https://github.com/fchollet/keras/tree/master/examples