Tensorflow 之模型内容可视化

• TensorFlow模型保存和提取方法

1. tensorflow实现

• 卷积神经网络CNN:Tensorflow实现(以及对卷积特征的可视化)

# 卷积网络的训练数据为MNIST(28*28灰度单色图像)
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.examples.tutorials.mnist import input_data

train_epochs = 100    # 训练轮数
batch_size   = 100     # 随机出去数据大小
display_step = 1       # 显示训练结果的间隔
learning_rate= 0.0001  # 学习效率
drop_prob    = 0.5     # 正则化,丢弃比例
fch_nodes    = 512     # 全连接隐藏层神经元的个数

# 网络模型需要的一些辅助函数
# 权重初始化(卷积核初始化)
# tf.truncated_normal()不同于tf.random_normal(),返回的值中不会偏离均值两倍的标准差
# 参数shpae为一个列表对象,例如[5, 5, 1, 32]对应
# 5,5 表示卷积核的大小, 1代表通道channel,对彩色图片做卷积是3,单色灰度为1
# 最后一个数字32,卷积核的个数,(也就是卷基层提取的特征数量)
#   显式声明数据类型,切记
def weight_init(shape):
    weights = tf.truncated_normal(shape, stddev=0.1,dtype=tf.float32)
    return tf.Variable(weights)

# 偏置的初始化
def biases_init(shape):
    biases = tf.random_normal(shape,dtype=tf.float32)
    return tf.Variable(biases)

# 随机选取mini_batch
def get_random_batchdata(n_samples, batchsize):
    start_index = np.random.randint(0, n_samples - batchsize)
    return (start_index, start_index + batchsize)

# 全连接层权重初始化函数xavier
def xavier_init(layer1, layer2, constant = 1):
    Min = -constant * np.sqrt(6.0 / (layer1 + layer2))
    Max = constant * np.sqrt(6.0 / (layer1 + layer2))
    return tf.Variable(tf.random_uniform((layer1, layer2), minval = Min, maxval = Max, dtype = tf.float32))

# 卷积
def conv2d(x, w):
    return tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME')

# 源码的位置在tensorflow/python/ops下nn_impl.py和nn_ops.py
# 这个函数接收两个参数,x 是图像的像素, w 是卷积核
# x 张量的维度[batch, height, width, channels]
# w 卷积核的维度[height, width, channels, channels_multiplier]
# tf.nn.conv2d()是一个二维卷积函数,
# stirdes 是卷积核移动的步长,4个1表示,在x张量维度的四个参数上移动步长
# padding 参数'SAME',表示对原始输入像素进行填充,卷积后映射的2D图像与原图大小相等
# 填充,是指在原图像素值矩阵周围填充0像素点
# 如果不进行填充,假设 原图为 32x32 的图像,卷积和大小为 5x5 ,卷积后映射图像大小 为 28x28

# 池化
def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

# 池化跟卷积的情况有点类似
# x 是卷积后,有经过非线性激活后的图像,
# ksize 是池化滑动张量
# ksize 的维度[batch, height, width, channels],跟 x 张量相同
# strides [1, 2, 2, 1],与上面对应维度的移动步长
# padding与卷积函数相同,padding='VALID',对原图像不进行0填充

# x 是手写图像的像素值,y 是图像对应的标签
x = tf.placeholder(tf.float32, [None, 784])
y = tf.placeholder(tf.float32, [None, 10])
# 把灰度图像一维向量,转换为28x28二维结构
x_image = tf.reshape(x, [-1, 28, 28, 1])
# -1表示任意数量的样本数,大小为28x28深度为一的张量
# 可以忽略(其实是用深度为28的,28x1的张量,来表示28x28深度为1的张量)

w_conv1 = weight_init([5, 5, 1, 16])                             # 5x5,深度为1,16个
b_conv1 = biases_init([16])
h_conv1 = tf.nn.relu(conv2d(x_image, w_conv1) + b_conv1)    # 输出张量的尺寸:28x28x16
h_pool1 = max_pool_2x2(h_conv1)                                   # 池化后张量尺寸:14x14x16
# h_pool1 , 14x14的16个特征图

w_conv2 = weight_init([5, 5, 16, 32])                             # 5x5,深度为16,32个
b_conv2 = biases_init([32])
h_conv2 = tf.nn.relu(conv2d(h_pool1, w_conv2) + b_conv2)    # 输出张量的尺寸:14x14x32
h_pool2 = max_pool_2x2(h_conv2)                                   # 池化后张量尺寸:7x7x32
# h_pool2 , 7x7的32个特征图

# h_pool2是一个7x7x32的tensor,将其转换为一个一维的向量
h_fpool2 = tf.reshape(h_pool2, [-1, 7*7*32])
# 全连接层,隐藏层节点为512个
# 权重初始化
w_fc1 = xavier_init(7*7*32, fch_nodes)
b_fc1 = biases_init([fch_nodes])
h_fc1 = tf.nn.relu(tf.matmul(h_fpool2, w_fc1) + b_fc1)

# 全连接隐藏层/输出层
# 为了防止网络出现过拟合的情况,对全连接隐藏层进行 Dropout(正则化)处理,在训练过程中随机的丢弃部分
# 节点的数据来防止过拟合.Dropout同把节点数据设置为0来丢弃一些特征值,仅在训练过程中,
# 预测的时候,仍使用全数据特征
# 传入丢弃节点数据的比例
#keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob=drop_prob)

# 隐藏层与输出层权重初始化
w_fc2 = xavier_init(fch_nodes, 10)
b_fc2 = biases_init([10])

# 未激活的输出
y_ = tf.add(tf.matmul(h_fc1_drop, w_fc2), b_fc2)
# 激活后的输出
y_out = tf.nn.softmax(y_)

# 交叉熵代价函数
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y * tf.log(y_out), reduction_indices = [1]))

# tensorflow自带一个计算交叉熵的方法
# 输入没有进行非线性激活的输出值 和 对应真实标签
#cross_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_, y))

# 优化器选择Adam(有多个选择)
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cross_entropy)

# 准确率
# 每个样本的预测结果是一个(1,10)的vector
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_out, 1))
# tf.cast把bool值转换为浮点数
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

# 全局变量进行初始化的Operation
init = tf.global_variables_initializer()

# 加载数据集MNIST
mnist = input_data.read_data_sets('MNIST/mnist', one_hot=True)
n_samples = int(mnist.train.num_examples)
total_batches = int(n_samples / batch_size)

# 会话
with tf.Session() as sess:
    sess.run(init)
    Cost = []
    Accuracy = []
    for i in range(train_epochs):

        for j in range(100):
            start_index, end_index = get_random_batchdata(n_samples, batch_size)

            batch_x = mnist.train.images[start_index: end_index]
            batch_y = mnist.train.labels[start_index: end_index]
            _, cost, accu = sess.run([ optimizer, cross_entropy,accuracy], feed_dict={x:batch_x, y:batch_y})
            Cost.append(cost)
            Accuracy.append(accu)
        if i % display_step ==0:
            print ('Epoch : %d ,  Cost : %.7f'%(i+1, cost))
    print ('training finished')
    # 代价函数曲线
    fig1,ax1 = plt.subplots(figsize=(10,7))
    plt.plot(Cost)
    ax1.set_xlabel('Epochs')
    ax1.set_ylabel('Cost')
    plt.title('Cross Loss')
    plt.grid()
    plt.show()
    # 准确率曲线
    fig7,ax7 = plt.subplots(figsize=(10,7))
    plt.plot(Accuracy)
    ax7.set_xlabel('Epochs')
    ax7.set_ylabel('Accuracy Rate')
    plt.title('Train Accuracy Rate')
    plt.grid()
    plt.show()
#----------------------------------各个层特征可视化-------------------------------
    # imput image
    fig2,ax2 = plt.subplots(figsize=(2,2))
    ax2.imshow(np.reshape(mnist.train.images[11], (28, 28)))
    plt.show()

    # 第一层的卷积输出的特征图
    input_image = mnist.train.images[11:12]
    conv1_16 = sess.run(h_conv1, feed_dict={x:input_image})     # [16, 28, 28 ,1]
    conv1_reshape = sess.run(tf.reshape(conv1_16, [16, 1, 28, 28]))
    fig3,ax3 = plt.subplots(nrows=1, ncols=16, figsize = (16,1))
    for i in range(16):
        ax3[i].imshow(conv1_reshape[i][0])                      # tensor的切片[batch, channels, row, column]

    plt.title('Conv1 16x28x28')
    plt.show()

    # 第一层池化后的特征图
    pool1_16 = sess.run(h_pool1, feed_dict={x:input_image})     # [16, 14, 14, 1]
    pool1_reshape = sess.run(tf.reshape(pool1_16, [16, 1, 14, 14]))
    fig4,ax4 = plt.subplots(nrows=1, ncols=16, figsize=(16,1))
    for i in range(16):
        ax4[i].imshow(pool1_reshape[i][0])

    plt.title('Pool1 16x14x14')
    plt.show()

    # 第二层卷积输出特征图
    conv2_32 = sess.run(h_conv2, feed_dict={x:input_image})          # [32, 14, 14, 1]
    conv2_reshape = sess.run(tf.reshape(conv2_32, [32, 1, 14, 14]))
    fig5,ax5 = plt.subplots(nrows=1, ncols=32, figsize = (32, 1))
    for i in range(32):
        ax5[i].imshow(conv2_reshape[i][0])
    plt.title('Conv2 32x14x14')
    plt.show()

    # 第二层池化后的特征图
    pool2_32 = sess.run(h_pool2, feed_dict={x:input_image})          #[32, 7, 7, 1]
    pool2_reshape = sess.run(tf.reshape(pool2_32, [32, 1, 7, 7]))
    fig6,ax6 = plt.subplots(nrows=1, ncols=32, figsize = (32, 1))
    plt.title('Pool2 32x7x7')
    for i in range(32):
        ax6[i].imshow(pool2_reshape[i][0])

    plt.show()

2.keras实现

• Essentials of Deep Learning: Visualizing Convolutional Neural Networks in Python
• https://github.com/keras-team/keras/blob/master/examples/conv_filter_visualization.py
• https://blog.keras.io/how-convolutional-neural-networks-see-the-world.html
• Neural network visualization toolkit for keras https://raghakot.github.io/keras-vis

'''Visualization of the filters of VGG16, via gradient ascent in input space.
This script can run on CPU in a few minutes.
Results example: http://i.imgur.com/4nj4KjN.jpg
'''
from __future__ import print_function

from scipy.misc import imsave
import numpy as np
import time
from keras.applications import vgg16
from keras import backend as K

# dimensions of the generated pictures for each filter.
img_width = 128
img_height = 128

# the name of the layer we want to visualize
# (see model definition at keras/applications/vgg16.py)
layer_name = 'block5_conv1'

# util function to convert a tensor into a valid image

def deprocess_image(x):
    # normalize tensor: center on 0., ensure std is 0.1
    x -= x.mean()
    x /= (x.std() + K.epsilon())
    x *= 0.1

    # clip to [0, 1]
    x += 0.5
    x = np.clip(x, 0, 1)

    # convert to RGB array
    x *= 255
    if K.image_data_format() == 'channels_first':
        x = x.transpose((1, 2, 0))
    x = np.clip(x, 0, 255).astype('uint8')
    return x

# build the VGG16 network with ImageNet weights
model = vgg16.VGG16(weights='imagenet', include_top=False)
print('Model loaded.')

model.summary()

# this is the placeholder for the input images
input_img = model.input

# get the symbolic outputs of each "key" layer (we gave them unique names).
layer_dict = dict([(layer.name, layer) for layer in model.layers[1:]])

def normalize(x):
    # utility function to normalize a tensor by its L2 norm
    return x / (K.sqrt(K.mean(K.square(x))) + K.epsilon())

kept_filters = []
for filter_index in range(200):
    # we only scan through the first 200 filters,
    # but there are actually 512 of them
    print('Processing filter %d' % filter_index)
    start_time = time.time()

    # we build a loss function that maximizes the activation
    # of the nth filter of the layer considered
    layer_output = layer_dict[layer_name].output
    if K.image_data_format() == 'channels_first':
        loss = K.mean(layer_output[:, filter_index, :, :])
    else:
        loss = K.mean(layer_output[:, :, :, filter_index])

    # we compute the gradient of the input picture wrt this loss
    grads = K.gradients(loss, input_img)[0]

    # normalization trick: we normalize the gradient
    grads = normalize(grads)

    # this function returns the loss and grads given the input picture
    iterate = K.function([input_img], [loss, grads])

    # step size for gradient ascent
    step = 1.

    # we start from a gray image with some random noise
    if K.image_data_format() == 'channels_first':
        input_img_data = np.random.random((1, 3, img_width, img_height))
    else:
        input_img_data = np.random.random((1, img_width, img_height, 3))
    input_img_data = (input_img_data - 0.5) * 20 + 128

    # we run gradient ascent for 20 steps
    for i in range(20):
        loss_value, grads_value = iterate([input_img_data])
        input_img_data += grads_value * step

        print('Current loss value:', loss_value)
        if loss_value <= 0.:
            # some filters get stuck to 0, we can skip them
            break

    # decode the resulting input image
    if loss_value > 0:
        img = deprocess_image(input_img_data[0])
        kept_filters.append((img, loss_value))
    end_time = time.time()
    print('Filter %d processed in %ds' % (filter_index, end_time - start_time))

# we will stich the best 64 filters on a 8 x 8 grid.
n = 8

# the filters that have the highest loss are assumed to be better-looking.
# we will only keep the top 64 filters.
kept_filters.sort(key=lambda x: x[1], reverse=True)
kept_filters = kept_filters[:n * n]

# build a black picture with enough space for
# our 8 x 8 filters of size 128 x 128, with a 5px margin in between
margin = 5
width = n * img_width + (n - 1) * margin
height = n * img_height + (n - 1) * margin
stitched_filters = np.zeros((width, height, 3))

# fill the picture with our saved filters
for i in range(n):
    for j in range(n):
        img, loss = kept_filters[i * n + j]
        stitched_filters[(img_width + margin) * i: (img_width + margin) * i + img_width,
                         (img_height + margin) * j: (img_height + margin) * j + img_height, :] = img

# save the result to disk
imsave('stitched_filters_%dx%d.png' % (n, n), stitched_filters)

3.tflearn

• https://github.com/tflearn/tflearn/blob/master/examples/basics/weights_persistence.py

""" An example showing how to save/restore models and retrieve weights. """

from __future__ import absolute_import, division, print_function

import tflearn

import tflearn.datasets.mnist as mnist

# MNIST Data
X, Y, testX, testY = mnist.load_data(one_hot=True)

# Model
input_layer = tflearn.input_data(shape=[None, 784], name='input')
dense1 = tflearn.fully_connected(input_layer, 128, name='dense1')
dense2 = tflearn.fully_connected(dense1, 256, name='dense2')
softmax = tflearn.fully_connected(dense2, 10, activation='softmax')
regression = tflearn.regression(softmax, optimizer='adam',
                                learning_rate=0.001,
                                loss='categorical_crossentropy')

# Define classifier, with model checkpoint (autosave)
model = tflearn.DNN(regression, checkpoint_path='model.tfl.ckpt')

# Train model, with model checkpoint every epoch and every 200 training steps.
model.fit(X, Y, n_epoch=1,
          validation_set=(testX, testY),
          show_metric=True,
          snapshot_epoch=True, # Snapshot (save & evaluate) model every epoch.
          snapshot_step=500, # Snapshot (save & evalaute) model every 500 steps.
          run_id='model_and_weights')

# ---------------------
# Save and load a model
# ---------------------

# Manually save model
model.save("model.tfl")

# Load a model
model.load("model.tfl")

# Or Load a model from auto-generated checkpoint
# >> model.load("model.tfl.ckpt-500")

# Resume training
model.fit(X, Y, n_epoch=1,
          validation_set=(testX, testY),
          show_metric=True,
          snapshot_epoch=True,
          run_id='model_and_weights')

# ------------------
# Retrieving weights
# ------------------

# Retrieve a layer weights, by layer name:
dense1_vars = tflearn.variables.get_layer_variables_by_name('dense1')
# Get a variable's value, using model `get_weights` method:
print("Dense1 layer weights:")
print(model.get_weights(dense1_vars[0]))
# Or using generic tflearn function:
print("Dense1 layer biases:")
with model.session.as_default():
    print(tflearn.variables.get_value(dense1_vars[1]))

# It is also possible to retrieve a layer weights through its attributes `W`
# and `b` (if available).
# Get variable's value, using model `get_weights` method:
print("Dense2 layer weights:")
print(model.get_weights(dense2.W))
# Or using generic tflearn function:
print("Dense2 layer biases:")
with model.session.as_default():
print(tflearn.variables.get_value(dense2.b))