在CNN上增長一層CAM告訴你CNN到底關注什麼

Cam(Class Activation Mapping)是一個頗有意思的算法，他可以將神經網絡到底在關注什麼可視化的表現出來。但同時它的實現卻又如此簡介，相比NIN，googLenet這些使用GAP（Global Average Pooling）用來代替全鏈接層，他卻將其輸出的權重和featuremap相乘，累加，將其用圖像表示出來。
其網絡架構以下

Class Activation Mapping具體論文

固然Cam的目的並不單單是將其表示出來，神經網絡所關注的地方，一般就是物體所在的地方，所以它能夠輔助訓練檢測網絡。
所以就有了PlacesNet。

論文

在這裏能夠體驗

網絡上基本都是基於AlexNet等網絡，其實任何網絡，只要加一層全局池化層就能夠幫助咱們將CNN關注什麼表示出來，所以我對Tensorflow官方Mnist的CNN網絡進行少許的修改，實現了CAM。只是將最後的全鏈接層，改成了全局池化層。

CAM的核心公式很簡單
\[ S_c = \sum_{k} {W_k^c} {\sum_{x,y} {f_k(x,y)}} \]

將全局平均池化層輸出的權重乘上feature map累加

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import matplotlib.pyplot as plt
import numpy as np
%matplotlib inline

mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)

/home/lyn/anaconda3/lib/python3.6/importlib/_bootstrap.py:205: RuntimeWarning: compiletime version 3.5 of module 'tensorflow.python.framework.fast_tensor_util' does not match runtime version 3.6
  return f(*args, **kwds)


Extracting /tmp/data/train-images-idx3-ubyte.gz
Extracting /tmp/data/train-labels-idx1-ubyte.gz
Extracting /tmp/data/t10k-images-idx3-ubyte.gz
Extracting /tmp/data/t10k-labels-idx1-ubyte.gz

在讀入數據後，設定基本的學習參數

# Training Parameters
learning_rate = 0.001
num_steps = 10000
batch_size = 128
display_step = 10

# Network Parameters
num_input = 784 # MNIST data input (img shape: 28*28)
num_classes = 10 # MNIST total classes (0-9 digits)

# tf Graph input
X = tf.placeholder(tf.float32, [None, num_input])
Y = tf.placeholder(tf.int32, [None, num_classes])

# Create some wrappers for simplicity
def conv2d(x, W, b, strides=1):
    # Conv2D wrapper, with bias and relu activation
    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
    x = tf.nn.bias_add(x, b)
    return tf.nn.relu(x)


def maxpool2d(x, k=2):
    # MaxPool2D wrapper
    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],
                          padding='SAME')

在實際使用中咱們須要得到得feature map，與全局池化層相乘並累加

def conv_layers(x,weights,biases):
    x = tf.reshape(x, shape=[-1, 28, 28, 1])

    # Convolution Layer
    conv1 = conv2d(x, weights['wc1'], biases['bc1'])
    # Max Pooling (down-sampling)
    conv1 = maxpool2d(conv1, k=2)
    # Convolution Layer
    conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])
    # Max Pooling (down-sampling)
    conv2 = maxpool2d(conv2, k=2)
    return conv2

def out_layer(conv2,weights,biases):
    gap = tf.nn.avg_pool(conv2,ksize=[1,7,7,1],strides=[1,7,7,1],padding="SAME")
    gap = tf.reshape(gap,[-1,128])
    out = tf.add(tf.matmul(gap, weights['out']), biases['out'])
    return out

def generate_heatmap(conv2,label,weights):
    conv2_resized = tf.image.resize_images(conv2,[28,28])
    label_w = tf.gather(tf.transpose(weights['out']),label)
    label_w = tf.reshape(label_w,[-1,128,1])
    conv2_resized = tf.reshape(conv2_resized,[-1,28*28,128])
    classmap = tf.matmul( conv2_resized, label_w )
    classmap = tf.reshape( classmap, [-1, 28,28] )
    return classmap

# Store layers weight & bias
weights = {
    # 5x5 conv, 1 input, 32 outputs
    'wc1': tf.Variable(tf.random_normal([5, 5, 1, 32])),
    # 5x5 conv, 32 inputs, 64 outputs
    'wc2': tf.Variable(tf.random_normal([5, 5, 32, 128])),
    'out': tf.Variable(tf.random_normal([128, num_classes]))
}

biases = {
    'bc1': tf.Variable(tf.random_normal([32])),
    'bc2': tf.Variable(tf.random_normal([128])),
    'bd1': tf.Variable(tf.random_normal([1024])),
    'out': tf.Variable(tf.random_normal([num_classes]))
}

# Construct model
conv2 = conv_layers(X, weights, biases)
logits = out_layer(conv2,weights, biases)
prediction = tf.nn.softmax(logits)
classmap = generate_heatmap(conv2,tf.argmax(prediction,1),weights)
loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
    logits=logits, labels=Y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
train_op = optimizer.minimize(loss_op)


# Evaluate model
correct_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

# Initialize the variables (i.e. assign their default value)
init = tf.global_variables_initializer()

sess = tf.Session()
sess.run(init)

for step in range(1, num_steps+1):
    batch_x, batch_y = mnist.train.next_batch(batch_size)
    # Run optimization op (backprop)
    sess.run(train_op, feed_dict={X: batch_x, Y: batch_y})
    if step % display_step == 0 or step == 1:
        # Calculate batch loss and accuracy
        loss, acc = sess.run([loss_op, accuracy], feed_dict={X: batch_x,
                                                             Y: batch_y})
        print("Step " + str(step) + ", Minibatch Loss= " + \
              "{:.4f}".format(loss) + ", Training Accuracy= " + \
              "{:.3f}".format(acc))

print("Optimization Finished!")

# Calculate accuracy for 256 MNIST test images
print("Testing Accuracy:", \
    sess.run(accuracy, feed_dict={X: mnist.test.images[:256],
                                  Y: mnist.test.labels[:256],}
            ))
sess.run(conv2,feed_dict={X:mnist.test.images[:10],Y:mnist.test.labels[:10]})
classmaps = sess.run(classmap,feed_dict={X:mnist.test.images[:10],Y:mnist.test.labels[:10]})

Testing Accuracy: 0.976562

for i in range(9):
    plt.subplot(33*10+i+1)
    plt.axis("off")
    plt.imshow(classmaps[i],cmap="gray")
    plt.title("label is"+str(np.argmax(mnist.test.labels[i])))

黑色爲正權重點，白色爲負權重點。

固然因爲網絡太淺，使用了全局平均池化之後，訓練時間大大增加，準確率也不如以前。