http://www.hcinnovation.tw/tc/index.php/k2-blog/item/66-tensorflow 課程講義3

1. TensorFlow深度學習快速上⼿手班
三、電腦視覺應⽤用
By Mark Chang

2. • 電腦視覺簡介
• 模型選擇與參數調整
• 影像識別實作

3. 電腦視覺簡介

4. 電腦視覺
• 電腦視覺是⼀一⾨門研究如何使機器「看」的科學
• ⽤用電腦代替⼈人眼對⺫⽬目標進⾏行識別、跟蹤和測量
等機器視覺，並進⼀一步做圖像處理。
• https://zh.wikipedia.org/wiki/%E8%AE
%A1%E7%AE%97%E6%9C%BA
%E8%A7%86%E8%A7%89

5. 影像識別
http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf

6. 物件偵測
http://papers.nips.cc/paper/5207-deep-neural-networks-for-object-detection.pdf

7. 影像補⿑齊
http://arxiv.org/abs/1601.06759

8. 藝術創作
http://arxiv.org/abs/1508.06576

9. 卷積神經網路

10. 影像識別
• 同⼀一個數字可能出現在圖⽚片中的不同部分
• 但這些圖⽚片所代表的數字相同

11. Local Connectivity
每個神經元只看到圖片中的一小區塊

12. Parameter Sharing
同一「種類」的神經元具有相同的weights

13. Parameter Sharing
不同「種類」的神經元具有不同的weights

14. 卷積神經網路
• Convolutional Layer
depth
widthwidthdepth
weights weights
height
shared weight

15. 卷積神經網路
• Stride
• Padding
Stride = 1
Stride = 2
Padding = 0
Padding = 1

16. 視覺認知
http://www.nature.com/neuro/journal/v8/n8/images/nn0805-975-F1.jpg

17. 特徵擷取

18. 卷積神經網路
• Pooling Layer
1
3
2
4
5
7
6
8
0
0
4
4
6
6
0
0
4
5
3
2
no overlap
no padding no weights
depth = 1
7
8
6
4
Maximum
Pooling
Average
Pooling

19. 卷積神經網路
Convolutional
Layer
Convolutional
Layer Pooling
Layer
Pooling
Layer
Receptive Fields
Receptive Fields
Input
Layer

20. 卷積神經網路
Input Layer
Convolutional
Layer with
Receptive Fields:
Max-pooling
Layer with
Width =3, Height = 3
Filter Responses
Filter Responses
Input Image

21. 影像識別實作

22. 卷積神經網路實作
https://github.com/ckmarkoh/ntc_deeplearning_tensorflow/
blob/master/sec3/convnet.ipynb

23. MNIST
• 數字識別
• 多元分類：0~9
https://www.tensorflow.org/versions/r0.7/images/MNIST.png

24. Create Variables Operators
def weight_variable(shape):
return tf.Variable(tf.truncated_normal(shape, stddev=0.1))
def bias_variable(shape):
return tf.Variable(tf.constant(0.1, shape=shape))
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME')

25. Computational Graph
x_ = tf.placeholder(tf.float32, [None, 784], name=x_)
y_ = tf.placeholder(tf.float32, [None, 10], name=y_”)
x_image = tf.reshape(x_, [-1,28,28,1])
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)
W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y= tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

26. 卷積神經網路
nx28x28x1
nx28x28x32
nx14x14x32
nx14x14x64
nx7x7x64
nx1024
nx10
x_image
h_conv1
h_pool1
h_conv2
h_pool2
h_fc1
y

27. Reshape
x_image = tf.reshape(x_, [-1,28,28,1])
x
n
784
n
28
1

28. Convolutional Layer
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
5
1
32
32
5x5
1
32
32
W_conv1
W_conv1
b_conv1
b_conv1

29. Convolutional Layer
tf.nn.conv2d(x, W , strides=[1, 1, 1, 1], padding='SAME')+b
1
5x5 1x1
28
28
28
28
strides=1
padding='SAME'
[ batch, in_height, in_width, in_channels ]

30. Convolutional Layer
tf.nn.conv2d(x, W , strides=[1, 1, 1, 1], padding='SAME')+b
nx28x28x1 nx28x28x32
28
28
28
28

31. ReLU
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
ReLU:
⇢
nin if nin 0
0 otherwise
-0.5 0.2 0.3 -0.1
0.2 -0.3 -0.4 -1.1
2.1 -2.1 0.1 1.2
0.2 3.0 -0.3 0.5
0 0.2 0.3 0
0.2 0 0 0
2.1 0 0.1 1.2
0.2 3.0 0 0.5

32. Pooling Layer
tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME')
1x2x2x1
1
1
1
1
2
2x2 1x1

33. Pooling Layer
tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME')
2
strides=2
padding='SAME'
28
28
14
14

34. Pooling Layer
h_pool1 = max_pool_2x2(h_conv1)
nx28x28x32 nx14x14x32
28
28
14
14

35. Reshape
h_pool2_
flat
n
7*7*64
7
64
n
h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])

36. GoogLeNet影像識別
https://github.com/ckmarkoh/ntc_deeplearning_tensorflow/
blob/master/sec3/googlenet.ipynb

37. GoogLeNet
http://www.cs.unc.edu/~wliu/papers/GoogLeNet.pdf
22 layers deep network

38. 訓練資料
• ILSVRC 2014 Classification Challenge
– http://www.image-net.org/challenges/
LSVRC/2014/
• Dataset:
1000 categories
– Training: 1,200,000
– Validation: 50,000
– Testing: 100,000

39. Inception Module

40. Load Computational Graph
model_fn = 'tensorflow_inception_graph.pb'
graph = tf.Graph()
sess = tf.InteractiveSession(graph=graph)
graph_def = tf.GraphDef.FromString(open(model_fn).read())
t_input = tf.placeholder(np.float32, name='input')
imagenet_mean = 139
t_preprocessed = tf.expand_dims(t_input - imagenet_mean, 0)
tf.import_graph_def(graph_def, {'input': t_preprocessed})
t_output = graph.get_tensor_by_name(import/output2:0)

41. Load Label
f = open(label.json)
labels = json.loads(.join(f.readlines()))
f.close()
1: kit fox, Vulpes macrotis,
2: English setter,
3: Siberian husky,
4: Australian terrier,
......
998: stole,
999: carbonara,
1000: dumbbell

42. Run Computational Graph
def load_image(imgfile):
return np.float32(PIL.Image.open(imgfile).resize((224,224)))
def get_class(image):
return labels[str(np.argmax(sess.run([t_output], {t_input:
load_image(image)})))]
print get_class('img/img1.jpg')
leaf beetle, chrysomelid

43. 講師資訊
• Email: ckmarkoh at gmail dot com
• Blog: http://cpmarkchang.logdown.com
• Github: https://github.com/ckmarkoh
Mark Chang
• Facebook: https://www.facebook.com/ckmarkoh.chang
• Slideshare: http://www.slideshare.net/ckmarkohchang
• Linkedin:
https://www.linkedin.com/pub/mark-chang/85/25b/847
43