API tools faq
paste
Advertisement
ERENARD63

Residual Network ++V2

ERENARD63
Apr 12th, 2018
661
0
Never
Add comment
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
Python 29.40 KB | None | 0 0
raw download clone embed print report
  1.  
  2. # coding: utf-8
  3.  
  4. # # Residual Networks
  5. #
  6. # Welcome to the second assignment of this week! You will learn how to build very deep convolutional networks, using Residual Networks (ResNets). In theory, very deep networks can represent very complex functions; but in practice, they are hard to train. Residual Networks, introduced by [He et al.](https://arxiv.org/pdf/1512.03385.pdf), allow you to train much deeper networks than were previously practically feasible.
  7. #
  8. # **In this assignment, you will:**
  9. # - Implement the basic building blocks of ResNets.
  10. # - Put together these building blocks to implement and train a state-of-the-art neural network for image classification.
  11. #
  12. # This assignment will be done in Keras.
  13. #
  14. # Before jumping into the problem, let's run the cell below to load the required packages.
  15.  
  16. # In[2]:
  17.  
  18. import numpy as np
  19. from keras import layers
  20. from keras.layers import Input, Add, Dense, Activation, ZeroPadding2D, BatchNormalization, Flatten, Conv2D, AveragePooling2D, MaxPooling2D, GlobalMaxPooling2D
  21. from keras.models import Model, load_model
  22. from keras.preprocessing import image
  23. from keras.utils import layer_utils
  24. from keras.utils.data_utils import get_file
  25. from keras.applications.imagenet_utils import preprocess_input
  26. import pydot
  27. from IPython.display import SVG
  28. from keras.utils.vis_utils import model_to_dot
  29. from keras.utils import plot_model
  30. from resnets_utils import *
  31. from keras.initializers import glorot_uniform
  32. import scipy.misc
  33. from matplotlib.pyplot import imshow
  34. get_ipython().magic('matplotlib inline')
  35.  
  36. import keras.backend as K
  37. K.set_image_data_format('channels_last')
  38. K.set_learning_phase(1)
  39.  
  40.  
  41. # ## 1 - The problem of very deep neural networks
  42. #
  43. # Last week, you built your first convolutional neural network. In recent years, neural networks have become deeper, with state-of-the-art networks going from just a few layers (e.g., AlexNet) to over a hundred layers.
  44. #
  45. # The main benefit of a very deep network is that it can represent very complex functions. It can also learn features at many different levels of abstraction, from edges (at the lower layers) to very complex features (at the deeper layers). However, using a deeper network doesn't always help. A huge barrier to training them is vanishing gradients: very deep networks often have a gradient signal that goes to zero quickly, thus making gradient descent unbearably slow. More specifically, during gradient descent, as you backprop from the final layer back to the first layer, you are multiplying by the weight matrix on each step, and thus the gradient can decrease exponentially quickly to zero (or, in rare cases, grow exponentially quickly and "explode" to take very large values).
  46. #
  47. # During training, you might therefore see the magnitude (or norm) of the gradient for the earlier layers descrease to zero very rapidly as training proceeds:
  48.  
  49. # <img src="images/vanishing_grad_kiank.png" style="width:450px;height:220px;">
  50. # <caption><center> <u> <font color='purple'> **Figure 1** </u><font color='purple'>  : **Vanishing gradient** <br> The speed of learning decreases very rapidly for the early layers as the network trains </center></caption>
  51. #
  52. # You are now going to solve this problem by building a Residual Network!
  53.  
  54. # ## 2 - Building a Residual Network
  55. #
  56. # In ResNets, a "shortcut" or a "skip connection" allows the gradient to be directly backpropagated to earlier layers:  
  57. #
  58. # <img src="images/skip_connection_kiank.png" style="width:650px;height:200px;">
  59. # <caption><center> <u> <font color='purple'> **Figure 2** </u><font color='purple'>  : A ResNet block showing a **skip-connection** <br> </center></caption>
  60. #
  61. # The image on the left shows the "main path" through the network. The image on the right adds a shortcut to the main path. By stacking these ResNet blocks on top of each other, you can form a very deep network.
  62. #
  63. # We also saw in lecture that having ResNet blocks with the shortcut also makes it very easy for one of the blocks to learn an identity function. This means that you can stack on additional ResNet blocks with little risk of harming training set performance. (There is also some evidence that the ease of learning an identity function--even more than skip connections helping with vanishing gradients--accounts for ResNets' remarkable performance.)
  64. #
  65. # Two main types of blocks are used in a ResNet, depending mainly on whether the input/output dimensions are same or different. You are going to implement both of them.
  66.  
  67. # ### 2.1 - The identity block
  68. #
  69. # The identity block is the standard block used in ResNets, and corresponds to the case where the input activation (say $a^{[l]}$) has the same dimension as the output activation (say $a^{[l+2]}$). To flesh out the different steps of what happens in a ResNet's identity block, here is an alternative diagram showing the individual steps:
  70. #
  71. # <img src="images/idblock2_kiank.png" style="width:650px;height:150px;">
  72. # <caption><center> <u> <font color='purple'> **Figure 3** </u><font color='purple'>  : **Identity block.** Skip connection "skips over" 2 layers. </center></caption>
  73. #
  74. # The upper path is the "shortcut path." The lower path is the "main path." In this diagram, we have also made explicit the CONV2D and ReLU steps in each layer. To speed up training we have also added a BatchNorm step. Don't worry about this being complicated to implement--you'll see that BatchNorm is just one line of code in Keras!
  75. #
  76. # In this exercise, you'll actually implement a slightly more powerful version of this identity block, in which the skip connection "skips over" 3 hidden layers rather than 2 layers. It looks like this:
  77. #
  78. # <img src="images/idblock3_kiank.png" style="width:650px;height:150px;">
  79. # <caption><center> <u> <font color='purple'> **Figure 4** </u><font color='purple'>  : **Identity block.** Skip connection "skips over" 3 layers.</center></caption>
  80. #
  81. # Here're the individual steps.
  82. #
  83. # First component of main path:
  84. # - The first CONV2D has $F_1$ filters of shape (1,1) and a stride of (1,1). Its padding is "valid" and its name should be `conv_name_base + '2a'`. Use 0 as the seed for the random initialization.
  85. # - The first BatchNorm is normalizing the channels axis.  Its name should be `bn_name_base + '2a'`.
  86. # - Then apply the ReLU activation function. This has no name and no hyperparameters.
  87. #
  88. # Second component of main path:
  89. # - The second CONV2D has $F_2$ filters of shape $(f,f)$ and a stride of (1,1). Its padding is "same" and its name should be `conv_name_base + '2b'`. Use 0 as the seed for the random initialization.
  90. # - The second BatchNorm is normalizing the channels axis.  Its name should be `bn_name_base + '2b'`.
  91. # - Then apply the ReLU activation function. This has no name and no hyperparameters.
  92. #
  93. # Third component of main path:
  94. # - The third CONV2D has $F_3$ filters of shape (1,1) and a stride of (1,1). Its padding is "valid" and its name should be `conv_name_base + '2c'`. Use 0 as the seed for the random initialization.
  95. # - The third BatchNorm is normalizing the channels axis.  Its name should be `bn_name_base + '2c'`. Note that there is no ReLU activation function in this component.
  96. #
  97. # Final step:
  98. # - The shortcut and the input are added together.
  99. # - Then apply the ReLU activation function. This has no name and no hyperparameters.
  100. #
  101. # **Exercise**: Implement the ResNet identity block. We have implemented the first component of the main path. Please read over this carefully to make sure you understand what it is doing. You should implement the rest.
  102. # - To implement the Conv2D step: [See reference](https://keras.io/layers/convolutional/#conv2d)
  103. # - To implement BatchNorm: [See reference](https://faroit.github.io/keras-docs/1.2.2/layers/normalization/) (axis: Integer, the axis that should be normalized (typically the channels axis))
  104. # - For the activation, use:  `Activation('relu')(X)`
  105. # - To add the value passed forward by the shortcut: [See reference](https://keras.io/layers/merge/#add)
  106.  
  107. # In[11]:
  108.  
  109. # GRADED FUNCTION: identity_block
  110.  
  111. def identity_block(X, f, filters, stage, block):
  112.     """
  113.    Implementation of the identity block as defined in Figure 3
  114.    
  115.    Arguments:
  116.    X -- input tensor of shape (m, n_H_prev, n_W_prev, n_C_prev)
  117.    f -- integer, specifying the shape of the middle CONV's window for the main path
  118.    filters -- python list of integers, defining the number of filters in the CONV layers of the main path
  119.    stage -- integer, used to name the layers, depending on their position in the network
  120.    block -- string/character, used to name the layers, depending on their position in the network
  121.    
  122.    Returns:
  123.    X -- output of the identity block, tensor of shape (n_H, n_W, n_C)
  124.    """
  125.    
  126.     # defining name basis
  127.     conv_name_base = 'res' + str(stage) + block + '_branch'
  128.     bn_name_base = 'bn' + str(stage) + block + '_branch'
  129.    
  130.     # Retrieve Filters
  131.     F1, F2, F3 = filters
  132.    
  133.     # Save the input value. You'll need this later to add back to the main path.
  134.     X_shortcut = X
  135.    
  136.     # First component of main path
  137.     X = Conv2D(filters = F1, kernel_size = (1, 1), strides = (1,1), padding = 'valid', name = conv_name_base + '2a', kernel_initializer = glorot_uniform(seed=0))(X)
  138.     X = BatchNormalization(axis = 3, name = bn_name_base + '2a')(X)
  139.     X = Activation('relu')(X)
  140.    
  141.     ### START CODE HERE ###
  142.    
  143.     # Second component of main path (≈3 lines)
  144.     X = Conv2D(filters = F2, kernel_size = (f, f), strides = (1,1), padding = 'same', name = conv_name_base + '2b', kernel_initializer = glorot_uniform(seed=0))(X)
  145.     X = BatchNormalization(axis = 3, name = bn_name_base + '2b')(X)
  146.     X = Activation('relu')(X)
  147.    
  148.     # Third component of main path (≈2 lines)
  149.     X = Conv2D(filters = F3, kernel_size = (1, 1), strides = (1,1), padding = 'valid', name = conv_name_base + '2c', kernel_initializer = glorot_uniform(seed=0))(X)
  150.     X = BatchNormalization(axis = 3, name = bn_name_base + '2c')(X)
  151.  
  152.     # Final step: Add shortcut value to main path, and pass it through a RELU activation (≈2 lines)
  153.     X = Add()([X, X_shortcut])
  154.     X = Activation('relu')(X)
  155.    
  156.     ### END CODE HERE ###
  157.    
  158.     return X
  159.  
  160.  
  161. # In[12]:
  162.  
  163. tf.reset_default_graph()
  164.  
  165. with tf.Session() as test:
  166.     np.random.seed(1)
  167.     A_prev = tf.placeholder("float", [3, 4, 4, 6])
  168.     X = np.random.randn(3, 4, 4, 6)
  169.     A = identity_block(A_prev, f = 2, filters = [2, 4, 6], stage = 1, block = 'a')
  170.     test.run(tf.global_variables_initializer())
  171.     out = test.run([A], feed_dict={A_prev: X, K.learning_phase(): 0})
  172.     print("out = " + str(out[0][1][1][0]))
  173.  
  174.  
  175. # **Expected Output**:
  176. #
  177. # <table>
  178. #     <tr>
  179. #         <td>
  180. #             **out**
  181. #         </td>
  182. #         <td>
  183. #            [ 0.94822985  0.          1.16101444  2.747859    0.          1.36677003]
  184. #         </td>
  185. #     </tr>
  186. #
  187. # </table>
  188.  
  189. # ## 2.2 - The convolutional block
  190. #
  191. # You've implemented the ResNet identity block. Next, the ResNet "convolutional block" is the other type of block. You can use this type of block when the input and output dimensions don't match up. The difference with the identity block is that there is a CONV2D layer in the shortcut path:
  192. #
  193. # <img src="images/convblock_kiank.png" style="width:650px;height:150px;">
  194. # <caption><center> <u> <font color='purple'> **Figure 4** </u><font color='purple'>  : **Convolutional block** </center></caption>
  195. #
  196. # The CONV2D layer in the shortcut path is used to resize the input $x$ to a different dimension, so that the dimensions match up in the final addition needed to add the shortcut value back to the main path. (This plays a similar role as the matrix $W_s$ discussed in lecture.) For example, to reduce the activation dimensions's height and width by a factor of 2, you can use a 1x1 convolution with a stride of 2. The CONV2D layer on the shortcut path does not use any non-linear activation function. Its main role is to just apply a (learned) linear function that reduces the dimension of the input, so that the dimensions match up for the later addition step.
  197. #
  198. # The details of the convolutional block are as follows.
  199. #
  200. # First component of main path:
  201. # - The first CONV2D has $F_1$ filters of shape (1,1) and a stride of (s,s). Its padding is "valid" and its name should be `conv_name_base + '2a'`.
  202. # - The first BatchNorm is normalizing the channels axis.  Its name should be `bn_name_base + '2a'`.
  203. # - Then apply the ReLU activation function. This has no name and no hyperparameters.
  204. #
  205. # Second component of main path:
  206. # - The second CONV2D has $F_2$ filters of (f,f) and a stride of (1,1). Its padding is "same" and it's name should be `conv_name_base + '2b'`.
  207. # - The second BatchNorm is normalizing the channels axis.  Its name should be `bn_name_base + '2b'`.
  208. # - Then apply the ReLU activation function. This has no name and no hyperparameters.
  209. #
  210. # Third component of main path:
  211. # - The third CONV2D has $F_3$ filters of (1,1) and a stride of (1,1). Its padding is "valid" and it's name should be `conv_name_base + '2c'`.
  212. # - The third BatchNorm is normalizing the channels axis.  Its name should be `bn_name_base + '2c'`. Note that there is no ReLU activation function in this component.
  213. #
  214. # Shortcut path:
  215. # - The CONV2D has $F_3$ filters of shape (1,1) and a stride of (s,s). Its padding is "valid" and its name should be `conv_name_base + '1'`.
  216. # - The BatchNorm is normalizing the channels axis.  Its name should be `bn_name_base + '1'`.
  217. #
  218. # Final step:
  219. # - The shortcut and the main path values are added together.
  220. # - Then apply the ReLU activation function. This has no name and no hyperparameters.
  221. #    
  222. # **Exercise**: Implement the convolutional block. We have implemented the first component of the main path; you should implement the rest. As before, always use 0 as the seed for the random initialization, to ensure consistency with our grader.
  223. # - [Conv Hint](https://keras.io/layers/convolutional/#conv2d)
  224. # - [BatchNorm Hint](https://keras.io/layers/normalization/#batchnormalization) (axis: Integer, the axis that should be normalized (typically the features axis))
  225. # - For the activation, use:  `Activation('relu')(X)`
  226. # - [Addition Hint](https://keras.io/layers/merge/#add)
  227.  
  228. # In[15]:
  229.  
  230. # GRADED FUNCTION: convolutional_block
  231.  
  232. def convolutional_block(X, f, filters, stage, block, s = 2):
  233.     """
  234.    Implementation of the convolutional block as defined in Figure 4
  235.    
  236.    Arguments:
  237.    X -- input tensor of shape (m, n_H_prev, n_W_prev, n_C_prev)
  238.    f -- integer, specifying the shape of the middle CONV's window for the main path
  239.    filters -- python list of integers, defining the number of filters in the CONV layers of the main path
  240.    stage -- integer, used to name the layers, depending on their position in the network
  241.    block -- string/character, used to name the layers, depending on their position in the network
  242.    s -- Integer, specifying the stride to be used
  243.    
  244.    Returns:
  245.    X -- output of the convolutional block, tensor of shape (n_H, n_W, n_C)
  246.    """
  247.    
  248.     # defining name basis
  249.     conv_name_base = 'res' + str(stage) + block + '_branch'
  250.     bn_name_base = 'bn' + str(stage) + block + '_branch'
  251.    
  252.     # Retrieve Filters
  253.     F1, F2, F3 = filters
  254.    
  255.     # Save the input value
  256.     X_shortcut = X
  257.  
  258.  
  259.     ##### MAIN PATH #####
  260.     # First component of main path
  261.     X = Conv2D(F1, (1, 1), strides = (s,s), padding = 'valid',name = conv_name_base + '2a', kernel_initializer = glorot_uniform(seed=0))(X)
  262.     X = BatchNormalization(axis = 3, name = bn_name_base + '2a')(X)
  263.     X = Activation('relu')(X)
  264.    
  265.     ### START CODE HERE ###
  266.  
  267.     # Second component of main path (≈3 lines)
  268.     X = Conv2D(filters = F2, kernel_size = (f, f), strides = (1,1), padding = 'same', name = conv_name_base + '2b', kernel_initializer = glorot_uniform(seed=0))(X)
  269.     X = BatchNormalization(axis = 3, name = bn_name_base + '2b')(X)
  270.     X = Activation('relu')(X)
  271.  
  272.     # Third component of main path (≈2 lines)
  273.     X = Conv2D(filters = F3, kernel_size = (1, 1), strides = (1,1), padding = 'valid', name = conv_name_base + '2c', kernel_initializer = glorot_uniform(seed=0))(X)
  274.     X = BatchNormalization(axis = 3, name = bn_name_base + '2c')(X)
  275.  
  276.     ##### SHORTCUT PATH #### (≈2 lines)
  277.     X_shortcut = Conv2D(filters = F3, kernel_size = (1, 1), strides = (s,s), padding = 'valid', name = conv_name_base + '1', kernel_initializer = glorot_uniform(seed=0))(X_shortcut)
  278.     X_shortcut = BatchNormalization(axis = 3, name = bn_name_base + '1')(X_shortcut)
  279.  
  280.     # Final step: Add shortcut value to main path, and pass it through a RELU activation (≈2 lines)
  281.     X = Add()([X, X_shortcut])
  282.     X = Activation('relu')(X)
  283.    
  284.     ### END CODE HERE ###
  285.    
  286.     return X
  287.  
  288.  
  289. # In[16]:
  290.  
  291. tf.reset_default_graph()
  292.  
  293. with tf.Session() as test:
  294.     np.random.seed(1)
  295.     A_prev = tf.placeholder("float", [3, 4, 4, 6])
  296.     X = np.random.randn(3, 4, 4, 6)
  297.     A = convolutional_block(A_prev, f = 2, filters = [2, 4, 6], stage = 1, block = 'a')
  298.     test.run(tf.global_variables_initializer())
  299.     out = test.run([A], feed_dict={A_prev: X, K.learning_phase(): 0})
  300.     print("out = " + str(out[0][1][1][0]))
  301.  
  302.  
  303. # **Expected Output**:
  304. #
  305. # <table>
  306. #     <tr>
  307. #         <td>
  308. #             **out**
  309. #         </td>
  310. #         <td>
  311. #            [ 0.09018463  1.23489773  0.46822017  0.0367176   0.          0.65516603]
  312. #         </td>
  313. #     </tr>
  314. #
  315. # </table>
  316.  
  317. # ## 3 - Building your first ResNet model (50 layers)
  318. #
  319. # You now have the necessary blocks to build a very deep ResNet. The following figure describes in detail the architecture of this neural network. "ID BLOCK" in the diagram stands for "Identity block," and "ID BLOCK x3" means you should stack 3 identity blocks together.
  320. #
  321. # <img src="images/resnet_kiank.png" style="width:850px;height:150px;">
  322. # <caption><center> <u> <font color='purple'> **Figure 5** </u><font color='purple'>  : **ResNet-50 model** </center></caption>
  323. #
  324. # The details of this ResNet-50 model are:
  325. # - Zero-padding pads the input with a pad of (3,3)
  326. # - Stage 1:
  327. #     - The 2D Convolution has 64 filters of shape (7,7) and uses a stride of (2,2). Its name is "conv1".
  328. #     - BatchNorm is applied to the channels axis of the input.
  329. #     - MaxPooling uses a (3,3) window and a (2,2) stride.
  330. # - Stage 2:
  331. #     - The convolutional block uses three set of filters of size [64,64,256], "f" is 3, "s" is 1 and the block is "a".
  332. #     - The 2 identity blocks use three set of filters of size [64,64,256], "f" is 3 and the blocks are "b" and "c".
  333. # - Stage 3:
  334. #     - The convolutional block uses three set of filters of size [128,128,512], "f" is 3, "s" is 2 and the block is "a".
  335. #     - The 3 identity blocks use three set of filters of size [128,128,512], "f" is 3 and the blocks are "b", "c" and "d".
  336. # - Stage 4:
  337. #     - The convolutional block uses three set of filters of size [256, 256, 1024], "f" is 3, "s" is 2 and the block is "a".
  338. #     - The 5 identity blocks use three set of filters of size [256, 256, 1024], "f" is 3 and the blocks are "b", "c", "d", "e" and "f".
  339. # - Stage 5:
  340. #     - The convolutional block uses three set of filters of size [512, 512, 2048], "f" is 3, "s" is 2 and the block is "a".
  341. #     - The 2 identity blocks use three set of filters of size [512, 512, 2048], "f" is 3 and the blocks are "b" and "c".
  342. # - The 2D Average Pooling uses a window of shape (2,2) and its name is "avg_pool".
  343. # - The flatten doesn't have any hyperparameters or name.
  344. # - The Fully Connected (Dense) layer reduces its input to the number of classes using a softmax activation. Its name should be `'fc' + str(classes)`.
  345. #
  346. # **Exercise**: Implement the ResNet with 50 layers described in the figure above. We have implemented Stages 1 and 2. Please implement the rest. (The syntax for implementing Stages 3-5 should be quite similar to that of Stage 2.) Make sure you follow the naming convention in the text above.
  347. #
  348. # You'll need to use this function:
  349. # - Average pooling [see reference](https://keras.io/layers/pooling/#averagepooling2d)
  350. #
  351. # Here're some other functions we used in the code below:
  352. # - Conv2D: [See reference](https://keras.io/layers/convolutional/#conv2d)
  353. # - BatchNorm: [See reference](https://keras.io/layers/normalization/#batchnormalization) (axis: Integer, the axis that should be normalized (typically the features axis))
  354. # - Zero padding: [See reference](https://keras.io/layers/convolutional/#zeropadding2d)
  355. # - Max pooling: [See reference](https://keras.io/layers/pooling/#maxpooling2d)
  356. # - Fully conected layer: [See reference](https://keras.io/layers/core/#dense)
  357. # - Addition: [See reference](https://keras.io/layers/merge/#add)
  358.  
  359. # In[17]:
  360.  
  361. # GRADED FUNCTION: ResNet50
  362.  
  363. def ResNet50(input_shape = (64, 64, 3), classes = 6):
  364.     """
  365.    Implementation of the popular ResNet50 the following architecture:
  366.    CONV2D -> BATCHNORM -> RELU -> MAXPOOL -> CONVBLOCK -> IDBLOCK*2 -> CONVBLOCK -> IDBLOCK*3
  367.    -> CONVBLOCK -> IDBLOCK*5 -> CONVBLOCK -> IDBLOCK*2 -> AVGPOOL -> TOPLAYER
  368.  
  369.    Arguments:
  370.    input_shape -- shape of the images of the dataset
  371.    classes -- integer, number of classes
  372.  
  373.    Returns:
  374.    model -- a Model() instance in Keras
  375.    """
  376.    
  377.     # Define the input as a tensor with shape input_shape
  378.     X_input = Input(input_shape)
  379.  
  380.    
  381.     # Zero-Padding
  382.     X = ZeroPadding2D((3, 3))(X_input)
  383.    
  384.     # Stage 1
  385.     X = Conv2D(64, (7, 7), strides = (2, 2), name = 'conv1', kernel_initializer = glorot_uniform(seed=0))(X)
  386.     X = BatchNormalization(axis = 3, name = 'bn_conv1')(X)
  387.     X = Activation('relu')(X)
  388.     X = MaxPooling2D((3, 3), strides=(2, 2))(X)
  389.  
  390.     # Stage 2
  391.     X = convolutional_block(X, f = 3, filters = [64, 64, 256], stage = 2, block='a', s = 1)
  392.     X = identity_block(X, 3, [64, 64, 256], stage=2, block='b')
  393.     X = identity_block(X, 3, [64, 64, 256], stage=2, block='c')
  394.  
  395.     ### START CODE HERE ###
  396.  
  397.     # Stage 3 (≈4 lines)
  398.     X = convolutional_block(X, f = 3, filters = [128,128, 512], stage = 3, block='a', s = 2)
  399.     X = identity_block(X, 3, [128,128, 512], stage=3, block='b')
  400.     X = identity_block(X, 3, [128,128, 512], stage=3, block='c')
  401.     X = identity_block(X, 3, [128,128, 512], stage=3, block='d')
  402.  
  403.     # Stage 4 (≈6 lines)
  404.     X = convolutional_block(X, f = 3, filters = [256,256, 1024], stage = 4, block='a', s = 2)
  405.     X = identity_block(X, 3, [256,256, 1024], stage=4, block='b')
  406.     X = identity_block(X, 3, [256,256, 1024], stage=4, block='c')
  407.     X = identity_block(X, 3, [256,256, 1024], stage=4, block='d')
  408.     X = identity_block(X, 3, [256,256, 1024], stage=4, block='e')
  409.     X = identity_block(X, 3, [256,256, 1024], stage=4, block='f')
  410.  
  411.     # Stage 5 (≈3 lines)
  412.     X = convolutional_block(X, f = 3, filters = [512,512, 2048], stage = 5, block='a', s = 2)
  413.     X = identity_block(X, 3, [512,512, 2048], stage=5, block='b')
  414.     X = identity_block(X, 3, [512,512, 2048], stage=5, block='c')
  415.  
  416.     # AVGPOOL (≈1 line). Use "X = AveragePooling2D(...)(X)"
  417.     X = AveragePooling2D(pool_size=(2, 2), strides=None, padding='valid', data_format=None)(X)
  418.     #AveragePooling2D(pool_size=(2, 2), strides=None, padding='valid', data_format=None)
  419.    
  420.     ### END CODE HERE ###
  421.  
  422.     # output layer
  423.     X = Flatten()(X)
  424.     X = Dense(classes, activation='softmax', name='fc' + str(classes), kernel_initializer = glorot_uniform(seed=0))(X)
  425.    
  426.    
  427.     # Create model
  428.     model = Model(inputs = X_input, outputs = X, name='ResNet50')
  429.  
  430.     return model
  431.  
  432.  
  433. # Run the following code to build the model's graph. If your implementation is not correct you will know it by checking your accuracy when running `model.fit(...)` below.
  434.  
  435. # In[18]:
  436.  
  437. model = ResNet50(input_shape = (64, 64, 3), classes = 6)
  438.  
  439.  
  440. # As seen in the Keras Tutorial Notebook, prior training a model, you need to configure the learning process by compiling the model.
  441.  
  442. # In[19]:
  443.  
  444. model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
  445.  
  446.  
  447. # The model is now ready to be trained. The only thing you need is a dataset.
  448.  
  449. # Let's load the SIGNS Dataset.
  450. #
  451. # <img src="images/signs_data_kiank.png" style="width:450px;height:250px;">
  452. # <caption><center> <u> <font color='purple'> **Figure 6** </u><font color='purple'>  : **SIGNS dataset** </center></caption>
  453. #
  454.  
  455. # In[20]:
  456.  
  457. X_train_orig, Y_train_orig, X_test_orig, Y_test_orig, classes = load_dataset()
  458.  
  459. # Normalize image vectors
  460. X_train = X_train_orig/255.
  461. X_test = X_test_orig/255.
  462.  
  463. # Convert training and test labels to one hot matrices
  464. Y_train = convert_to_one_hot(Y_train_orig, 6).T
  465. Y_test = convert_to_one_hot(Y_test_orig, 6).T
  466.  
  467. print ("number of training examples = " + str(X_train.shape[0]))
  468. print ("number of test examples = " + str(X_test.shape[0]))
  469. print ("X_train shape: " + str(X_train.shape))
  470. print ("Y_train shape: " + str(Y_train.shape))
  471. print ("X_test shape: " + str(X_test.shape))
  472. print ("Y_test shape: " + str(Y_test.shape))
  473.  
  474.  
  475. # Run the following cell to train your model on 2 epochs with a batch size of 32. On a CPU it should take you around 5min per epoch.
  476.  
  477. # In[21]:
  478.  
  479. model.fit(X_train, Y_train, epochs = 2, batch_size = 32)
  480.  
  481.  
  482. # **Expected Output**:
  483. #
  484. # <table>
  485. #     <tr>
  486. #         <td>
  487. #             ** Epoch 1/2**
  488. #         </td>
  489. #         <td>
  490. #            loss: between 1 and 5, acc: between 0.2 and 0.5, although your results can be different from ours.
  491. #         </td>
  492. #     </tr>
  493. #     <tr>
  494. #         <td>
  495. #             ** Epoch 2/2**
  496. #         </td>
  497. #         <td>
  498. #            loss: between 1 and 5, acc: between 0.2 and 0.5, you should see your loss decreasing and the accuracy increasing.
  499. #         </td>
  500. #     </tr>
  501. #
  502. # </table>
  503.  
  504. # Let's see how this model (trained on only two epochs) performs on the test set.
  505.  
  506. # In[22]:
  507.  
  508. preds = model.evaluate(X_test, Y_test)
  509. print ("Loss = " + str(preds[0]))
  510. print ("Test Accuracy = " + str(preds[1]))
  511.  
  512.  
  513. # **Expected Output**:
  514. #
  515. # <table>
  516. #     <tr>
  517. #         <td>
  518. #             **Test Accuracy**
  519. #         </td>
  520. #         <td>
  521. #            between 0.16 and 0.25
  522. #         </td>
  523. #     </tr>
  524. #
  525. # </table>
  526.  
  527. # For the purpose of this assignment, we've asked you to train the model only for two epochs. You can see that it achieves poor performances. Please go ahead and submit your assignment; to check correctness, the online grader will run your code only for a small number of epochs as well.
  528.  
  529. # After you have finished this official (graded) part of this assignment, you can also optionally train the ResNet for more iterations, if you want. We get a lot better performance when we train for ~20 epochs, but this will take more than an hour when training on a CPU.
  530. #
  531. # Using a GPU, we've trained our own ResNet50 model's weights on the SIGNS dataset. You can load and run our trained model on the test set in the cells below. It may take ≈1min to load the model.
  532.  
  533. # In[23]:
  534.  
  535. model = load_model('ResNet50.h5')
  536.  
  537.  
  538. # In[24]:
  539.  
  540. preds = model.evaluate(X_test, Y_test)
  541. print ("Loss = " + str(preds[0]))
  542. print ("Test Accuracy = " + str(preds[1]))
  543.  
  544.  
  545. # ResNet50 is a powerful model for image classification when it is trained for an adequate number of iterations. We hope you can use what you've learnt and apply it to your own classification problem to perform state-of-the-art accuracy.
  546. #
  547. # Congratulations on finishing this assignment! You've now implemented a state-of-the-art image classification system!
  548.  
  549. # ## 4 - Test on your own image (Optional/Ungraded)
  550.  
  551. # If you wish, you can also take a picture of your own hand and see the output of the model. To do this:
  552. #     1. Click on "File" in the upper bar of this notebook, then click "Open" to go on your Coursera Hub.
  553. #     2. Add your image to this Jupyter Notebook's directory, in the "images" folder
  554. #     3. Write your image's name in the following code
  555. #     4. Run the code and check if the algorithm is right!
  556.  
  557. # In[ ]:
  558.  
  559. img_path = 'images/my_image.jpg'
  560. img = image.load_img(img_path, target_size=(64, 64))
  561. x = image.img_to_array(img)
  562. x = np.expand_dims(x, axis=0)
  563. x = preprocess_input(x)
  564. print('Input image shape:', x.shape)
  565. my_image = scipy.misc.imread(img_path)
  566. imshow(my_image)
  567. print("class prediction vector [p(0), p(1), p(2), p(3), p(4), p(5)] = ")
  568. print(model.predict(x))
  569.  
  570.  
  571. # You can also print a summary of your model by running the following code.
  572.  
  573. # In[ ]:
  574.  
  575. model.summary()
  576.  
  577.  
  578. # Finally, run the code below to visualize your ResNet50. You can also download a .png picture of your model by going to "File -> Open...-> model.png".
  579.  
  580. # In[ ]:
  581.  
  582. plot_model(model, to_file='model.png')
  583. SVG(model_to_dot(model).create(prog='dot', format='svg'))
  584.  
  585.  
  586. # <font color='blue'>
  587. # **What you should remember:**
  588. # - Very deep "plain" networks don't work in practice because they are hard to train due to vanishing gradients.  
  589. # - The skip-connections help to address the Vanishing Gradient problem. They also make it easy for a ResNet block to learn an identity function.
  590. # - There are two main type of blocks: The identity block and the convolutional block.
  591. # - Very deep Residual Networks are built by stacking these blocks together.
  592.  
  593. # ### References
  594. #
  595. # This notebook presents the ResNet algorithm due to He et al. (2015). The implementation here also took significant inspiration and follows the structure given in the github repository of Francois Chollet:
  596. #
  597. # - Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun - [Deep Residual Learning for Image Recognition (2015)](https://arxiv.org/abs/1512.03385)
  598. # - Francois Chollet's github repository: https://github.com/fchollet/deep-learning-models/blob/master/resnet50.py
  599. #
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement
Public Pastes
Advertisement
We use cookies for various purposes including analytics. By continuing to use Pastebin, you agree to our use of cookies as described in the Cookies Policy.  OK, I Understand
Not a member of Pastebin yet?
Sign Up, it unlocks many cool features!