kaggle-rsna-pneumonia-detection-challenge

CNN References:

• 關於影像辨識，所有你應該知道的深度學習模型 | Medium @2018-02-04
• 一文讀懂：R-CNN、Fast R-CNN、Faster R-CNN、YOLO、SSD @2018-05-02
• 如何评价 Kaiming He 最新的 Mask R-CNN? | 知乎 @2017-03-23
• Splash of Color: Instance Segmentation with Mask R-CNN and TensorFlow | Medium @2018-05-20
• Deep Learning in Computer Vision | Coursera

Keras Tutorial

• How to Use the Keras Functional API for Deep Learning @2017-10-27

   from keras.models import Model
   from keras.layers import Input
   from keras.layers import Dense
   visible = Input(shape=(10,))
   hidden1 = Dense(10, activation='relu')(visible)
   hidden2 = Dense(20, activation='relu')(hidden1)
   hidden3 = Dense(10, activation='relu')(hidden2)
   output = Dense(1, activation='sigmoid')(hidden3)
   model = Model(inputs=visible, outputs=output)
   print(model.summary())

   Layer (type)                 Output Shape              Param #
   ===========================================================================
   input_1 (InputLayer)         (None, 10)                0
   ___________________________________________________________________________
   dense_1 (Dense)              (None, 10)                110 =(10+1)*10
   ___________________________________________________________________________
   dense_2 (Dense)              (None, 20)                220 =(10+1)*20
   ___________________________________________________________________________
   dense_3 (Dense)              (None, 10)                210 =(20+1)*10
   ___________________________________________________________________________
   dense_4 (Dense)              (None, 1)                 11  =(10+1)*1
   ===========================================================================
   Total params: 551
   Trainable params: 551
   Non-trainable params: 0
  

• How to calculate the number of parameters for convolutional neural network?

   from keras.models import Model
   from keras.layers import Input
   from keras.layers import Dense
   from keras.layers.convolutional import Conv2D
   from keras.layers.pooling import MaxPooling2D
   visible = Input(shape=(64, 64, 1))
   # fitler num: 32, filter shape: (4, 4)
   conv1 = Conv2D(32, kernel_size=4, activation='relu')(visible)
   pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
   conv2 = Conv2D(16, kernel_size=4, activation='relu')(pool1)
   pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
   hidden1 = Dense(10, activation='relu')(pool2)
   output = Dense(1, activation='sigmoid')(hidden1)
   model = Model(inputs=visible, outputs=output)
   print(model.summary())

   Layer (type)                 Output Shape              Param #
   ===========================================================================
   input_1 (InputLayer)         (None, 64, 64, 1)         0
   ___________________________________________________________________________
   # filter number 32, size 4*4. Say k=32, m=4, n=4, c=1
   conv2d_1 (Conv2D)            (None, 61, 61, 32)        544  = (4*4*1+1)*32
                                       61=64-4+1                 (m*n*c+1)*k
   ___________________________________________________________________________
   max_pooling2d_1 (MaxPooling2 (None, 30, 30, 32)        0
                                       30=61/2
   ___________________________________________________________________________
   # filter number 16, size 4*4. Say k=16, m=4, n=4, c=32
   conv2d_2 (Conv2D)            (None, 27, 27, 16)        8208 = (4*4*32+1)*16
                                       27=30-4+1                 (m*n*c+1)*k
   ___________________________________________________________________________
   max_pooling2d_2 (MaxPooling2 (None, 13, 13, 16)        0
                                       13=27/2
   ___________________________________________________________________________
   dense_1 (Dense)              (None, 13, 13, 10)        170  = (16+1)*10
   ___________________________________________________________________________
   dense_2 (Dense)              (None, 13, 13, 1)         11
   ===========================================================================
   Total params: 8,933
   Trainable params: 8,933
   Non-trainable params: 0